betaFeedback

WorksComputer Science

T2ST: Data-Efficient Spatio-Temporal Foundation Models via Temporal Data and Knowledge Transfer

Abstract

Spatio-temporal foundation models (STFMs) require large-scale training corpora, yet synchronized observations and spatial metadata are costly to acquire and curate, limiting the availability of suitable data. By contrast, time-series data are abundant and easier to collect. Since observations at each location form a time series, this motivates a natural question: can abundant time-series data enhance STFMs when spatio-temporal observations are scarce? To bridge this gap, we propose T2ST, a data-efficient training framework that augments limited spatio-temporal observations with selected auxiliary time series and transfers knowledge from a frozen time-series foundation model. Our key insight is that temporal pattern learning can benefit from abundant external time series, while spatial dependency learning remains grounded in real spatio-temporal observations. Guided by this insight, T2ST retrieves matched sequences to reinforce sparsely observed temporal patterns and selects reliable complementary sequences to broaden temporal coverage. Temporal relation distillation aligns relationships among historical patches, while multi-scale prediction distillation transfers forecasting behavior at fine and coarse temporal resolutions, enabling knowledge transfer across heterogeneous architectures. Using only 10% of the original spatio-temporal data, T2ST enables STFMs to achieve overall forecasting performance competitive with or superior to their full-data counterparts across six long-term and twelve short-term forecasting datasets. These results establish a new paradigm for data-efficient training of spatio-temporal foundation models. Code is available at https://anonymous.4open.science/r/T2ST-65F3/.

Introduction

Spatio-temporal foundation models (STFMs) aim to support forecasting across diverse application domains, including intelligent transportation [1, 2, 3, 4], energy management [5], and environmental monitoring [6]. By pre-training on heterogeneous spatio-temporal datasets, these models move beyond the conventional one-model-per-dataset paradigm and enable knowledge sharing across cities and sensing systems [4, 7, 8, 9]. However, the scarcity of spatio-temporal training data remains a major bottleneck for STFMs. Building large-scale corpora involves collecting observations across multiple locations and curating both the measurements and their associated spatial information. The resulting acquisition and curation costs make these corpora difficult to expand. This scarcity motivates data-efficient STFM training: maintaining forecasting performance with fewer spatio-temporal observations.

In contrast, time-series data can be collected from diverse sources without requiring explicit spatial information [10]. This flexibility has supported substantially larger pre-training corpora for time-series foundation models (TSFMs) [11, 12], such as Moirai, Chronos, and TimesFM [13, 14, 15]. As summarized in Table I, reported TSFM corpora contain tens to hundreds of billions of time points, compared with hundreds of millions to over a billion for representative STFMs. Pre-training on these large and diverse corpora enables TSFMs to learn temporal regularities, including trends, periodicities, and dependencies across time scales. These regularities are also relevant to spatio-temporal forecasting, since observations at each location form a time series. We therefore ask: Can abundant time-series data and pretrained temporal knowledge reduce the amount of structured spatio-temporal data needed to train effective STFMs? To investigate this question, we supplement limited spatio-temporal training data with auxiliary time series and transfer temporal knowledge from a pretrained TSFM to the STFM.

ModelPoints (B)
Spatio-temporal foundation models (STFMs)
OpenCity [7]≈0.15\approx0.15
CompactST [4]0.290.29
WorldST [9]>1> 1
Time-series foundation models (TSFMs)
Moirai [13]>27> 27
Chronos [14]≈84\approx84
TimesFM [15]≈100\approx100

Table I. PRE-TRAINING DATA SCALES OF REPRESENTATIVE SPATIO-TEMPORAL AND TIME-SERIES FOUNDATION MODELS. THE B DENOTES 10910^9 REPORTED OBSERVATIONS OR TIME POINTS.

Effectively using these temporal resources requires addressing two challenges under limited spatial supervision. The first is data selection. Auxiliary corpora contain heterogeneous temporal dynamics, not all of which are equally useful for STFM training. Selecting only similar sequences, however, may restrict temporal coverage and overlook complementary patterns. For a foundation model intended to transfer across domains, similarity alone is therefore an incomplete criterion for selecting useful temporal data. The second is knowledge transfer. Applying a pretrained TSFM independently to each location does not explicitly model spatial interactions, while transferring its temporal knowledge into an STFM is complicated by differences in architecture, hidden dimensions, and temporal resolution. This raises the challenge of enriching temporal learning with external resources while keeping spatial learning grounded in the available structured observations.

These challenges motivate the design of T2ST, a data-efficient training framework that integrates auxiliary time-series selection with temporal knowledge distillation from a frozen TSFM into an STFM. Our key insight is that selected external time series and pretrained temporal knowledge can enrich an STFM’s temporal learning, while real spatio-temporal observations continue to provide supervision for spatial dependencies. This insight allows temporal learning to draw on resources beyond the limited spatio-temporal corpus, with the available structured observations supporting joint temporal and spatial modeling.

To address the data selection challenge, T2ST uses the limited spatio-temporal observations as a reference to construct two auxiliary subsets in a shared temporal representation space. Matched sequences are retrieved according to their similarity to the temporal patterns represented in the available reference, providing additional examples of familiar temporal patterns that are sparsely sampled in the available observations. Diverse sequences contribute complementary temporal patterns beyond this limited reference, broadening the coverage available for learning. For the diverse subset, teacher reliability filtering retains candidates that the TSFM can forecast reliably, while temporal pattern discovery identifies coherent groups of additional patterns. This selection strategy balances relevance to the observed dynamics with exposure to a broader range of temporal patterns.

To address the knowledge transfer challenge, T2ST distills temporal knowledge through two complementary objectives. Temporal relation distillation aligns relationships among historical patches rather than directly matching hidden features, accommodating differences in teacher and student representations. Multi-scale prediction distillation aligns forecasts at fine and coarse temporal resolutions, transferring both local variations and broader temporal trends. During training, auxiliary sequences are processed as separate single-node inputs through the student’s shared temporal pathway, while real spatio-temporal samples are processed with their spatial structure to support temporal and spatial learning.

In summary, the contributions of this study are as follows:

  • We introduce a data-efficient training paradigm for STFMs, in which abundant auxiliary time-series data enrich temporal pattern learning while limited real spatio-temporal observations provide spatial supervision. This broadens the sources of training information available to STFMs beyond structured spatio-temporal corpora.

  • We propose T2ST, a framework that integrates systematic selection of matched and diverse time series with knowledge distillation from a pretrained TSFM. Temporal relation distillation and multi-scale prediction distillation transfer temporal knowledge across heterogeneous teacher and student architectures.

  • We conduct extensive experiments across different spatio-temporal foundation models and datasets. Using only one-sixth of the full spatio-temporal training set, T2ST achieves performance competitive with or superior to that of counterparts trained on the full dataset, demonstrating its effectiveness, generality, and potential to reduce the structured-data cost of future STFMs.

Spatio-Temporal Foundation Models

Existing work on spatio-temporal foundation models broadly follows two directions [16]. The first adapts pretrained large language models (LLMs) to spatio-temporal prediction. UrbanGPT, for example, combines a spatio-temporal dependency encoder with instruction tuning to integrate spatio-temporal observations into an LLM [17]. The second develops models specifically for spatio-temporal data and pretrains them directly on heterogeneous spatio-temporal corpora. UniST combines masked pretraining with knowledge-guided prompting for prediction across urban scenarios [8], while OpenCity integrates Transformer-based temporal modeling with graph-based spatial modeling for traffic forecasting across cities [7]. CompactST develops compact models for scalable multi-domain pretraining [4], and UrbanFM investigates data and architecture scaling for urban spatio-temporal forecasting [9]. These directly pretrained models demonstrate strong transfer across cities and tasks, but their pretraining still relies on substantial spatio-temporal corpora.

To alleviate this constraint, temporal representations are often strengthened prior to or alongside spatial modeling. STEP pretrains a masked time-series encoder for a downstream spatio-temporal graph neural network, while BigST extracts features from long node-wise histories for traffic forecasting [18, 19]. Time Tracker separates channel-wise pretraining from inter-variable modeling; FactoST pretrain temporal backbones before spatial adaptation [20, 21].

Building on these efforts, T2ST focuses on strengthening existing STFMs under a limited spatio-temporal data budget. It combines selected external time series with guidance from an independently pretrained TSFM, retaining the STFM backbone during joint training.

Knowledge Distillation for Forecasting

Knowledge distillation (KD) transfers knowledge from a trained teacher to a student [22, 23, 24]. KD methods are broadly categorized as response-based, feature-based, and relation-based, aligning predictions, intermediate representations, and relationships among representations, respectively [25]. Relation-based methods transfer structural knowledge without directly aligning representation spaces [26, 27, 28].

Recent efforts have explored knowledge distillation for time-series and spatio-temporal forecasting [29, 30, 31, 32]. For time-series forecasting, TimeDistill transfers multi-scale and multi-period patterns from Transformer and CNN based teachers to lightweight MLP students [33]. DistilTS further explores distillation from pretrained TSFMs, combining horizon-aware supervision with temporal representation alignment to train compact forecasters [34]. In spatio-temporal forecasting, LightST transfers graph-structured spatial and temporal knowledge from a GNN teacher to an efficient MLP student for traffic prediction [35]. These distillation methods primarily address the computational and storage costs of powerful forecasting models by transferring their knowledge to compact students. However, efficient model deployment does not directly address another key challenge in STFM training: the limited availability of observations.

T2ST bridges this gap by distilling pretrained temporal knowledge into an STFM using abundant temporal-only data, while learning spatial dependencies from the available spatio-temporal observations. This allows temporal learning to draw on resources beyond the limited structured corpus, with the aim of reducing the amount of spatio-temporal data needed to train an effective STFM.

Preliminaries

Spatio-Temporal and Time-Series Data

Spatio-temporal (ST) data consist of observations collected over time from spatially related entities, such as road sensors or urban regions. Spatial structure is represented by a graph G=(V,E,A)\mathcal{G} = (\mathcal{V}, \mathcal{E}, \mathcal{A}), where V\mathcal{V} contains NN spatial entities, E\mathcal{E} denotes their relations, and A∈RN×N\mathcal{A} \in\mathbb{R}^{N \times N} is the adjacency matrix. The observations are denoted by S∈RT×N×C\mathbf{S} \in\mathbb{R}^{T \times N \times C}, where TT is the number of timestamps and CC-dimensional vector indicating CC features. Each St,:,:∈RN×C\mathbf{S}_{t,:,:} \in\mathbb{R}^{N \times C} records the measurements of all entities at time step tt. For forecasting, we denote the historical observations by Xt∈RL×N×C\mathbf{X}_t \in\mathbb{R}^{L \times N \times C} and the future targets by Yt∈RH×N×C\mathbf{Y}_t \in\mathbb{R}^{H \times N \times C}, covering time steps t−L+1t - L + 1 to tt and t+1t + 1 to t+Ht + H, respectively. Here, LL and HH denote the history length and forecasting horizon.

Time-series (TS) data are a time-ordered sequence of observations, denoted by Z∈RT×D\mathbf{Z} \in\mathbb{R}^{T \times D}, where DD is the number of observed variables. Each row zt∈RD\mathbf{z}_t \in\mathbb{R}^{D} records the values of these variables, such as electricity load or temperature, at time step tt. In particular, each node-wise trajectory S:,v,:∈RT×C\mathbf{S}_{:,v,:} \in\mathbb{R}^{T \times C} extracted from spatio-temporal data is itself a multivariate time series, or a univariate series when C=1C = 1.

Spatio-Temporal and Time-Series Foundation Models

The STFM is pretrained on observations from multiple spatial domains to learn reusable spatial and temporal patterns. For spatio-temporal forecasting, its prediction is expressed as

Y^=fθ(X,G)(1)\widehat{\mathbf{Y}} = f_{\theta}(\mathbf{X}, \mathcal{G}) \tag*{(1)}

where θ\theta is shared parameters across domains.

The TSFM is pretrained on a large collection of time series to learn transferable temporal patterns. Given an input time series u\mathbf{u}, its prediction is denoted by

v^=gϕ(u)(2)\widehat{\mathbf{v}} = g_{\phi}(\mathbf{u}) \tag*{(2)}

where ϕ\phi denotes the parameters of the TSFM.

Problem Formulation

We aim to train an STFM from limited synchronized observations using auxiliary temporal data and a pretrained TSFM. Across all ST training datasets used in our long-term experiments, eligible windows are selected consecutively from the end of each source’s training split, without random sampling across time. Let KK be the selected window count and MM the total eligible window count over the complete source timelines. The window budget is

rwin=KM.(3)r_{\mathrm{win}} = \frac{K}{M}. \tag*{(3)}

The 10% setting uses rwin≈0.1r_{\mathrm{win}} \approx0.1.

For k>0k > 0 consecutive selected windows from a source, let LL and HH denote the history and forecast lengths in its native time steps. The union of their input and target timestamps forms a contiguous block containing unique timestamps.

Q=k+L+H−1(4)Q = k + L + H - 1 \tag*{(4)}

Method

In this section, we present T2ST, a data-efficient training framework for spatio-temporal foundation models. T2ST augments limited spatio-temporal observations with selected temporal-only data and transfers temporal knowledge from a pretrained time-series foundation model through distillation. We first provide an overview of the framework and and then describe its key components in detail.

Overall Framework

As illustrated in Figure 1, T2ST consists of an offline temporal data selection stage and a distillation training stage. Given limited spatio-temporal observations and an auxiliary time-series corpus, it selects complementary temporal data and transfers knowledge from a pretrained TSFM to an STFM student. The main components are outlined below.

Figure 1. Overview of the T2ST framework. (a) Temporal data selection: a shared encoder maps spatio-temporal and auxiliary time-series windows into a common representation space. A calibrated spatio-temporal reference guides matched selection, while teacher reliability filtering and novel pattern discovery guide diversity selection. Selected embeddings identify the source windows retrieved by index to construct DM\mathcal{D}_{\mathrm{M}} and DD\mathcal{D}_{\mathrm{D}}. (b) Knowledge distillation: a frozen TSFM guides a STFM student through temporal relation distillation and multi-scale prediction distillation. The student is trained on DST\mathcal{D}_{\mathrm{ST}}, DM\mathcal{D}_{\mathrm{M}}, and DD\mathcal{D}_{\mathrm{D}}, sharing temporal modeling parameters across both input types and bypassing spatial interactions for temporal-only inputs. Snowflakes and flames indicate frozen and trainable parameters, respectively.

Temporal Data Selection. A shared encoder maps node-wise sequences from the spatio-temporal dataset and auxiliary time-series windows into a shared representation space. Using the spatio-temporal data as a reference, we construct a matched subset DM\mathcal{D}_{\mathrm{M}} containing similar temporal patterns and a diverse subset DD\mathcal{D}_{\mathrm{D}} containing additional patterns that the teacher can forecast reliably. Together, these subsets supplement the limited spatio-temporal dataset DST\mathcal{D}_{\mathrm{ST}}.

Teacher-Student Setup. The pretrained TSFM serves as a frozen teacher, processing each node-wise and each auxiliary temporal sequence independently to provide intermediate temporal representations and forecasts. The STFM student shares its temporal modeling parameters across both sequence types. It incorporates spatial interactions when processing spatio-temporal samples and bypasses these interactions for temporal-only samples, allowing auxiliary temporal sequences to support temporal learning without introducing artificial spatial dependencies.

Knowledge Distillation. We distill teacher knowledge via two complementary components: Temporal Relation Distillation and Multi-Scale Prediction Distillation.

Temporal Relation Distillation. This module transfers relationships among temporal patches across selected teacher–student layers. By aligning pairwise similarities derived from intermediate representations, it encourages the student to preserve the teacher’s temporal relational structure while accommodating differences in hidden dimensions and patch counts.

Multi-Scale Prediction Distillation. This module aligns teacher and student forecasts at multiple temporal resolutions through temporal aggregation. It complements temporal relation distillation by transferring forecasting behavior at the output level, capturing both fine temporal variations and broader trends.

The student is jointly trained on DST\mathcal{D}_{\mathrm{ST}}, DM\mathcal{D}_{\mathrm{M}}, and DD\mathcal{D}_{\mathrm{D}} using ground-truth supervision and both distillation objectives.

Temporal Data Selection

As shown in Figure 1(a), a shared pretrained encoder maps ST and TS windows into a common representation space. Temporal pattern discovery and exemplar selection establish an ST exemplar reference, and reference calibration determines the matching threshold. Auxiliary time series candidates are then scored by their proximity to this reference: nearby candidates undergo matched selection, while distant candidates pass through teacher reliability filtering, novel pattern discovery, and diversity selection. The selected embeddings are used to retrieve the corresponding temporal windows by index.

Figure 2. ST reference exemplar construction and threshold calibration.

Shared Temporal Representation

We construct encoder pretraining samples from the limited spatio-temporal data DST\mathcal{D}_{\mathrm{ST}} by treating each node’s observations as an independent time series, omitting spatial relations. Let DST′\mathcal{D}'_{\mathrm{ST}} denote a fixed collection of these sampled node-wise windows.

We consider three alternative implementations of the pretrained encoder EψE_{\psi} in Figure 1(a): TS2Vec [36], TS-TCC [37], and CoST [38]. Each encoder is pretrained separately on the same DST′\mathcal{D}'_{\mathrm{ST}}, with the preprocessing described below, using its own self-supervised objective and is then frozen throughout selection. Within each encoder configuration, the same frozen EψE_{\psi} processes both ST and TS windows, placing samples from the two sources in a shared representation space.

For a training window u=(x,y)u=(x,y), where x∈RL×1x\in\mathbb{R}^{L\times1} and y∈RH×1y\in\mathbb{R}^{H\times1} are the history and observed target, we concatenate them into a trajectory [x;y]∈R(L+H)×1[x;y]\in\mathbb{R}^{(L+H)\times1}. Mean aggregation to a common sampling interval produces s(u)∈RT×1s(u)\in\mathbb{R}^{T\times1}, where TT is the aligned sequence length. We then compute

s~(u)=s(u)−μ(s(u))σ(s(u))(5)\widetilde{s}(u)=\frac{s(u)-\mu(s(u))}{\sigma(s(u))} \tag*{(5)}
h(u)=Pool⁡(Eψ(s~(u)))(6)h(u)=\operatorname{Pool}\left(E_{\psi}(\widetilde{s}(u))\right) \tag*{(6)}
z(u)=h(u)∥h(u)∥2(7)z(u)=\frac{h(u)}{\lVert h(u)\rVert_{2}} \tag*{(7)}

where μ(s(u))\mu(s(u)) and σ(s(u))\sigma(s(u)) are the scalar mean and standard deviation of the aggregated trajectory. The encoder produces Eψ(s~(u))∈RTψ×dE_{\psi}(\widetilde{s}(u))\in\mathbb{R}^{T_{\psi}\times d}, where TψT_{\psi} is the encoder’s output sequence length. The operator Pool performs global max pooling, yielding h(u)∈Rdh(u)\in\mathbb{R}^{d}. The final L2L_{2} normalization produces z(u)z(u). These vectors z(u)z(u) form the ST Embeddings or TS Embeddings in Figure 1(a) and are used for subsequent pattern discovery and reference proximity scoring.

ST Exemplar Construction and Threshold Calibration

We construct ST reference exemplars to represent temporal patterns in the limited ST data and calibrate a proximity threshold to support both matched and diversity selection. Following Figure 1(a), this stage comprises temporal pattern discovery, exemplar selection, reference calibration, and reference proximity scoring.

Temporal pattern discovery. As illustrated in Figure 2(a), temporal pattern discovery maps ST reference embeddings to cluster assignments while preserving the identity of each window. Using a fixed random seed, we partition the available node-wise ST training windows WST\mathcal{W}_{\mathrm{ST}} into disjoint reference and calibration sets, Wref\mathcal{W}_{\mathrm{ref}} and Wcal\mathcal{W}_{\mathrm{cal}}. The reference set is used to fit the clustering transformation, discover temporal patterns, and select exemplars; the calibration set is reserved for subsequent threshold estimation.

Let Wref={ui}i=1nref\mathcal{W}_{\mathrm{ref}}=\{u_i\}_{i=1}^{n_{\mathrm{ref}}}, where nref=∣Wref∣n_{\mathrm{ref}}=|\mathcal{W}_{\mathrm{ref}}|. Stacking the normalized embeddings z(ui)z(u_i) as rows gives Zref∈Rnref×dZ_{\mathrm{ref}}\in\mathbb{R}^{n_{\mathrm{ref}}\times d}. The row labels uiu_i in Figure 2(a) identify the original windows associated with these embeddings. We fit principal component analysis (PCA) [39] on ZrefZ_{\mathrm{ref}}, with reference mean zˉST=nref−1∑i=1nrefz(ui)∈Rd\bar{z}_{\mathrm{ST}}=n_{\mathrm{ref}}^{-1}\sum_{i=1}^{n_{\mathrm{ref}}}z(u_i)\in\mathbb{R}^{d} and loading matrix PST∈Rd×qP_{\mathrm{ST}}\in\mathbb{R}^{d\times q} containing the leading qq principal directions. The individual and stacked PCA coordinates are

a(u)=PSTT(z(u)−zˉST)∈Rq(8)a(u)=P_{\mathrm{ST}}^{\mathsf{T}}\left(z(u)-\bar{z}_{\mathrm{ST}}\right)\in\mathbb{R}^{q} \tag*{(8)}

where zˉST∈Rd\bar{z}_{\mathrm{ST}}\in\mathbb{R}^{d} is the mean reference embedding and PST∈Rd×qP_{\mathrm{ST}}\in\mathbb{R}^{d\times q} contains the leading qq principal directions. This operation centers each embedding and projects its feature dimension from dd to qq, with qq capped at the effective rank of the centered reference matrix. The resulting clustering coordinates form a matrix Aref∈Rnref×qA_{\mathrm{ref}}\in\mathbb{R}^{n_{\mathrm{ref}}\times q}.

We apply HDBSCAN [40], [41] to the rows of ArefA_{\mathrm{ref}} using Euclidean distance, fixed neighborhood and minimum-cluster-size parameters, and excess-of-mass cluster extraction. It determines the number of clusters KK and assigns each reference window a label

Ck={ui∈Wref:ci=k},k=−1,1,…,K(9)\mathcal{C}_{k}=\{u_i\in\mathcal{W}_{\mathrm{ref}}:c_i=k\},\quad k=-1,1,\ldots,K \tag*{(9)}

@page 5/14 starts_mid=0 ends_mid=1

Temporal Data Selection

As shown in Figure 1(a), a shared pretrained encoder maps ST and TS windows into a common representation space. Temporal pattern discovery and exemplar selection establish an ST exemplar reference, and reference calibration determines the matching threshold. Auxiliary time series candidates are then scored by their proximity to this reference: nearby candidates undergo matched selection, while distant candidates pass through teacher reliability filtering, novel pattern discovery, and diversity selection. The selected embeddings are used to retrieve the corresponding temporal windows by index.

Figure 2 (PDF p. 5). ST reference exemplar construction and threshold calibration.

Shared Temporal Representation

We construct encoder pretraining samples from the limited spatio-temporal data DST\mathcal{D}_{\mathrm{ST}} by treating each node’s observations as an independent time series, omitting spatial relations. Let DST′\mathcal{D}'_{\mathrm{ST}} denote a fixed collection of these sampled node-wise windows.

We consider three alternative implementations of the pretrained encoder EψE_{\psi} in Figure 1(a): TS2Vec [36], TS-TCC [37], and CoST [38]. Each encoder is pretrained separately on the same DST′\mathcal{D}'_{\mathrm{ST}}, with the preprocessing described below, using its own self-supervised objective and is then frozen throughout selection. Within each encoder configuration, the same frozen EψE_{\psi} processes both ST and TS windows, placing samples from the two sources in a shared representation space.

For a training window u=(x,y)u=(x,y), where x∈RL×1x\in\mathbb{R}^{L\times1} and y∈RH×1y\in\mathbb{R}^{H\times1} are the history and observed target, we concatenate them into a trajectory [x;y]∈R(L+H)×1[x;y]\in\mathbb{R}^{(L+H)\times1}. Mean aggregation to a common sampling interval produces s(u)∈RT×1s(u)\in\mathbb{R}^{T\times1}, where TT is the aligned sequence length. We then compute

s~(u)=s(u)−μ(s(u))σ(s(u))(10)\widetilde{s}(u)=\frac{s(u)-\mu(s(u))}{\sigma(s(u))} \tag*{(10)}
h(u)=Pool⁡(Eψ(s~(u)))(11)h(u)=\operatorname{Pool}\left(E_{\psi}(\widetilde{s}(u))\right) \tag*{(11)}
z(u)=h(u)∥h(u)∥2(12)z(u)=\frac{h(u)}{\lVert h(u)\rVert_{2}} \tag*{(12)}

where μ(s(u))\mu(s(u)) and σ(s(u))\sigma(s(u)) are the scalar mean and standard deviation of the aggregated trajectory. The encoder produces Eψ(s~(u))∈RTψ×dE_{\psi}(\widetilde{s}(u))\in\mathbb{R}^{T_{\psi}\times d}, where TψT_{\psi} is the encoder’s output sequence length. The operator Pool performs global max pooling, yielding h(u)∈Rdh(u)\in\mathbb{R}^{d}. The final L2L_{2} normalization produces z(u)z(u). These vectors z(u)z(u) form the ST Embeddings or TS Embeddings in Figure 1(a) and are used for subsequent pattern discovery and reference proximity scoring.

ST Exemplar Construction and Threshold Calibration

We construct ST reference exemplars to represent temporal patterns in the limited ST data and calibrate a proximity threshold to support both matched and diversity selection. Following Figure 1(a), this stage comprises temporal pattern discovery, exemplar selection, reference calibration, and reference proximity scoring.

Temporal pattern discovery. As illustrated in Figure 2(a), temporal pattern discovery maps ST reference embeddings to cluster assignments while preserving the identity of each window. Using a fixed random seed, we partition the available node-wise ST training windows WST\mathcal{W}_{\mathrm{ST}} into disjoint reference and calibration sets, Wref\mathcal{W}_{\mathrm{ref}} and Wcal\mathcal{W}_{\mathrm{cal}}. The reference set is used to fit the clustering transformation, discover temporal patterns, and select exemplars; the calibration set is reserved for subsequent threshold estimation.

Let Wref={ui}i=1nref\mathcal{W}_{\mathrm{ref}}=\{u_i\}_{i=1}^{n_{\mathrm{ref}}}, where nref=∣Wref∣n_{\mathrm{ref}}=|\mathcal{W}_{\mathrm{ref}}|. Stacking the normalized embeddings z(ui)z(u_i) as rows gives Zref∈Rnref×dZ_{\mathrm{ref}}\in\mathbb{R}^{n_{\mathrm{ref}}\times d}. The row labels uiu_i in Figure 2(a) identify the original windows associated with these embeddings. We fit principal component analysis (PCA) [39] on ZrefZ_{\mathrm{ref}}, with reference mean zˉST=nref−1∑i=1nrefz(ui)∈Rd\bar{z}_{\mathrm{ST}}=n_{\mathrm{ref}}^{-1}\sum_{i=1}^{n_{\mathrm{ref}}}z(u_i)\in\mathbb{R}^{d} and loading matrix PST∈Rd×qP_{\mathrm{ST}}\in\mathbb{R}^{d\times q} containing the leading qq principal directions. The individual and stacked PCA coordinates are

a(u)=PSTT(z(u)−zˉST)∈Rq(13)a(u)=P_{\mathrm{ST}}^{\mathsf{T}}\left(z(u)-\bar{z}_{\mathrm{ST}}\right)\in\mathbb{R}^{q} \tag*{(13)}

where zˉST∈Rd\bar{z}_{\mathrm{ST}}\in\mathbb{R}^{d} is the mean reference embedding and PST∈Rd×qP_{\mathrm{ST}}\in\mathbb{R}^{d\times q} contains the leading qq principal directions. This operation centers each embedding and projects its feature dimension from dd to qq, with qq capped at the effective rank of the centered reference matrix. The resulting clustering coordinates form a matrix Aref∈Rnref×qA_{\mathrm{ref}}\in\mathbb{R}^{n_{\mathrm{ref}}\times q}.

We apply HDBSCAN [40], [41] to the rows of ArefA_{\mathrm{ref}} using Euclidean distance, fixed neighborhood and minimum-cluster-size parameters, and excess-of-mass cluster extraction. It determines the number of clusters KK and assigns each reference window a label

Ck={ui∈Wref:ci=k},k=−1,1,…,K(14)\mathcal{C}_{k}=\{u_i\in\mathcal{W}_{\mathrm{ref}}:c_i=k\},\quad k=-1,1,\ldots,K \tag*{(14)}

@page 5/14 starts_mid=0 ends_mid=1

Temporal Data Selection

As shown in Figure 1(a), a shared pretrained encoder maps ST and TS windows into a common representation space. Temporal pattern discovery and exemplar selection establish an ST exemplar reference, and reference calibration determines the matching threshold. Auxiliary time series candidates are then scored by their proximity to this reference: nearby candidates undergo matched selection, while distant candidates pass through teacher reliability filtering, novel pattern discovery, and diversity selection. The selected embeddings are used to retrieve the corresponding temporal windows by index.

Figure 2 (PDF p. 5, second). ST reference exemplar construction and threshold calibration.

Shared Temporal Representation

We construct encoder pretraining samples from the limited spatio-temporal data DST\mathcal{D}_{\mathrm{ST}} by treating each node’s observations as an independent time series, omitting spatial relations. Let DST′\mathcal{D}'_{\mathrm{ST}} denote a fixed collection of these sampled node-wise windows.

We consider three alternative implementations of the pretrained encoder EψE_{\psi} in Figure 1(a): TS2Vec [36], TS-TCC [37], and CoST [38]. Each encoder is pretrained separately on the same DST′\mathcal{D}'_{\mathrm{ST}}, with the preprocessing described below, using its own self-supervised objective and is then frozen throughout selection. Within each encoder configuration, the same frozen EψE_{\psi} processes both ST and TS windows, placing samples from the two sources in a shared representation space.

For a training window u=(x,y)u=(x,y), where x∈RL×1x\in\mathbb{R}^{L\times1} and y∈RH×1y\in\mathbb{R}^{H\times1} are the history and observed target, we concatenate them into a trajectory [x;y]∈R(L+H)×1[x;y]\in\mathbb{R}^{(L+H)\times1}. Mean aggregation to a common sampling interval produces s(u)∈RT×1s(u)\in\mathbb{R}^{T\times1}, where TT is the aligned sequence length. We then compute

s~(u)=s(u)−μ(s(u))σ(s(u))(15)\widetilde{s}(u)=\frac{s(u)-\mu(s(u))}{\sigma(s(u))} \tag*{(15)}
h(u)=Pool⁡(Eψ(s~(u)))(16)h(u)=\operatorname{Pool}\left(E_{\psi}(\widetilde{s}(u))\right) \tag*{(16)}
z(u)=h(u)∥h(u)∥2(17)z(u)=\frac{h(u)}{\lVert h(u)\rVert_{2}} \tag*{(17)}

where μ(s(u))\mu(s(u)) and σ(s(u))\sigma(s(u)) are the scalar mean and standard deviation of the aggregated trajectory. The encoder produces Eψ(s~(u))∈RTψ×dE_{\psi}(\widetilde{s}(u))\in\mathbb{R}^{T_{\psi}\times d}, where TψT_{\psi} is the encoder’s output sequence length. The operator Pool performs global max pooling, yielding h(u)∈Rdh(u)\in\mathbb{R}^{d}. The final L2L_{2} normalization produces z(u)z(u). These vectors z(u)z(u) form the ST Embeddings or TS Embeddings in Figure 1(a) and are used for subsequent pattern discovery and reference proximity scoring.

ST Exemplar Construction and Threshold Calibration

We construct ST reference exemplars to represent temporal patterns in the limited ST data and calibrate a proximity threshold to support both matched and diversity selection. Following Figure 1(a), this stage comprises temporal pattern discovery, exemplar selection, reference calibration, and reference proximity scoring.

Temporal pattern discovery. As illustrated in Figure 2(a), temporal pattern discovery maps ST reference embeddings to cluster assignments while preserving the identity of each window. Using a fixed random seed, we partition the available node-wise ST training windows WST\mathcal{W}_{\mathrm{ST}} into disjoint reference and calibration sets, Wref\mathcal{W}_{\mathrm{ref}} and Wcal\mathcal{W}_{\mathrm{cal}}. The reference set is used to fit the clustering transformation, discover temporal patterns, and select exemplars; the calibration set is reserved for subsequent threshold estimation.

Let Wref={ui}i=1nref\mathcal{W}_{\mathrm{ref}}=\{u_i\}_{i=1}^{n_{\mathrm{ref}}}, where nref=∣Wref∣n_{\mathrm{ref}}=|\mathcal{W}_{\mathrm{ref}}|. Stacking the normalized embeddings z(ui)z(u_i) as rows gives Zref∈Rnref×dZ_{\mathrm{ref}}\in\mathbb{R}^{n_{\mathrm{ref}}\times d}. The row labels uiu_i in Figure 2(a) identify the original windows associated with these embeddings. We fit principal component analysis (PCA) [39] on ZrefZ_{\mathrm{ref}}, with reference mean zˉST=nref−1∑i=1nrefz(ui)∈Rd\bar{z}_{\mathrm{ST}}=n_{\mathrm{ref}}^{-1}\sum_{i=1}^{n_{\mathrm{ref}}}z(u_i)\in\mathbb{R}^{d} and loading matrix PST∈Rd×qP_{\mathrm{ST}}\in\mathbb{R}^{d\times q} containing the leading qq principal directions. The individual and stacked PCA coordinates are

a(u)=PSTT(z(u)−zˉST)∈Rq(18)a(u)=P_{\mathrm{ST}}^{\mathsf{T}}\left(z(u)-\bar{z}_{\mathrm{ST}}\right)\in\mathbb{R}^{q} \tag*{(18)}

where zˉST∈Rd\bar{z}_{\mathrm{ST}}\in\mathbb{R}^{d} is the mean reference embedding and PST∈Rd×qP_{\mathrm{ST}}\in\mathbb{R}^{d\times q} contains the leading qq principal directions. This operation centers each embedding and projects its feature dimension from dd to qq, with qq capped at the effective rank of the centered reference matrix. The resulting clustering coordinates form a matrix Aref∈Rnref×qA_{\mathrm{ref}}\in\mathbb{R}^{n_{\mathrm{ref}}\times q}.

We apply HDBSCAN [40], [41] to the rows of ArefA_{\mathrm{ref}} using Euclidean distance, fixed neighborhood and minimum-cluster-size parameters, and excess-of-mass cluster extraction. It determines the number of clusters KK and assigns each reference window a label

Ck={ui∈Wref:ci=k},k=−1,1,…,K(19)\mathcal{C}_{k}=\{u_i\in\mathcal{W}_{\mathrm{ref}}:c_i=k\},\quad k=-1,1,\ldots,K \tag*{(19)}

@page 5/14 starts_mid=0 ends_mid=1

Temporal Data Selection

As shown in Figure 1(a), a shared pretrained encoder maps ST and TS windows into a common representation space. Temporal pattern discovery and exemplar selection establish an ST exemplar reference, and reference calibration determines the matching threshold. Auxiliary time series candidates are then scored by their proximity to this reference: nearby candidates undergo matched selection, while distant candidates pass through teacher reliability filtering, novel pattern discovery, and diversity selection. The selected embeddings are used to retrieve the corresponding temporal windows by index.

Figure 2 (PDF p. 5, third). ST reference exemplar construction and threshold calibration.

Shared Temporal Representation

We construct encoder pretraining samples from the limited spatio-temporal data DST\mathcal{D}_{\mathrm{ST}} by treating each node’s observations as an independent time series, omitting spatial relations. Let DST′\mathcal{D}'_{\mathrm{ST}} denote a fixed collection of these sampled node-wise windows.

We consider three alternative implementations of the pretrained encoder EψE_{\psi} in Figure 1(a): TS2Vec [36], TS-TCC [37], and CoST [38]. Each encoder is pretrained separately on the same DST′\mathcal{D}'_{\mathrm{ST}}, with the preprocessing described below, using its own self-supervised objective and is then frozen throughout selection. Within each encoder configuration, the same frozen EψE_{\psi} processes both ST and TS windows, placing samples from the two sources in a shared representation space.

For a training window u=(x,y)u=(x,y), where x∈RL×1x\in\mathbb{R}^{L\times1} and y∈RH×1y\in\mathbb{R}^{H\times1} are the history and observed target, we concatenate them into a trajectory [x;y]∈R(L+H)×1[x;y]\in\mathbb{R}^{(L+H)\times1}. Mean aggregation to a common sampling interval produces s(u)∈RT×1s(u)\in\mathbb{R}^{T\times1}, where TT is the aligned sequence length. We then compute

s~(u)=s(u)−μ(s(u))σ(s(u))(20)\widetilde{s}(u)=\frac{s(u)-\mu(s(u))}{\sigma(s(u))} \tag*{(20)}
h(u)=Pool⁡(Eψ(s~(u)))(21)h(u)=\operatorname{Pool}\left(E_{\psi}(\widetilde{s}(u))\right) \tag*{(21)}
z(u)=h(u)∥h(u)∥2(22)z(u)=\frac{h(u)}{\lVert h(u)\rVert_{2}} \tag*{(22)}

where μ(s(u))\mu(s(u)) and σ(s(u))\sigma(s(u)) are the scalar mean and standard deviation of the aggregated trajectory. The encoder produces Eψ(s~(u))∈RTψ×dE_{\psi}(\widetilde{s}(u))\in\mathbb{R}^{T_{\psi}\times d}, where TψT_{\psi} is the encoder’s output sequence length. The operator Pool performs global max pooling, yielding h(u)∈Rdh(u)\in\mathbb{R}^{d}. The final L2L_{2} normalization produces z(u)z(u). These vectors z(u)z(u) form the ST Embeddings or TS Embeddings in Figure 1(a) and are used for subsequent pattern discovery and reference proximity scoring.

ST Exemplar Construction and Threshold Calibration

We construct ST reference exemplars to represent temporal patterns in the limited ST data and calibrate a proximity threshold to support both matched and diversity selection. Following Figure 1(a), this stage comprises temporal pattern discovery, exemplar selection, reference calibration, and reference proximity scoring.

Temporal pattern discovery. As illustrated in Figure 2(a), temporal pattern discovery maps ST reference embeddings to cluster assignments while preserving the identity of each window. Using a fixed random seed, we partition the available node-wise ST training windows WST\mathcal{W}_{\mathrm{ST}} into disjoint reference and calibration sets, Wref\mathcal{W}_{\mathrm{ref}} and Wcal\mathcal{W}_{\mathrm{cal}}. The reference set is used to fit the clustering transformation, discover temporal patterns, and select exemplars; the calibration set is reserved for subsequent threshold estimation.

Let Wref={ui}i=1nref\mathcal{W}_{\mathrm{ref}}=\{u_i\}_{i=1}^{n_{\mathrm{ref}}}, where nref=∣Wref∣n_{\mathrm{ref}}=|\mathcal{W}_{\mathrm{ref}}|. Stacking the normalized embeddings z(ui)z(u_i) as rows gives Zref∈Rnref×dZ_{\mathrm{ref}}\in\mathbb{R}^{n_{\mathrm{ref}}\times d}. The row labels uiu_i in Figure 2(a) identify the original windows associated with these embeddings. We fit principal component analysis (PCA) [39] on ZrefZ_{\mathrm{ref}}, with reference mean zˉST=nref−1∑i=1nrefz(ui)∈Rd\bar{z}_{\mathrm{ST}}=n_{\mathrm{ref}}^{-1}\sum_{i=1}^{n_{\mathrm{ref}}}z(u_i)\in\mathbb{R}^{d} and loading matrix PST∈Rd×qP_{\mathrm{ST}}\in\mathbb{R}^{d\times q} containing the leading qq principal directions. The individual and stacked PCA coordinates are

a(u)=PSTT(z(u)−zˉST)∈Rq(23)a(u)=P_{\mathrm{ST}}^{\mathsf{T}}\left(z(u)-\bar{z}_{\mathrm{ST}}\right)\in\mathbb{R}^{q} \tag*{(23)}

where zˉST∈Rd\bar{z}_{\mathrm{ST}}\in\mathbb{R}^{d} is the mean reference embedding and PST∈Rd×qP_{\mathrm{ST}}\in\mathbb{R}^{d\times q} contains the leading qq principal directions. This operation centers each embedding and projects its feature dimension from dd to qq, with qq capped at the effective rank of the centered reference matrix. The resulting clustering coordinates form a matrix Aref∈Rnref×qA_{\mathrm{ref}}\in\mathbb{R}^{n_{\mathrm{ref}}\times q}.

We apply HDBSCAN [40], [41] to the rows of ArefA_{\mathrm{ref}} using Euclidean distance, fixed neighborhood and minimum-cluster-size parameters, and excess-of-mass cluster extraction. It determines the number of clusters KK and assigns each reference window a label

Ck={ui∈Wref:ci=k},k=−1,1,…,K(24)\mathcal{C}_{k}=\{u_i\in\mathcal{W}_{\mathrm{ref}}:c_i=k\},\quad k=-1,1,\ldots,K \tag*{(24)}

Here, ci=−1c_i=-1 denotes density noise, as shown in the cluster assignments in Figure 2(a), which is excluded from this selection. Each CkC_k contains the original windows assigned to one temporal pattern cluster and supplies candidates for exemplar selection.

Exemplar selection. Reference exemplars are constructed by random sampling stratified across the KK clusters. Selecting at least one exemplar from each of the KK clusters requires an exemplar budget BE≥KB_E \ge K. Using a fixed random seed, we first sample one window uniformly from each cluster. And then expand the exemplar set by drawing additional windows uniformly and without replacement from the clusters in turn, skipping exhausted clusters until the exemplar budget is reached or no candidates remain.

Let Ek⊆Ck\mathcal{E}_k \subseteq C_k contain the exemplars selected from cluster kk. The resulting ST reference exemplar set is

E=⋃k=1KEk\mathcal{E}=\bigcup_{k=1}^{K}\mathcal{E}_k
M=∣E∣=min⁡{BE,∑k=1K∣Ck∣}(25)M=|\mathcal{E}|=\min\left\{B_E,\sum_{k=1}^{K}|C_k|\right\} \tag*{(25)}

Each e∈Ee \in\mathcal{E} is a context–target window with normalized embedding z(e)∈Rdz(e) \in\mathbb{R}^{d}. Stacking these vectors gives the ST reference embeddings in Figure 1(a), a matrix in RM×d\mathbb{R}^{M \times d}. The sampling procedure retains at least one exemplar from each cluster and distributes additional exemplars across the discovered patterns.

Reference calibration. As illustrated in Figure 2(b), reference calibration computes the distance from each calibration window to its nearest ST exemplar, sorts these distances, and determines the reference threshold. We define the proximity of a window uu to the ST reference as

d(u)=min⁡e∈E∥z(u)−z(e)∥2(26)d(u)=\min_{e\in\mathcal{E}}\|z(u)-z(e)\|_2 \tag*{(26)}

Distances are computed using the normalized embeddings zz, while the PCA coordinates aa are used for clustering. We evaluate d(u)d(u) on the calibration windows and set

τ=Q1−α({d(u):u∈Wcal}),0<α<1(27)\tau=Q_{1-\alpha}\left(\{d(u):u\in\mathcal{W}_{\mathrm{cal}}\}\right),\qquad0<\alpha<1 \tag*{(27)}

where ncal=∣Wcal∣n_{\mathrm{cal}}=|\mathcal{W}_{\mathrm{cal}}| and Q1−αQ_{1-\alpha} denotes the empirical (1−α)(1-\alpha)-quantile, returning the ⌈(1−α)ncal⌉\lceil(1-\alpha)n_{\mathrm{cal}}\rceil-th smallest calibration distance. Consequently, at least a fraction 1−α1-\alpha of the calibration windows satisfy d(u)≤τd(u)\le\tau.

Reference proximity scoring. With E\mathcal{E} and τ\tau fixed, we compute d(u)d(u) for every auxiliary candidate u∈DTSu\in\mathcal{D}_{\mathrm{TS}}. These distances correspond to the scores attached to TS embeddings in Figure 1(a). Together with τ\tau, they guide matched selection within the reference neighborhood and distant-candidate screening for diversity selection. Both branches therefore use the same ST reference and proximity measure.

Matched Data Selection

Matched data selection retrieves auxiliary windows close to the ST reference. Using the calibrated threshold, we define the matched pool as

UM={u∈DTS:d(u)≤τ}.(28)\mathcal{U}_{\mathrm{M}}=\{u\in\mathcal{D}_{\mathrm{TS}}:d(u)\le\tau\}. \tag*{(28)}

Each candidate is assigned to the cluster of its closest exemplar. Within each resulting bucket, candidates are ranked by increasing d(u)d(u). We cycle through nonempty buckets in a fixed seeded order, taking the closest remaining eligible candidate from each, until the budget BMB_{\mathrm{M}} is reached. The selected windows form the matched subset DM⊆UM\mathcal{D}_{\mathrm{M}}\subseteq\mathcal{U}_{\mathrm{M}}. This limits domination by a single reference pattern.

Diverse Data Selection

Following Figure 1(a), this branch comprises teacher reliability filtering, novel pattern discovery, and diversity selection. Its input pool contains auxiliary windows outside the expanded ST reference neighborhood:

UD0={u∈DTS:d(u)>τ+δ},δ>0(29)\mathcal{U}_{\mathrm{D}}^{0}=\{u\in\mathcal{D}_{\mathrm{TS}}:d(u)>\tau+\delta\},\qquad\delta>0 \tag*{(29)}

after excluding windows overlapping DM\mathcal{D}_{\mathrm{M}}. The margin δ\delta separates the two branches; windows with τ<d(u)≤τ+δ\tau<d(u)\le\tau+\delta are used by neither.

Teacher reliability filtering. Distance from the ST reference alone does not ensure reliable teacher supervision. For each u=(x,y)∈UD0u=(x,y)\in\mathcal{U}_{\mathrm{D}}^{0}, the frozen teacher predicts y^T=gϕ(x)\hat{y}^{T}=g_{\phi}(x), and we measure its forecasting error:

ℓT(u)=H−1∑h=1H∣y^hT−yh∣max⁡{σ(x),ϵs}(30)\ell_T(u)=\frac{H^{-1}\sum_{h=1}^{H}\left|\hat{y}^{T}_{h}-y_h\right|}{\max\{\sigma(x),\epsilon_s\}} \tag*{(30)}

where y^hT\hat{y}^{T}_{h} and yhy_h denote the hh-th scalar entries of y^T\hat{y}^{T} and yy, respectively. Here, σ(x)\sigma(x) is the history standard deviation and ϵs>0\epsilon_s>0 is a scale floor for source ss. As in reference calibration, we set an acceptance threshold using the ST calibration windows:

η=Q1−β({ℓT(u):u∈Wcal})\eta=Q_{1-\beta}\left(\{\ell_T(u):u\in\mathcal{W}_{\mathrm{cal}}\}\right)
UD1={u∈UD0:ℓT(u)≤η}(31)\mathcal{U}_{\mathrm{D}}^{1}=\{u\in\mathcal{U}_{\mathrm{D}}^{0}:\ell_T(u)\le\eta\} \tag*{(31)}

Here, 0<β<10<\beta<1. This retains candidates whose teacher error lies within the calibrated tolerance.

Novel pattern discovery. After teacher reliability filtering, we fit a separate PCA transformation, again capped at the effective rank, on {z(u):u∈UD1}\{z(u):u\in\mathcal{U}_{\mathrm{D}}^{1}\} and apply HDBSCAN with specified minimum cluster size and neighborhood parameters to obtain {Gj}j=1J\{\mathcal{G}_j\}_{j=1}^{J}. Refitting PCA here allows clustering to retain variation specific to the distant pool. We exclude density-noise points and retain members with strength at least p0p_0. Consequently, novelty means a coherent candidate pattern outside the calibrated ST reference neighborhood.

Diversity selection. As in matched selection, we visit nonempty cluster pools in a fixed seeded round-robin order and apply the same overlap exclusion and tie-breaking rules. On a cluster’s first visit, we select its eligible medoid, minimizing the sum of distances to the other eligible members. On subsequent visits, we choose the eligible window whose minimum distance to the already selected windows in that cluster is largest. All distances are Euclidean distances between the normalized embeddings zz. Selection stops at budget BDB_{\mathrm{D}} or candidate exhaustion. The selected windows form the diverse subset DD⊆UD1\mathcal{D}_{\mathrm{D}}\subseteq\mathcal{U}_{\mathrm{D}}^{1}, with ∣DD∣≤BD|\mathcal{D}_{\mathrm{D}}|\le B_{\mathrm{D}}. Together, DM\mathcal{D}_{\mathrm{M}} and DD\mathcal{D}_{\mathrm{D}} supplement DST\mathcal{D}_{\mathrm{ST}} during joint training. All selection operations are offline and are summarized in Algorithm 1.

Algorithm 1 Offline Temporal Data Selection in T2ST
Require: Training pools $\mathcal{D}_{\mathrm{ST}}, \mathcal{D}_{\mathrm{TS}}$; encoder $E_{\psi}$; frozen teacher $g_{\phi}$; budgets $B_E, B_M, B_D$; selection settings
Ensure: Matched subset $\mathcal{D}_M$ and diverse subset $\mathcal{D}_D$
Records are identified by dataset, series row, and window start.
Use a fixed seed for splits, ties, and round-robin cluster visits.
1: Construct $\mathcal{W}_{\mathrm{ST}}$; pretrain $E_{\psi}$ on $\mathcal{D}_{\mathrm{ST}}$ and freeze it
2: Embed $\mathcal{W}_{\mathrm{ST}} \cup \mathcal{D}_{\mathrm{TS}}$ using (5)–(7)
3: Split $\mathcal{W}_{\mathrm{ST}}$ into $\mathcal{W}_{\mathrm{ref}}$ and $\mathcal{W}_{\mathrm{cal}}$
4: Fit PCA–HDBSCAN to build $\mathcal{E}$ by (10)
5: Compute $d,\tau$ by (11)–(12); initialize $\mathcal{D}_M,\mathcal{D}_D \leftarrow \varnothing$
6: Form $\mathcal{U}_M$ by (13); group by nearest-exemplar cluster into buckets $\mathcal{B}_k$, sorted by increasing $d(u)$
7: while $|\mathcal{D}_M| < B_M$ and $\bigcup_k \mathcal{B}_k \ne \varnothing$ do
8:   Visit the next $\mathcal{B}_k$; add its closest remaining record $u$ to $\mathcal{D}_M$
9:   Remove the selected record identity from all $\mathcal{B}_i$
10: end while
11: Form $\mathcal{U}_D^0$ by (14), excluding identities already in $\mathcal{D}_M$
12: Compute $\ell_T$, calibrate $\eta$, and form $\mathcal{U}_D^1$ by (15)–(16)
13: if $\mathcal{U}_D^1$ cannot support clustering, return $(\mathcal{D}_M,\varnothing)$
14: Refit PCA–HDBSCAN on the reliable embeddings to obtain nonnoise clusters $\mathcal{G}_j$
15: Retain members with strength $\ge p_0$ in $\mathcal{Q}_j \subseteq \mathcal{G}_j$;
16: while $|\mathcal{D}_D| < B_D$ and $\bigcup_j \mathcal{Q}_j \ne \varnothing$ do
17:   Visit the next nonempty $\mathcal{Q}_j$; choose $u$ as its medoid if $\mathcal{S}_j=\varnothing$, otherwise maximize its minimum distance to $\mathcal{S}_j$
18:   Add $u$ to $\mathcal{S}_j$ and $\mathcal{D}_D$; remove its record identity from all $\mathcal{Q}_i$
19: end while
20: return $\mathcal{D}_M,\mathcal{D}_D$

Knowledge Distillation

After temporal data selection, T2ST trains an STFM student using the limited spatio-temporal dataset DST\mathcal{D}_{\mathrm{ST}} and the selected auxiliary subsets DM\mathcal{D}_M and DD\mathcal{D}_D. A frozen TSFM provides two complementary forms of supervision: relationships among historical temporal patches and predictions at multiple temporal resolutions. The former transfers the teacher’s organization of temporal context, while the latter transfers its forecasting behavior. We first describe the teacher and the student, followed by the knowledge transfer objectives.

Frozen TSFM Teacher

Let gϕg_{\phi} denote a pretrained TSFM with fixed parameters ϕ\phi. We describe the scalar forecasting setting used in our implementation, with X∈RL×N×1X \in\mathbb{R}^{L \times N \times1} and Y∈RH×N×1Y \in\mathbb{R}^{H \times N \times1}. For a spatio-temporal sample, the teacher processes each node history xn=X:,n,1x_n = X_{:,n,1} independently:

y^nT=gϕ(xn),n=1,…,N(32)\hat{y}^{T}_{n} = g_{\phi}(x_n), \qquad n=1,\ldots,N \tag*{(32)}

Stacking these outputs gives Y^T∈RH×N×1\hat{Y}^{T} \in\mathbb{R}^{H \times N \times1}. An auxiliary temporal window is processed in the same way with N=1N=1. Thus, the teacher supplies temporal supervision for both data types without requiring spatial structure.

In addition to forecasts, we extract hidden representations Hℓ,nT∈RPT×dTH^{T}_{\ell,n} \in\mathbb{R}^{P_T \times d_T} from selected teacher layers, where PTP_T is the number of history patches and dTd_T is the hidden dimension. We instantiate the teacher from a local copy of the official Google Research TimesFM-2.5 checkpoint and extract representations from three layers. We perform no additional teacher pretraining or fine-tuning. All teacher parameters remain frozen during offline reliability filtering and student distillation.

STFM Student with Shared Temporal Modeling

The student fθf_{\theta} retains the temporal and spatial modeling components of its STFM backbone. It supports two input-dependent computation paths:

Y^S={fθST(X,G),(X,Y)∈DSTfθT(X),(X,Y)∈DM∪DD(33)\hat{Y}^{S} = \begin{cases} f_{\theta}^{\mathrm{ST}}(X,\mathcal{G}), & (X,Y) \in\mathcal{D}_{\mathrm{ST}} \\ f_{\theta}^{T}(X), & (X,Y) \in\mathcal{D}_{M} \cup\mathcal{D}_{D} \end{cases} \tag*{(33)}

Here, fθSTf_{\theta}^{\mathrm{ST}} and fθTf_{\theta}^{T} denote two execution paths of the same student, sharing temporal modeling and forecasting parameters. The ST path uses the available spatial structure G\mathcal{G}. The temporal-only path bypasses cross-node interactions and omits graph-dependent spatial embeddings. Auxiliary training therefore updates the shared temporal pathway, while ST samples provide supervision for spatial modeling.

For relation distillation, the student exposes temporal representations Hi,nS∈RPS×dSH^{S}_{i,n} \in\mathbb{R}^{P_S \times d_S} at selected layers. In the OpenCity instantiation, these features are taken after temporal modeling and before graph convolution within each block. On ST inputs, they can retain spatial information from context embeddings and preceding blocks. The formulation permits another STFM backbone, such as UrbanFM, provided that its adapter exposes temporally ordered historical features, enables temporal-only processing with shared parameters, and produces forecasts for the same target horizon.

Knowledge Transfer Objectives

We transfer temporal knowledge from the frozen TSFM teacher to the STFM student through two complementary objectives: temporal relation distillation and multi-scale prediction distillation. The former preserves relationships among intermediate temporal patches, while the latter aligns teacher and student forecasts at multiple temporal resolutions, providing supervision at both the representation and prediction levels.

Temporal relation distillation. Directly matching hidden features is restrictive because the teacher and student can differ in both temporal resolution and representation dimension. Instead, we align pairwise patch similarities separately for each sample and node. Let P\mathcal{P} be the set of selected student–teacher layer pairs (i,ℓ)(i,\ell). Per pair, adaptive average pooling (AAP) aligns the student’s patch axis to the teacher’s patch count:

H~i,nS=AAP⁡PT(Hi,nS)∈RPT×dS(34)\widetilde{H}^{S}_{i,n} = \operatorname{AAP}_{P_T}(H^{S}_{i,n}) \in\mathbb{R}^{P_T \times d_S} \tag*{(34)}

The operator is the identity when PS=PTP_S=P_T. It acts only along the temporal patch axis, leaving the student’s feature dimension unchanged.

Let N(H)\mathcal{N}(H) normalize each row hh as h/max⁡(∥h∥2,ϵ)h/\max(\lVert h\rVert_2,\epsilon), with ϵ>0\epsilon>0. The temporal relation matrices are

Ri,nS=N(H~i,nS)N(H~i,nS)TR^{S}_{i,n} = \mathcal{N}(\widetilde{H}^{S}_{i,n})\mathcal{N}(\widetilde{H}^{S}_{i,n})^{\mathsf{T}}
Rℓ,nT=N(Hℓ,nT)N(Hℓ,nT)T(35)R^{T}_{\ell,n} = \mathcal{N}(H^{T}_{\ell,n})\mathcal{N}(H^{T}_{\ell,n})^{\mathsf{T}} \tag*{(35)}

Both matrices belong to RPT×PT\mathbb{R}^{P_T \times P_T}, and each entry measures the similarity of two temporal patches within the same node history. For a mini-batch of BB samples, the relation loss is

Lrel=1∣P∣BNPT2∑(i,ℓ)∈P∑b=1B∑n=1N∥Ri,b,nS−Rℓ,b,nT∥F2(36)\mathcal{L}_{\mathrm{rel}} = \frac{1}{|\mathcal{P}|BNP_T^2} \sum_{(i,\ell)\in\mathcal{P}} \sum_{b=1}^{B} \sum_{n=1}^{N} \left\lVert R^{S}_{i,b,n} - R^{T}_{\ell,b,n} \right\rVert_F^2 \tag*{(36)}

For the three-block OpenCity configuration, we use P={(1,6),(2,13),(3,20)}\mathcal{P} = \{(1, 6), (2, 13), (3, 20)\}.

Multi-scale prediction distillation. Relation alignment constrains historical representations but does not directly constrain future forecasts. We therefore align teacher and student predictions at multiple temporal resolutions.

For an aggregation factor rr, define nonoverlapping temporal averaging as

[Ar(Y^)]b,k,n,1=1r∑j=1r[Y^]b,(k−1)r+j,n,1(37)[\mathcal{A}_r(\widehat{Y})]_{b,k,n,1} = \frac{1}{r}\sum_{j=1}^{r}[\widehat{Y}]_{b,(k-1)r+j,n,1} \tag*{(37)}

where k=1,…,H/rk = 1, \ldots, H/r. We use R={1,2,4,8}\mathcal{R} = \{1, 2, 4, 8\} and forecasting horizons divisible by every factor. The loss is

Lms=1∣R∣∑r∈RMAE⁡(Ar(Y^S),Ar(Y^T))(38)\mathcal{L}_{\mathrm{ms}} = \frac{1}{|\mathcal{R}|}\sum_{r\in\mathcal{R}}\operatorname{MAE}\left(\mathcal{A}_r(\widehat{Y}^{S}), \mathcal{A}_r(\widehat{Y}^{T})\right) \tag*{(38)}

The r=1r = 1 term transfers step-level forecasting behavior, while coarser averages constrain broader temporal variations. Each resolution contributes equally after averaging its prediction errors.

Joint training and inference. Both distillation losses are applied to ST, matched, and diversity batches. For a batch B\mathcal{B}, we combine them with ground-truth forecasting supervision:

L(B)=Lsup(B)+λrelLrel(B)+λmsLms(B)(39)\mathcal{L}(\mathcal{B}) = \mathcal{L}_{\mathrm{sup}}(\mathcal{B}) + \lambda_{\mathrm{rel}}\mathcal{L}_{\mathrm{rel}}(\mathcal{B}) + \lambda_{\mathrm{ms}}\mathcal{L}_{\mathrm{ms}}(\mathcal{B}) \tag*{(39)}

where Lsup\mathcal{L}_{\mathrm{sup}} is the student’s supervised forecasting loss and λrel,λms\lambda_{\mathrm{rel}}, \lambda_{\mathrm{ms}} balance the two transfer objectives.

Mixed training is organized around ST batches. Let IM,tI_{M,t} and ID,tI_{D,t} indicate whether matched and diverse data are scheduled at training event tt, respectively, with ID,t≤IM,tI_{D,t} \le I_{M,t}. With combined updates, the event objective is

Jt=L(BST,t)+IM,tγ~M∑j=1KML(BM,t,j)+ID,tγ~D,tL(BD,t)(40)\begin{aligned} J_t &= \mathcal{L}(\mathcal{B}_{\mathrm{ST},t}) \\ &\quad+ I_{M,t}\widetilde{\gamma}_M\sum_{j=1}^{K_M}\mathcal{L}(\mathcal{B}_{M,t,j}) \\ &\quad+ I_{D,t}\widetilde{\gamma}_{D,t}\mathcal{L}(\mathcal{B}_{D,t}) \tag*{(40)} \end{aligned}

Here, KMK_M is the number of matched batches per scheduled event and γM\gamma_M is their per-batch weight. To control the diversity contribution, let At=γM∑jL(BM,t,j)A_t = \gamma_M\sum_j\mathcal{L}(\mathcal{B}_{M,t,j}) and Dt=L(BD,t)D_t = \mathcal{L}(\mathcal{B}_{D,t}). On diversity events, its effective weight is

γ~D,t=γDmin⁡{1,ρAtγDDt+ϵ}(41)\widetilde{\gamma}_{D,t} = \gamma_D\min\left\{1,\frac{\rho A_t}{\gamma_DD_t+\epsilon}\right\} \tag*{(41)}

where 0<ρ<10 < \rho< 1 caps the weighted diversity loss relative to the weighted matched loss, and 0<γD<γM0 < \gamma_D < \gamma_M. This coefficient is treated as constant during backpropagation. Sampling intervals and loss weights therefore control auxiliary exposure and its contribution to training. Ground-truth supervision anchors the student to observed outcomes while the frozen teacher supplies additional temporal guidance.

Experiment

We evaluate forecasting with limited structured observations through five research questions:

  • RQ1 (Overall Effectiveness): How does T2ST compare with its STFM backbones trained on the same limited ST subset or the full ST training set?

  • RQ2 (Temporal Data Selection): How does the composition of auxiliary temporal data and the choice of selection encoder affect T2ST’s forecasting performance?

  • RQ3 (Batch Mixing Ratio): How does the ratio of ST to matched temporal batches affect forecasting?

  • RQ4 (ST Data Budget): How does T2ST’s forecasting performance vary across different ST data budgets?

  • RQ5 (Distillation Weight Sensitivity): How sensitive is T2ST to the loss weights?

Experiment Setup

Evaluation Settings and Data Budgets

1) Evaluation Settings and Data Budgets: We consider two evaluation settings, summarized in Table II. The OpenCity setting [7] evaluates long-term forecasting on CAD3, CAD5, PEMS07M, TrafficSH, CHI-TAXI, and NYC-BIKE. It uses one day of observations to predict the next day, corresponding to 288 steps for five-minute data. The UrbanFM setting [9] uses a 12-step history and a 12-step horizon on the five-minute data packages of EvalST, whose 12 datasets cover flow, occupancy, speed, and grid-based urban observations.

DatasetCategoryData format# RegionsInterval
(a) Short-term: L=12L = 12, H=12H = 12
PEMS-03FlowSensor-based3585 min
PEMS-04FlowSensor-based3075 min
PEMS-07FlowSensor-based8835 min
PEMS-08FlowSensor-based1705 min
OCC-PairsOccupancySensor-based1065 min
OCC-HamburgOccupancySensor-based2405 min
PEMS-BAYSpeedSensor-based3255 min
METR-LASpeedSensor-based2075 min
Traffic-SHRoadGrid-based8965 min
Bike-NYCBikeGrid-based2005 min
Taxi-NYCTaxiGrid-based2005 min
Tdrive-BJTrajectoryGrid-based8965 min
(b) Long-term: one-day history and horizon
CAD3Traffic flowSensor-based4805 min
CAD5Traffic flowSensor-based2115 min
CHI-TAXITaxi demandGrid-based7730 min
NYC-BIKEBike trajectoryGrid-based54030 min
TrafficSHTraffic speedGrid-based89630 min
PEMS07MTraffic speedSensor-based2285 min

Table II. Evaluation datasets for short-term forecasting and long-term forecasting.

In pre-training, long-term ST data follow OpenCity, and short-term ST data follow UrbanFM’s WorldST collection. The 10% ST budget is measured against eligible samples on the complete source timelines, with all selected samples drawn from the training splits. T2ST supplements this restricted subset DST\mathcal{D}_{\mathrm{ST}} with matched and diverse temporal windows from the GiftEvalPretrain corpus [42].

Baselines and Model Configurations

2) Baselines and Model Configurations: In both settings, we apply T2ST as a training framework to two STFM backbones, OpenCity and UrbanFM. For each backbone, we compare the T2ST-trained model with its counterparts trained solely on the restricted or the full ST data. Long-term baselines also include ten full-shot task-specific predictors, such as PDFormer [43], AGCRN [44], MTGNN [45], and GWN [46]. Short-term baselines include TimesFM++, Moirai, TimeMoE [47], Chronos, and FactoST [21].

T2ST uses a frozen TimesFM 2.5 teacher and TS2Vec [36] for offline temporal-data selection. Before distillation, the STFM student is initialized from a checkpoint trained on the restricted 10% ST subset. The student is then distilled using temporal-relation and multi-scale prediction losses. The encoder study replaces TS2Vec with TS-TCC [37] or CoST [38].

Evaluation Metrics

We report mean absolute error (MAE) and root mean squared error (RMSE) for all datasets. For short-term forecasting on grid-based datasets, we additionally report mean absolute percentage error (MAPE).

Overall Forecasting Performance (RQ1)

Long-Term Forecasting

Table III reports long-term forecasting results on six datasets, comparing T2ST with full-shot task-specific model and the corresponding backbone models trained with 10% or full ST data.

ModelCAD3CAD5PEMS07MTrafficSHCHI-TAXINYC-BIKE
MAERMSEMAERMSEMAERMSEMAERMSEMAERMSEMAERMSE
PDFormer20.2836.4312.8921.184.628.360.771.234.0312.827.3313.01
STWA21.6537.5514.4324.144.548.571.422.493.7011.497.9714.35
MSDR22.7437.1515.4429.524.818.611.843.553.5510.398.0813.97
AGCRN18.7231.9312.8823.784.618.631.332.453.6011.318.1314.46
MTGNN17.5928.9211.7020.304.528.060.811.263.279.876.4811.49
STSGCN21.8834.5213.8722.324.568.051.663.334.8714.406.8511.98
GWN16.9428.8110.6919.75<u>4.17</u><u>7.84</u>0.761.393.5611.277.6113.56
ASTGNN23.6039.3512.5821.234.398.210.691.093.2810.326.4411.62
STGCN20.2434.3413.7627.494.448.371.603.143.099.548.0113.94
TGCN19.5630.8213.0721.564.888.381.792.654.0211.707.7513.49
UrbanFM(10%)20.8636.5313.4822.545.4810.500.660.932.686.357.6814.53
UrbanFM20.2535.8812.2921.075.139.930.550.842.857.857.3313.93
T2ST (UrbanFM)19.6434.5412.3020.994.889.310.550.882.787.477.5114.28
OpenCity (10%)17.9731.6211.7320.334.598.32<u>0.58</u>0.89<u>2.29</u>5.716.2011.24
OpenCity<u>16.53</u><u>28.34</u>11.30<u>19.42</u>4.538.300.550.851.994.646.1311.22
T2ST (OpenCity)16.2527.72<u>10.74</u>18.514.167.810.55<u>0.85</u>2.34<u>5.41</u>6.0510.92

Table III. Long-term forecasting results. The first ten methods are full-shot baselines. Best and second-best results are shown in bold and underlined, respectively, with light purple cell backgrounds.

As shown in Table III, T2ST-trained OpenCity matches or outperforms full-data OpenCity in both MAE and RMSE on five of six datasets while using the 10% ST budget. On PEMS07M, MAE decreases from 4.53 to 4.16 and RMSE from 8.30 to 7.81, corresponding to reductions of 8.17% and 5.90%, respectively. Both metrics also improve on CAD3, CAD5, and NYC-BIKE.

Under the same restricted ST budget, T2ST-trained UrbanFM reduces MAE on three datasets and RMSE on four compared with full-data UrbanFM. On PEMS07M, MAE decreases from 5.13 to 4.88 and RMSE from 9.93 to 9.31, yielding reductions of 4.87% and 6.24%, respectively. Its performance broadly comparable to that of the full-data model.

Short-Term Forecasting

Table IV reports results on four grid-based datasets: TrafficSH, NYC-BIKE, NYC-TAXI, and Tdrive-BJ. With OpenCity as the backbone, the T2ST framework improves all three metrics over OpenCity across these datasets. Compared with full-data OpenCity, it achieves lower errors in nine of the twelve dataset–metric entries while using the 10% ST budget. On TrafficSH, MAE decreases from 0.44 to 0.41 and RMSE from 0.97 to 0.88, corresponding to reductions of 6.82% and 9.28%, respectively.

ModelTrafficSHNYC-BIKENYC-TAXITdrive-BJ
MAERMSEMAPEMAERMSEMAPEMAERMSEMAPEMAERMSEMAPE
TimesFM++0.491.147.313.987.3765.528.9520.0449.5013.9030.4818.26
Moirai0.460.918.761.635.8785.187.5822.7263.3321.7352.1631.99
TimeMoE0.43<u>0.74</u><u>6.65</u>3.908.0864.2010.0523.5989.8521.4036.6730.76
Chronos0.450.806.774.087.6457.629.4821.69<u>43.22</u>16.9135.0120.79
FactoST0.400.705.843.777.1154.368.8421.6842.3317.2635.2320.61
OpenCity (10%)0.430.937.571.264.5062.155.4917.3250.0712.6537.5718.36
OpenCity0.440.977.651.294.6765.255.4517.0452.0111.9933.8717.69
T2ST (OpenCity)<u>0.41</u>0.887.111.234.4360.695.2816.7747.3712.2535.49<u>17.92</u>
UrbanFM (10%)0.500.957.791.194.11<u>56.45</u>5.5117.6254.63<u>11.37</u><u>29.16</u>25.18
UrbanFM0.470.907.54<u>1.20</u>4.1157.295.4317.3355.3612.3830.2633.92
T2ST (UrbanFM)0.470.947.671.19<u>4.24</u>58.95<u>5.34</u><u>16.89</u>50.7810.5328.8619.09

Table IV. Short-term forecasting results. Best and second-best results are shown in bold and underlined, respectively, with light purple cell backgrounds.

With UrbanFM as the backbone, the T2ST framework matches or outperforms full-data UrbanFM in MAE on all four grid-based datasets under the same ST budget. On Tdrive-BJ, MAE and RMSE decrease by 14.94% and 4.63%, respectively, while MAPE falls from 33.92% to 19.09%.

On sensor-based datasets (Table VI), the T2ST framework achieves broadly comparable performance to the full-data models for both backbones, with lower errors in more than half of the comparisons, while using the 10% ST budget.

DatasetUrbanFM (Full)T2ST (UrbanFM)OpenCity (Full)T2ST (OpenCity)
MAERMSEMAERMSEMAERMSEMAERMSE
OCC-Pairs1.352.951.292.921.282.901.373.05
OCC-Hamburg6.2310.386.2010.416.1910.716.1110.45
PEMS-Bay2.495.632.325.312.586.072.365.75
METR-LA6.1713.336.0513.156.4114.295.8613.67
PEMS-0321.5334.3321.4034.5021.6135.3922.2534.93
PEMS-0427.0841.9827.5042.6728.3145.0229.2545.15
PEMS-0731.5247.9331.9548.5332.4951.4532.9750.46
PEMS-0822.4734.9722.7035.5323.5638.6124.1338.16

Table VI. Forecasting performance of UrbanFM, OpenCity, and their T2ST variants. Full denotes the full ST training set; T2ST variants use the 10% ST budget.

Effect of Temporal Data Selection (RQ2)

Matched, Diverse, and Random Temporal Data

Table V compares four distillation settings under a fixed 10% ST budget: using only ST data, adding randomly sampled temporal data DM\mathcal{D}_{M} and DD\mathcal{D}_{D}, adding matched data DM\mathcal{D}_{M}, and adding both matched and diverse data DM\mathcal{D}_{M} and DD\mathcal{D}_{D}. Matched selection yields lower MAE and RMSE than equal-budget random sampling on all six datasets, supporting the selection of temporal patterns relevant to ST observations.

SettingCAD3CAD5PEMS07MTrafficSHCHI-TAXINYC-BIKE
MAERMSEMAERMSEMAERMSEMAERMSEMAERMSEMAERMSE
w/o DM\mathcal{D}_M and DD\mathcal{D}_D17.0728.8511.0518.964.227.840.560.852.335.356.0910.90
w/o DD\mathcal{D}_D16.5228.0510.8118.584.197.790.560.862.345.336.0510.88
w/ random DM\mathcal{D}_M and DD\mathcal{D}_D16.8128.0611.1719.034.267.890.590.892.445.576.4711.36
w/ DM\mathcal{D}_M and DD\mathcal{D}_D16.2527.7210.7418.514.167.810.550.852.345.416.0510.92

Table V. Ablation of temporal-data composition for distillation with a fixed 10% ST-data budget. The random setting replaces both with randomly sampled temporal data. Bold and underlined values indicate the best and second-best results, respectively.

Adding diverse data DD\mathcal{D}_{D} further reduces MAE on four datasets and matches the matched-only result on the other two. The combined setting achieves the best or tied-best MAE on five of six datasets and outperforms random sampling in every reported metric. On CAD3, MAE and RMSE reach 16.25 and 27.72, representing reductions of 4.80% and 3.92%, respectively, relative to distillation using only ST data.

Choice of Selection Encoder

Table VII replaces the frozen encoder used to construct temporal-data embeddings with TS2Vec, TS-TCC, or CoST. This changes the representation used for retrieval and clustering.

DatasetTS2VecTS-TCCCoST
MAERMSEMAERMSEMAERMSE
CAD316.2527.7216.5727.9416.4728.13
CAD510.7418.5110.9918.9310.8218.68
PEMS07M4.167.814.297.954.177.73
TrafficSH0.550.850.560.860.560.86
CHI-TAXI2.345.412.325.332.315.23
NYC-BIKE6.0510.926.1210.976.0710.90

Table VII. Effect of different temporal encoders.

TS2Vec yields the lowest MAE on five of six datasets and the lowest RMSE on CAD3, CAD5, and TrafficSH. CoST performs best on CHI-TAXI for both metrics and gives the lowest RMSE on PEMS07M and NYC-BIKE. Across all six datasets, all three encoder configurations match or outperform the OpenCity baseline trained on the full training set. The gains are therefore observed with multiple selection encoders, although the most effective representation depends on the target dataset and metric.

Visualization of Temporal Patterns

Figure 3 visualizes TS2Vec embeddings of temporal windows from the 10% ST subset and the selected diverse data DD\mathcal{D}_{D}. Each point represents a temporal window, with colors indicating the five largest HDBSCAN clusters in each panel. The clusters highlight recurring patterns in the ST observations and coherent temporal structure in the diverse data, providing a qualitative complement to the ablation results in Table V.

Figure 3. Five largest temporal clusters in (a) the 10% ST subset and (b) the diverse data DD\mathcal{D}_{D}.

Effect of the Batch Mixing Ratio (RQ3)

Figure 4 varies the scheduling ratio between ST and matched temporal batches, where a:ba:b schedules bb matched batches for every aa ST batches; ST-only distillation (1 : 0) provides the reference. Among the tested mixed ratios from 4 : 1 to 1 : 2, 2 : 1 achieves the lowest MAE on CAD3, CAD5, PEMS07M, and TrafficSH. CHI-TAXI attains its lowest MAE at both 3 : 1 and 1 : 1, while NYC-BIKE favors 1 : 1, indicating dataset-dependent scheduling preferences. We use 2 : 1 as the common setting in the main experiments.

Figure 4. MAE under different ratios of ST to matched temporal batches. Horizontal dashed lines indicate the baseline for distillation using only ST data.

Effect of the ST Data Budget (RQ4)

Figure 5 reports the performance of T2ST (OpenCity) after auxiliary time-series augmentation and knowledge distillation under different ST-data budgets. For each budget of 10%, 30%, or 60%, we supplement the corresponding ST subset with auxiliary time-series data and distill knowledge from a frozen TSFM into the STFM student using both data types.

Figure 5. Forecasting performance of T2ST (OpenCity) with 10%, 30%, and 60% ST-data budgets, each supplemented with auxiliary time-series data for knowledge distillation. The 60% budget denotes the full training split.

These percentages refer only to the ST-data budget, with 60% corresponding to the complete training split. Under this training procedure, increasing the ST budget does not consistently improve forecasting accuracy. Compared with the 30% configuration, the 10% configuration achieves lower MAE and RMSE on CAD5, TrafficSH, and NYC-BIKE.

The comparison with the full 60% training budget further supports the data efficiency of the 10% setting. TrafficSH and NYC-BIKE attain their lowest MAE and RMSE at 10%, while PEMS07M also performs better at 10% than at 60%. On CAD5, increasing the structured-data budget sixfold reduces MAE from 10.74 to 10.55 and RMSE from 18.51 to 18.19, corresponding to improvements of only 1.77% and 1.73%, respectively. Larger budgets benefit CAD3 and CHI-TAXI more consistently, but their MAE reductions from 10% to 60% remain approximately 2.22% and 4.70%. These results indicate that, with auxiliary time-series data and pretrained temporal knowledge incorporated through distillation, additional ST observations provide limited and dataset-dependent marginal gains. The 10% configuration therefore offers a favorable accuracy–data-efficiency trade-off, using one-sixth of the full structured-data budget while achieving comparable or better accuracy on several datasets.

Effect of Distillation Loss Weights (RQ5)

We examine how the weights of temporal-relation distillation λrel\lambda_{\mathrm{rel}} and multi-scale prediction distillation λms\lambda_{\mathrm{ms}} in Eq. (39) affect the STFM under the 10% ST budget.

Figure 6 shows that, MAE on CAD3 first decreases and then increases as either distillation weight grows. With λms\lambda_{\mathrm{ms}} fixed at 0.5, the lowest MAE is 16.15 at λrel=5\lambda_{\mathrm{rel}} = 5. With λrel\lambda_{\mathrm{rel}} fixed at 3, the lowest MAE is 16.25 at λms=0.5\lambda_{\mathrm{ms}} = 0.5.

Figure 6. Sensitivity to distillation loss weights on CAD3.

Conclusion

We presented T2ST, a data-efficient framework that combines selected auxiliary time series and knowledge from a frozen time-series foundation model to reduce STFM dependence on structured observations. Matched sequences reinforce observed patterns, while diverse sequences broaden temporal coverage through teacher reliability filtering. Temporal relation and multi-scale prediction distillation enhance the student’s shared temporal pathway, while real spatio-temporal samples provide spatial supervision. Experiments with STFM on six long-term and twelve short-term datasets demonstrate competitive forecasting performance under the limited ST-data, matching or exceeding full-data counterparts on multiple datasets. These results support temporal-data selection and knowledge transfer as a practical approach to data-efficient STFM training.

References

References

  1. [1]Y. Li, R. Yu, C. Shahabi, and Y. Liu, “Diffusion convolutional recurrent neural network: Data-driven traffic forecasting,” arXiv preprint arXiv:1707.01926, 2017.
  2. [2]M. Ma, J. Hu, C. S. Jensen, F. Teng, P. Han, Z. Xu, and T. Li, “Learning time-aware graph structures for spatially correlated time series forecasting,” in 2024 IEEE 40th international conference on data engineering (ICDE). IEEE, 2024, pp. 4435–4448.arxiv.org/abs/2312.16403
  3. [3]Y. Fang, Y. Qin, H. Luo, F. Zhao, B. Xu, L. Zeng, and C. Wang, “When spatio-temporal meet wavelets: Disentangled traffic forecasting via efficient spectral graph attention networks,” in 2023 IEEE 39th international conference on data engineering (ICDE). IEEE, 2023, pp. 517–529.
  4. [4]J. Han, H. Wang, H. Xiong, and H. Liu, “Scalable pre-training of compact urban spatio-temporal predictive models on large-scale multi-domain data,” Proc. VLDB Endow., vol. 18, no. 7, pp. 2149–2158, 2025.DOI
  5. [5]J. Zhou, X. Lu, Y. Xiao, J. Su, J. Lyu, Y. Ma, and D. Dou, “Sdwpf: A dataset for spatial dynamic wind power forecasting challenge at kdd cup 2022,” arXiv preprint arXiv:2208.04360, 2022.DOI
  6. [6]Y. Zheng, F. Liu, and H.-P. Hsieh, “U-air: When urban air quality inference meets big data,” in Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining, 2013, pp. 1436–1444.
  7. [7]Z. Li, L. Xia, L. Shi, Y. Xu, D. Yin, and C. Huang, “Open spatio-temporal foundation models for traffic prediction,” ACM Transactions on Intelligent Systems and Technology, 2025.DOI
  8. [8]Y. Yuan, J. Ding, J. Feng, D. Jin, and Y. Li, “Unist: A prompt-empowered universal model for urban spatio-temporal prediction,” in Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, 2024, pp. 4095–4106.DOI
  9. [9]W. Chen, Y. Wu, J. Chen, X. Zhou, and Y. Liang, “Urbanfm: Scaling urban spatio-temporal foundation models,” in Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, 2026, pp. 10 635–10 645.
  10. [10]Z. Lai, D. Zhang, H. Li, C. S. Jensen, H. Lu, and Y. Zhao, “Lightstics: A lightweight framework for correlated time series forecasting,” Proceedings of the ACM on Management of Data, vol. 1, no. 2, pp. 1–26, 2023.
  11. [11]M. Goswami, K. Szafer, A. Choudhry, Y. Cai, S. Li, and A. Dubrawski, “MOMENT: A family of open time-series foundation models,” in Proceedings of the 41st International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, R. Salakhutdinov, Z. Kolter, K. Heller, A. Weller, N. Oliver, J. Scarlett, and F. Berkenkamp, Eds., vol. 235. PMLR, 21–27 Jul 2024, pp. 16 115–16 152. [Online]. Available: https://proceedings.mlr.press/v235/goswami24a.htmlarxiv.org/abs/2402.03885
  12. [12]Y. Liu, H. Zhang, C. Li, X. Huang, J. Wang, and M. Long, “Timer: Generative pre-trained transformers are large time series models,” in Proceedings of the 41st International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, R. Salakhutdinov, Z. Kolter, K. Heller, A. Weller, N. Oliver, J. Scarlett, and F. Berkenkamp, Eds., vol. 235. PMLR, 21–27 Jul 2024, pp. 32 369–32 399. [Online]. Available: https://proceedings.mlr.press/v235/liu24bc.htmlarxiv.org/abs/2402.02368
  13. [13]G. Woo, C. Liu, A. Kumar, C. Xiong, S. Savarese, and D. Sahoo, “Unified training of universal time series forecasting transformers,” arXiv preprint arXiv:2402.02592, 2024.arxiv.org/abs/2402.02592
  14. [14]A. F. Ansari, O. Schur, J. Küken, A. Auer, B. Han, P. Mercado, S. S. Rangapuram, H. Shen, L. Stella, X. Zhang et al., “Chronos-2: From univariate to universal forecasting,” arXiv preprint arXiv:2510.15821, 2025.arxiv.org/abs/2510.15821
  15. [15]A. Das, W. Kong, R. Sen, and Y. Zhou, “A decoder-only foundation model for time-series forecasting,” arXiv preprint arXiv:2310.10688, 2023.arxiv.org/abs/2310.10688
  16. [16]Y. Liang, H. Wen, Y. Xia, M. Jin, B. Yang, F. Salim, Q. Wen, S. Pan, and G. Cong, “Foundation models for spatio-temporal data science: A tutorial and survey,” in Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, 2025, pp. 6063–6073.
  17. [17]Z. Li, L. Xia, J. Tang, Y. Xu, L. Shi, L. Xia, D. Yin, and C. Huang, “Urbangpt: Spatio-temporal large language models,” in Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, 2024, pp. 5351–5362.arxiv.org/abs/2403.00813
  18. [18]Z. Shao, Z. Zhang, F. Wang, and Y. Xu, “Pre-training enhanced spatial-temporal graph neural network for multivariate time series forecasting,” in Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, 2022, pp. 1567–1577.DOI
  19. [19]J. Han, W. Zhang, H. Liu, T. Tao, N. Tan, and H. Xiong, “Bigst: Linear complexity spatio-temporal graph neural network for traffic forecasting on large-scale road networks,” Proceedings of the VLDB Endowment, vol. 17, no. 5, pp. 1081–1090, 2024.DOI
  20. [20]A. Liang, Y. Sun, X. Shi, and K. Li, “Time tracker: Mixture-of-experts-enhanced foundation time series forecasting model with decoupled training pipelines,” arXiv preprint arXiv:2505.15151, 2025.arxiv.org/abs/2505.15151
  21. [21]S. Zhong, J. Qiu, Y. Wu, X. Zou, Z. Rao, B. Yang, C. Guo, H. Xu, and Y. Jiang, “Learning to factorize spatio-temporal foundation models,” Advances in Neural Information Processing Systems, vol. 38, pp. 88 232–88 260, 2026.DOI
  22. [22]G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” arXiv preprint arXiv:1503.02531, 2015.arxiv.org/abs/1503.02531
  23. [23]T. Luo, Z. Fang, K. Duan, L. Chen, P. Feng, and M. Lu, “Towards online spatio-temporal prediction: A knowledge distillation driven continual learning approach,” in 2025 IEEE 41st International Conference on Data Engineering (ICDE). IEEE, 2025, pp. 2642–2655.
  24. [24]A. Romero, N. Ballas, S. E. Kahou, A. Chassang, C. Gatta, and Y. Bengio, “Fitnets: Hints for thin deep nets,” 2015. [Online]. Available: https://arxiv.org/abs/1412.6550
  25. [25]J. Gou, B. Yu, S. J. Maybank, and D. Tao, “Knowledge distillation: A survey,” International journal of computer vision, vol. 129, no. 6, pp. 1789–1819, 2021.DOI
  26. [26]F. Tung and G. Mori, “Similarity-preserving knowledge distillation,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 1365–1374.arxiv.org/abs/1907.09682
  27. [27]W. Park, D. Kim, Y. Lu, and M. Cho, “Relational knowledge distillation,” in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2019, pp. 3962–3971.arxiv.org/abs/1904.05068
  28. [28]W. Wang, H. Bao, S. Huang, L. Dong, and F. Wei, “Minilmv2: Multi-head self-attention relation distillation for compressing pretrained transformers,” in Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, 2021, pp. 2140–2151.arxiv.org/abs/2012.15828
  29. [29]Y. Jin, K. Chen, and Q. Yang, “Transferable graph structure learning for graph-based traffic forecasting across cities,” in Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining, 2023, pp. 1032–1043.DOI
  30. [30]R. Dey, A. Matin, N. Orwick, Y. Zhang, S. Pallickara, and S. L. Pallickara, “When to trust, how to distill: Multi-foundation model guidance for lightweight, robust scientific time series forecasting,” in Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, 2026, pp. 10 807–10 818.DOI
  31. [31]W. Wang, Y. Shen, Y. Tan, and Y. Chen, “S^2-kd: Semantic-spectral knowledge distillation spatiotemporal forecasting,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 2, 2026, pp. 1195–1203.
  32. [32]Y. Li, C. Yang, H. Zeng, Z. Dong, Z. An, Y. Xu, Y. Tian, and H. Wu, “Frequency-aligned knowledge distillation for lightweight spatiotemporal forecasting,” in 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025, pp. 7262–7272.arxiv.org/abs/2507.02939
  33. [33]J. Ni, Z. Liu, S. Wang, M. Jin, and W. Jin, “Timedistill: Efficient long-term time series forecasting with mlp via cross-architecture distillation,” in Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1, 2026, pp. 1113–1124.arxiv.org/abs/2502.15016
  34. [34]Y. Li, K. Ding, C. Yang, S.-Y. Chen, and Y. Tian, “Distilling time series foundation models for efficient forecasting,” in ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026, pp. 4631–4635.
  35. [35]Q. Zhang, X. Gao, H. Wang, S. M. Yiu, and H. Yin, “Efficient traffic prediction through spatio-temporal distillation,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 1, 2025, pp. 1093–1101.arxiv.org/abs/2501.10459
  36. [36]Z. Yue, Y. Wang, J. Duan, T. Yang, C. Huang, Y. Tong, and B. Xu, “Ts2vec: Towards universal representation of time series,” in Proceedings of the AAAI conference on artificial intelligence, vol. 36, no. 8, 2022, pp. 8980–8987.arxiv.org/abs/2106.10466
  37. [37]E. Elder, M. Ragab, Z. Chen, M. Wu, C. K. Kwoh, X. Li, and C. Guan, “Time-series representation learning via temporal and contextual contrasting,” in Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, 2021, pp. 2352–2359.
  38. [38]G. Woo, C. Liu, D. Sahoo, A. Kumar, and S. Hoi, “CoST: Contrastive learning of disentangled seasonal-trend representations for time series forecasting,” in International Conference on Learning Representations, 2022.arxiv.org/abs/2202.01575
  39. [39]J. Shlens, “A tutorial on principal component analysis,” arXiv preprint arXiv:1404.1100, 2014.
  40. [40]L. McInnes and J. Healy, “Accelerated hierarchical density based clustering,” in 2017 IEEE international conference on data mining workshops (ICDMW). IEEE, 2017, pp. 33–42.DOI
  41. [41]R. J. Campello, D. Moulavi, and J. Sander, “Density-based clustering based on hierarchical density estimates,” in Pacific-Asia conference on knowledge discovery and data mining. Springer, 2013, pp. 160–172.DOI
  42. [42]T. Aksu, G. Woo, J. Liu, X. Liu, C. Liu, S. Savarese, C. Xiong, and D. Sahoo, “Gift-eval: A benchmark for general time series forecasting model evaluation,” arxiv preprint arxiv:2410.0393, 2024.
  43. [43]J. Jiang, C. Han, W. X. Zhao, and J. Wang, “Pdformer: Propagation delay-aware dynamic long-range transformer for traffic flow prediction,” in Proceedings of the AAAI conference on artificial intelligence, vol. 37, no. 4, 2023, pp. 4365–4373.arxiv.org/abs/2301.07945
  44. [44]L. Bai, J. Yao, C. Li, X. Wang, and C. Wang, “Adaptive graph convolutional recurrent network for traffic forecasting,” Advances in neural information processing systems, vol. 33, pp. 17804–17815, 2020.arxiv.org/abs/2007.02842
  45. [45]Z. Wu, S. Pan, G. Long, J. Jiang, X. Chang, and C. Zhang, “Connecting the dots: Multivariate time series forecasting with graph neural networks,” in Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, 2020, pp. 753–763.arxiv.org/abs/2005.11650
  46. [46]Z. Wu, S. Pan, G. Long, J. Jiang, and C. Zhang, “Graph wavenet for deep spatial-temporal graph modeling,” arXiv preprint arXiv:1906.00121, 2019.arxiv.org/abs/1906.00121
  47. [47]X. Shi, S. Wang, Y. Nie, D. Li, Z. Ye, Q. Wen, and M. Jin, “Time-moe: Billion-scale time series foundation models with mixture of experts,” in International conference on learning representations, vol. 2025, 2025, pp. 34635–34667.

Paper details

vlnc:W_fy2nk6ra
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents