betaFeedback

WorksComputer Science

TriGen: A Generative Paradigm for Trigger-Induced Recommendation

Abstract

Generative recommendation typically operates under the User-Induced Recommendation (UIR) setting, predicting target items by autoregressively generating semantic IDs as a passive continuation of long-term user history. However, this paradigm overlooks Trigger-Induced Recommendation (TIR), an essential industrial scenario where a user click triggers an immediate follow-up feed. In TIR, recommendations must dynamically align with the instant trigger intent without compromising historical preference continuity. Traditional TIR methods rely on discriminative scoring and cannot guide token-level generation, while existing generative models lack mechanisms to reconcile this structural tension. To address this challenge, we propose TriGen, establishing a tailored generative paradigm for trigger-induced recommendation. TriGen introduces calibrated tail injection to re-encode local history, adaptively arbitrates between persistent tastes and immediate impulses via dynamic gating, and applies position-specific transformations to sustain level-aware trigger guidance across decoding steps. Extensive offline evaluations confirm consistent gains over competitive baselines. Furthermore, online A/B tests across six country markets demonstrate significant revenue and GMV improvements, leading to the full production deployment of TriGen.

CCS Concepts

• Information systems → Recommender systems; Personalization; • Computing methodologies → Neural networks.

Introduction

In recent years, Generative Recommendation (GR) has emerged as an important paradigm in recommender systems [5, 16, 31]. Instead of scoring items from a predefined candidate pool, semantic-ID-based GR models directly predict target items by autoregressively generating their discrete Semantic IDs (SIDs) [16]. This formulation unifies item representation with generation, showing strong capabilities in modeling complex user behavior sequences. Many representative GR methods formulate recommendation as predicting the next item from user history [16, 31], a setting we refer to as User-Induced Recommendation (UIR). In practical search, advertising, and recommendation platforms, another important scenario is Trigger-Induced Recommendation (TIR) [17, 27, 28]. As illustrated in Figure 1, when a user interacts with a specific product in a general feed, the platform immediately opens a follow-up feed organized around that item. In this context, the clicked item is no longer an ordinary historical interaction. Instead, it serves as an explicit trigger that conveys the user’s immediate shopping intent, which should actively steer the direction of follow-up recommendations.

Trigger-Induced Recommendation in an e-commerce application

Figure 1. Trigger-Induced Recommendation (TIR) in an e-commerce application. The item the user clicks (b) in the general feed (a) becomes the trigger of the follow-up feed (c), where it is pinned above the recommendations organized around it.

Although Trigger-Induced Recommendation (TIR) represents a cornerstone scenario across industrial digital platforms, trigger modeling remains largely unexplored in generative recommendation (GR). Existing GR frameworks inherently presume behavioral continuity, lacking dedicated mechanisms to accommodate sudden shifts in user intent. Meanwhile, representative TIR solutions are developed for discriminative CTR prediction using trigger-aware representations [17, 27, 28]. Because these methods produce only scalar ranking scores or static representations, they cannot steer discrete, token-by-token autoregressive generation. While extending GR to TIR marks a fundamental transition from discriminative scoring to condition-controlled generation, it inevitably exposes an essential structural tension between historical preference continuity and trigger-driven intent redirection. An overemphasis on continuity reduces the model to a conventional history-based system that fails to capture immediate user impulses. Conversely, an exclusive reliance on the trigger risks confining candidate generation to an overly narrow semantic space, disregarding persistent user preferences and causing severe overspecialization.

Reconciling this structural tension within an autoregressive framework introduces three challenges across encoding, representation, and decoding. First, during sequence encoding, standard self-attention allows extensive interaction histories to easily dilute a single trigger token, preventing it from effectively contextualizing recent user actions without an explicit mechanism to calibrate its influence. Second, the interplay between persistent preferences and immediate intent varies dynamically across requests. A user action may reflect either casual exploration where historical taste should prevail, or a decisive shift toward a new interest requiring strict trigger alignment, making static fusion schemes inadequate. Third, during autoregressive decoding, semantic IDs unfold sequentially from broad categories to fine-grained attributes. Conditioning only at the initial step inevitably weakens over successive generation steps and cannot accommodate the varying semantic granularity of each token, underscoring the need for persistent, position-aware guidance throughout the generation trajectory.

To address these challenges, we propose TriGen, an end-to-end framework that establishes a tailored generative paradigm for trigger-induced recommendation. As illustrated in Figure 2, TriGen systematically reconciles preference continuity and intent redirection across the entire autoregressive pipeline through three synergistic designs. At the sequence encoding stage, Calibrated Tail Injection places repeated trigger representations at the history tail, using RoPE to model relative positions and the repetition count to regulate trigger influence. At the representation stage, Adaptive Dual-Path Fusion decouples persistent historical preferences from immediate trigger impulses, employing an input-dependent gating module to dynamically arbitrate their balance on a per-request basis. Finally, at the autoregressive decoding stage, Position-Specific Trigger-PFFN incorporates level-aware transformations directly into the decoder layers, sustaining persistent trigger guidance that adapts to the evolving semantic granularity of each decoding step.

Overview of Trigen architecture

Figure 2. Overall architecture of TriGen. (a) Calibrated Tail Injection: The trigger is appended to the sequence tail with NN repetitions, placing it adjacent to recent behaviors. RoPE models relative positions, while NN controls the repetition count to regulate trigger influence. (b) Adaptive Dual-Path Fusion: Parallel pathways decouple trigger-conditioned steering from history-preserved preference, dynamically balanced by an input-dependent gating router. (c) Position-Specific Trigger-PFFN: Step-specific feed-forward networks project the fused representation into level-aware condition signals, maintaining persistent, coarse-to-fine guidance throughout autoregressive decoding.

We evaluate TriGen on both a large-scale industrial e-commerce dataset and a public benchmark. Offline experimental results demonstrate that TriGen consistently outperforms competitive sequential, generative, and discriminative TIR baselines. Furthermore, online A/B tests across six international markets achieve traffic-weighted overall lifts of +3.27% in Revenue and +0.90% in GMV. TriGen is now fully deployed in production, serving primary online traffic.

The main contributions of this work are summarized as follows:

  • A New Generative Paradigm for TIR: While TIR has historically relied on discriminative matching methods, our work pioneers an intent-controlled autoregressive sequence generation framework under preference continuity constraints. To the best of our knowledge, we are the first to systematically investigate generative recommendation in the TIR scenario.

  • The TriGen Framework: We identify the core structural tension between preference continuity and intent redirection in TIR-GR and propose TriGen to resolve it. The framework achieves strength-calibrated local reinterpretation via tail injection with RoPE, adaptively balances dual-path preferences through a gating router, and enforces persistent hierarchical intent guidance during decoding via position-specific TriggerPFFN.

  • Empirical Results and Deployment: Offline evaluations on large-scale industrial and public datasets demonstrate that TriGen consistently outperforms competitive sequential, generative, and discriminative baselines. Crucially, rigorous online A/B testing across six country markets confirms statistically significant business gains in both Revenue and Gross Merchandise Value, culminating in the full deployment of TriGen in production to serve primary online traffic.

Related Work

Trigger-Induced Recommendation

Industrial recommendation often separates candidate generation from ranking [2]. Within this setting, DIN activates historical behaviors relevant to a target item [33], while DIEN models the evolution of interests through behavior sequences [32]. Sequential recommenders likewise learn preferences from interaction histories using recurrent networks [6], causal self-attention [9], or bidirectional sequence modeling [19]. These approaches provide useful history-modeling mechanisms, but a clicked trigger introduces an additional, explicit signal of immediate intent.

DIHN studies Trigger-Induced Recommendation and highlights historical behaviors through trigger–target fusion and user-intent estimation [17]. DIAN explicitly distinguishes trigger-driven entry from habitual entry and balances trigger-based and trigger-free predictions according to estimated intention [27]. DEI2N further models evolving instant interest using temporal behavior information and trigger–target interactions [28]. These methods motivate the need to coordinate historical preferences with trigger intent. Their original formulations predict CTRs instead of item-token sequences, leaving open how to incorporate trigger information throughout semantic-ID generation.

Generative Recommendation

Transformer models [23] and unified text-to-text modeling [15] provide building blocks for sequence generation. P5 formulates multiple recommendation tasks within a unified language-processing framework [5]. In document retrieval, DSI and NCI generate document identifiers [21, 25], SEAL retrieves documents through generated text substrings [1], and GenRet learns discrete document identifiers [20]. These works establish different ways to connect sequence generation with retrieval; document retrieval and personalized next-item recommendation nevertheless have different conditioning signals and evaluation settings.

For semantic-ID-based recommendation, vector quantization and residual quantization provide discrete representations [10, 22]. TIGER represents items with semantic IDs and autoregressively generates the identifier of the next item [16]. LC-Rec integrates language and collaborative semantics through item indexing and alignment tuning [31]. LETTER improves item tokenization with semantic, collaborative, and diversity regularization [24]; ETEGRec jointly learns the tokenizer and recommender [11]; and ED2 coordinates dual dynamic indices for sequential recommendation [29].

Other directions include behavior–semantic collaboration in EAGER [26], multimodal generation in MMGRec [12], and parallel prediction of long semantic IDs in RPG [7]. RPG demonstrates that semantic-ID recommendation need not use token-by-token autoregressive decoding. At industrial scale, OneRec unifies retrieval and ranking with preference alignment [4], while HSTU formulates recommendation as sequential transduction [30]. Our focus is complementary: we study how an explicit trigger should influence history encoding, preference fusion, and each decoding step in autoregressive semantic-ID recommendation.

Preliminaries

Generative Recommendation with Semantic IDs

Generative recommendation reframes item recommendation as an autoregressive token generation task. Instead of retrieving items from a large candidate pool, the model directly predicts the discrete tokens of the target item [16].

Formally, each item ii is represented by a SID, which is a hierarchical sequence of LL discrete tokens:

ci=[ci(1),ci(2),…,ci(L)],ci(l)∈Vl,(1)c_i = [c_i^{(1)}, c_i^{(2)}, \ldots, c_i^{(L)}], \quad c_i^{(l)} \in\mathcal{V}_l, \tag*{(1)}

where ci(l)c_i^{(l)} denotes the token at the ll-th level, and Vl\mathcal{V}_l is the corresponding codebook vocabulary. Typically obtained via residual quantization methods such as RQ-VAE [10, 16], these tokens encode successive refinements of an item representation. This motivates level-specific conditioning, without requiring every level to correspond to an explicitly labeled category or attribute.

Let Su=[i1,i2,…,iT]S_u = [i_1, i_2, \ldots, i_T] denote the historical interaction sequence of user uu. Standard generative recommenders predict the target item iT+1i_{T+1} by autoregressively generating its SID tokens conditioned on the history:

P(ciT+1∣Su)=∏l=1LP(ciT+1(l)∣Su,ciT+1(<l)),(2)P(c_{i_{T+1}} \mid S_u) = \prod_{l=1}^{L} P\left(c_{i_{T+1}}^{(l)} \mid S_u, c_{i_{T+1}}^{(<l)}\right), \tag*{(2)}

where ciT+1(<l)=[ciT+1(1),…,ciT+1(l−1)]c_{i_{T+1}}^{(<l)} = [c_{i_{T+1}}^{(1)}, \ldots, c_{i_{T+1}}^{(l-1)}] represents the prefix tokens generated before level ll.

Trigger-Aware Generative Recommendation

In practical platforms, recommendations are often triggered by a specific user interaction, such as clicking an item in the main feed to enter a related follow-up feed. We define this clicked item as the trigger item itri_{\mathrm{tr}}, represented by its SID ctr=[ctr(1),…,ctr(L)]c_{\mathrm{tr}} = [c_{\mathrm{tr}}^{(1)}, \ldots, c_{\mathrm{tr}}^{(L)}]. The trigger reflects the user’s immediate interest and should explicitly guide subsequent recommendations.

Unlike standard generative recommendation that relies solely on history, Trigger-Induced Recommendation (TIR) requires the model to generate the target item conditioned on both user history and the trigger item:

P(ciT+1∣Su,ctr)=∏l=1LP(ciT+1(l)∣Su,ctr,ciT+1(<l)).(3)P(c_{i_{T+1}} \mid S_u, c_{\mathrm{tr}}) = \prod_{l=1}^{L} P\left(c_{i_{T+1}}^{(l)} \mid S_u, c_{\mathrm{tr}}, c_{i_{T+1}}^{(<l)}\right). \tag*{(3)}

This formulation highlights a core challenge: the trigger ctrc_{\mathrm{tr}} must not act as a mere static feature. Instead, it must actively steer the generation trajectory across all decoding levels l∈{1,…,L}l \in\{1, \ldots, L\} while preserving the user’s long-term historical preferences.

Methodology

To systematically reconcile preference continuity and intent redirection, we propose TriGen, an encoder-decoder generative framework tailored for trigger-induced recommendation. As illustrated in Figure 2, TriGen operationalizes this trade-off across the entire generation pipeline through three synergistic designs. At the sequence encoding stage, Calibrated Tail Injection places repeated trigger representations at the history tail, using RoPE to model relative positions and the repetition count to regulate trigger influence. At the representation stage, Adaptive Dual-Path Fusion decouples persistent historical preferences from immediate trigger impulses, employing an input-dependent gating module to dynamically balance both paths for each request. Finally, at the autoregressive decoding stage, Position-Specific Trigger-PFFN incorporates level-aware transformations directly into the decoder layers, sustaining persistent trigger guidance that adapts to the evolving semantic granularity of each decoding step. We elaborate on each component below.

Calibrated Tail Injection

To address local reinterpretation and strength control in TIR, we propose Calibrated Tail Injection, which incorporates the trigger directly into the sequence tail within encoder self-attention. The core intuition is that an immediate trigger click does not reshape the entire interaction history uniformly; rather, it primarily modulates a localized subset of behaviors, particularly recent and relevant ones. Placing the trigger inside the sequence rather than treating it as an external global condition enables direct attention interactions with this relevant local context.

Formally, let H=(h1,h2,…,hM)H=(h_1,h_2,\ldots,h_M) denote the historical sequence representation within a fixed context window of length MM, where each hj∈Rdh_j\in\mathbb{R}^d, and let etrg∈Rde^{trg}\in\mathbb{R}^d be the representation of the trigger item. To explicitly calibrate the trigger’s intervention strength over historical behaviors, we introduce NN consecutive copies of the trigger at the tail of the sequence, yielding:

H~=(hN+1,…,hM,etrg,…,etrg).(4)\widetilde{H}=(h_{N+1},\ldots,h_M,e^{trg},\ldots,e^{trg}). \tag*{(4)}

This design coordinates two complementary mechanisms. Placing the trigger at the sequence tail makes it positionally adjacent to the user’s most recent behaviors. Combined with the relative-position modeling of Rotary Position Embedding (RoPE) [18], this arrangement is intended to facilitate interactions between the trigger and recent behaviors. Meanwhile, the repetition factor NN regulates the query-key capacity occupied by the trigger in self-attention: a larger NN empowers the trigger to steer the context toward immediate intent, whereas a smaller NN prevents it from overriding historical preferences.

Finally, H~\widetilde{H} is processed by a bidirectional Transformer encoder to produce the trigger-aware contextual sequence O=(o1,o2,…,oM)O=(o_1,o_2,\ldots,o_M), which provides representations for both the Adaptive Dual-Path Fusion module and the decoder cross-attention.

Adaptive Dual-Path Fusion

While Calibrated Tail Injection regulates the trigger’s localized intervention strength within encoder self-attention, the resulting representations inevitably couple persistent historical preferences with immediate trigger impulses. Disentangling these two factors is crucial in trigger-induced recommendation, as user behavior may reflect both habitual preferences and trigger-induced intent [27]. If the system over-emphasizes the trigger item, recommendations risk collapsing into myopic over-reaction or being misled by accidental clicks; conversely, if it adheres strictly to historical habits, the redirective signal of the trigger is effectively lost. Consequently, the generative condition must be explicitly decoupled and adaptively reconciled on a per-request basis. To this end, we propose Adaptive Dual-Path Fusion (Figure 2, Stage 3), which decouples condition modeling into a trigger-guided instant path and a history-preserving baseline path, dynamically arbitrating their synthesis via an input-dependent gating router.

Decoupled Dual-Path Condition Modeling. To prevent mutual interference between transient impulses and baseline habits, we bypass the coupled encoder sequence and extract two complementary condition pathways directly from the unperturbed historical representations. Specifically, the trigger path projects the trigger item into distinct step-specific queries to retrieve historical behaviors relevant to the immediate intent. Symmetrically, the history path uses a set of learnable default queries (marked as “Default” in Figure 2) to summarize the user’s intrinsic preference prototypes independently of the trigger. Formally, for each decoding step, the condition vectors from both paths are extracted through attention over the raw historical sequence:

tl=Softmax⁡((Wltretr)(HWK)Td)(HWV),hl=Softmax⁡(qldef(HWK)Td)(HWV).(5)\begin{aligned} t_l &= \operatorname{Softmax}\left(\frac{\left(W_l^{tr}e^{tr}\right)\left(HW_K\right)^{\mathsf{T}}}{\sqrt{d}}\right)\left(HW_V\right),\\ h_l &= \operatorname{Softmax}\left(\frac{q_l^{def}\left(HW_K\right)^{\mathsf{T}}}{\sqrt{d}}\right)\left(HW_V\right). \tag*{(5)} \end{aligned}

where l∈{1,…,L}l \in\{1,\ldots,L\} indexes the decoding step, etrg∈Rde^{trg} \in\mathbb{R}^{d} is the trigger representation, and H∈RM×dH \in\mathbb{R}^{M \times d} denotes the pristine historical sequence. Here, Wltr∈Rd×dW_l^{tr} \in\mathbb{R}^{d \times d} is the step-specific projection matrix, qldef∈Rdq_l^{def} \in\mathbb{R}^{d} is the learnable default query, and WK,WV∈Rd×dW_K,W_V \in\mathbb{R}^{d \times d} project historical behaviors into shared key and value spaces. Stacking these vectors yields the trigger condition sequence T=[t1,…,tL]T = [t_1,\ldots,t_L] and the history condition sequence Hcond=[h1,…,hL]H_{cond} = [h_1,\ldots,h_L], capturing immediate redirection and baseline continuity.

Hierarchical Gated Routing. Because the intensity of intent redirection naturally varies across requests, a static combination of both paths is inadequate. We introduce an input-dependent router to learn dynamic fusion weights on a per-request basis. The router first condenses each pathway into a global profile via sequence-level mean pooling, computes step-wise gating coefficients, and synthesizes the calibrated condition vector through a position-wise convex combination:

w=[w1,…,wL]=σ(Wr[1L∑j=1Ltj∥1L∑j=1Lhj]+br),cl=wltl+(1−wl)hl,∀l∈{1,…,L},(6)\begin{aligned} \mathbf{w} &= [w_1,\ldots,w_L] = \sigma\left(W_r\left[\frac{1}{L}\sum_{j=1}^{L}t_j \mathbin{\Vert} \frac{1}{L}\sum_{j=1}^{L}h_j\right]+b_r\right),\\ c_l &= w_l t_l + (1-w_l)h_l,\quad\forall l \in\{1,\ldots,L\}, \tag*{(6)} \end{aligned}

where [⋅∥⋅][\cdot\mathbin{\Vert} \cdot] denotes vector concatenation, Wr∈RL×2dW_r \in\mathbb{R}^{L \times2d} and br∈RLb_r \in\mathbb{R}^{L} are learnable routing parameters, and σ(⋅)\sigma(\cdot) is the element-wise sigmoid function. The resulting sequence C=[c1,…,cL]∈RL×dC = [c_1,\ldots,c_L] \in\mathbb{R}^{L \times d} establishes a unified condition signal that balances instant redirection with historical continuity, which is subsequently passed to Stage 4 to steer autoregressive generation.

Position-Specific Trigger-PFFN

While encoder-side representations and dual-path condition fusion offer essential global constraints, steering the target generation also requires step-level guidance on the decoder side. In generative recommendation, the target item is represented as a Semantic ID and generated token by token in an autoregressive manner. Residual-quantized Semantic IDs encode an item representation through successive quantization levels [10, 16]. Earlier tokens provide an initial approximation, while later tokens encode residual refinements; their semantic roles can therefore differ across positions. Consequently, different decoding steps naturally require distinct forms of trigger modulation instead of a uniform signal across all positions. To provide step-tailored guidance throughout the generative process, we introduce the Position-Specific Trigger Feed-Forward Network (Trigger-PFFN) on the decoder side (Figure 2, Stage 4).

Taking the calibrated condition sequence C=[c1,…,cL]∈RL×dC = [c_1,\ldots,c_L] \in\mathbb{R}^{L \times d} from Stage 3 as input, Trigger-PFFN first maps each condition vector clc_l through a dedicated feed-forward layer to adapt to the semantic granularity of the corresponding position:

c~l=MLP⁡l(cl),l∈{1,…,L},(7)\widetilde{c}_l = \operatorname{MLP}_l(c_l),\quad l \in\{1,\ldots,L\}, \tag*{(7)}

where MLP⁡l(⋅)\operatorname{MLP}_l(\cdot) is an independent multi-layer perceptron learned specifically for the ll-th Semantic ID level. By using independent projection parameters for each position, the model can flexibly tailor the trigger condition, providing broad categorical guidance for early decoding steps and detailed attribute control for later steps.

To maintain continuous guidance across the entire generation process, we incorporate the modulated condition vectors C~=[c~1,…,c~L]∈RL×d\widetilde{C} = [\widetilde{c}_1,\ldots,\widetilde{c}_L] \in\mathbb{R}^{L \times d} directly into the decoder input representations. Under the standard teacher-forcing protocol with the ground-truth Semantic ID sequence (y1,y2,…,yL)(y_1,y_2,\ldots,y_L), the right-shifted decoder input at step ll is formed as:

xldec={e(BOS)+c~1,l=1,e(yl−1)+c~l,l∈{2,…,L},(8)x_l^{dec} = \begin{cases} e(\mathrm{BOS})+\widetilde{c}_1, & l=1,\\ e(y_{l-1})+\widetilde{c}_l, & l\in\{2,\ldots,L\}, \end{cases} \tag*{(8)}

where e(⋅)∈Rde(\cdot) \in\mathbb{R}^{d} denotes the token embedding lookup function and BOS\mathrm{BOS} represents the start-of-sequence token. By adding the position-specific condition vector c~l\widetilde{c}_l to the input at each position, the trigger guidance accompanies the target prefix at every decoding step. The resulting sequence Xdec=[x1dec,…,xLdec]∈RL×dX^{dec} = [x_1^{dec},\ldots,x_L^{dec}] \in\mathbb{R}^{L \times d} is then processed by causal self-attention and cross-attention over the encoder output O∈RM×dO \in\mathbb{R}^{M \times d}, enabling the decoder to autoregressively predict the target tokens with hierarchy-aware trigger steering.

Experiments

Experimental Setup

Datasets

We use industrial and public datasets. The industrial dataset is collected from a real Trigger-Induced Recommendation scenario on a major e-commerce platform. It contains large-scale user behavior and feedback logs. The dataset covers real exposure, click, and follow-up feedback signals, supporting industrial TIR evaluation.

For the public benchmark, we use the Industrial_and_Scientific category from Amazon Reviews 2018 [14]. We use the 5-core version and keep interactions from 2016-10 to 2018-11. Because this public dataset does not provide native TIR annotations, we construct simulated triggers on it. Specifically, we treat the most recent interaction in the user history as the trigger and use it to predict the subsequent target item. This assumes that the most recent interaction reflects instant interest. The statistics of the public dataset are reported in Table 5.

StatisticTrainValidTest
beforesamples36,2594,5324,533
beforeusers7,0342,0111,987
beforeavg. ∣H∣|H|3.774.644.96
beforemax. ∣H∣|H|101010
aftersamples29,2254,1394,266
afterdropped7,034393267
afteravg. ∣H∣|H|3.443.994.20
aftermax. ∣H∣|H|999

Table 5. Statistics of the public dataset before and after the trigger rewriting. “After” is the sample set used by every model; avg. ∣H∣|H| there excludes the trigger, which trigger-free baselines still read as the last history item.

Evaluation Metrics

We use Recall@K and NDCG@K [8] as offline evaluation metrics, where K∈{5,10}K \in\{5,10\}. The model first generates the Semantic ID sequence of the target item and then maps it back to the final item. For online evaluation, we report business results in six country markets, namely Indonesia, Malaysia, Singapore, the Philippines, Thailand, and Vietnam. In addition, we report the weighted aggregated GMV and REV across these six countries.

Compared Methods

To evaluate the effectiveness of Tri-Gen and investigate whether naive trigger conditioning suffices, we compare our framework against four categories of baselines:

(1) TIR-specific discriminative methods, including DIHN [17], DEI2N [28], and DIAN [27], which model trigger information under matching or CTR-scoring paradigms; (2) Standard sequential recommendation methods, including GRU4Rec [6] and SASRec [9], which rely solely on historical interaction sequences without trigger modeling; (3) Generative and sequential-transduction baselines, including TIGER [16], HSTU [30], and LC-Rec [31]; and (4) Simple TIR-GR baselines, comprising GR + trigger append, GR + trigger conditioning, and GR + trigger-history attention. These baselines adopt the identical generative backbone, codebook, and parameter budget as TriGen to examine whether straightforward trigger injection can deliver comparable gains. Detailed configurations of all compared methods are provided in Appendix § 8.

5.1.4 Implementation Details. All methods are trained and evaluated on the same data split. We use AdamW [13] for optimization and beam search for inference. Semantic IDs are fixed discrete codes and are not updated during training. Configurations are provided in Appendix § 10.

Overall Performance

Table 1 compares all methods on the industrial and Amazon datasets. Overall, TriGen achieves the best or highly competitive performance on both datasets. These results support explicit trigger modeling in generative recommendation.

CategoryMethodIndustrial DatasetAmazon Industrial_and_Scientific
R@5R@10N@5N@10R@5R@10N@5N@10
TIR-specificDIHN1.546.391.082.634.786.833.033.69
DIAN1.596.801.152.824.936.943.644.26
DEI2N1.617.201.192.985.177.464.084.81
SequentialGRU4Rec1.667.261.223.027.329.976.036.88
SASRec2.167.841.573.399.3411.887.198.01
Generative (Trigger-free)HSTU2.587.131.833.289.3511.647.187.91
TIGER2.397.531.743.389.7512.417.518.36
LC-Rec2.467.301.793.3410.2012.987.958.85
Simple TIR-GRGR + trigger append4.3512.153.125.6210.2713.267.938.89
GR + trigger conditioning4.4611.443.285.5110.2412.997.858.74
GR + trigger-history attention3.339.802.414.476.819.384.725.54
OursTriGen5.5815.504.167.3310.8313.228.459.21

Table 1. Table 1: Overall performance comparison on the industrial dataset and the Amazon Industrial_and_Scientific benchmark. Results report the mean performance over 5 seeds (corresponding standard deviations are detailed in Appendix § 11, Table 7). R@KR@K and N@KN@K denote Recall@KK and NDCG@KK, respectively. The best performance in each column is highlighted in bold, and the second-best is underlined.

Compared with TIR-specific methods, TriGen directly models how the trigger affects hierarchical Semantic ID generation within a generative framework, combining instant interest with generation. Compared with standard sequential recommendation methods such as GRU4Rec and SASRec, TriGen’s gains indicate that history alone is insufficient to capture current intent in TIR. Compared with generative recommendation baselines such as TIGER, HSTU, and LC-Rec, the consistent gains of TriGen suggest that TIR also requires dedicated trigger modeling.

Crucially, TriGen consistently outperforms the Simple TIR-GR baselines across both datasets. Naive strategies such as GR + trigger append and GR + trigger conditioning fail to calibrate trigger influence against long histories and cannot accommodate the coarse-to-fine granularity of Semantic IDs, limiting gains under noisy triggers. The decisive margin between TriGen and these simple baselines rules out the hypothesis that any arbitrary trigger conditioning scheme suffices, confirming that our tailored architectural designs provide benefits far beyond the mere presence of trigger information.

Ablation Study

To examine the contributions of TriGen’s components, we evaluate the full model against four variants on the industrial dataset under identical backbone configurations: (1) no trigger, which removes the trigger item entirely and relies solely on user history for recommendation; (2) w/o Calibrated Tail Injection, which excludes the trigger from the encoder historical sequence; (3) w/o Adaptive Dual-Path Fusion, which removes the dual-path condition channels and the gating router, and instead feeds the original trigger embedding into the retained position-specific Trigger-PFFN; and (4) w/o Position-Specific Trigger-PFFN, which removes the position-specific trigger projections and completely eliminates trigger conditioning from the decoder.

As shown in Table 2, the full TriGen model performs best across all metrics, supporting the contribution of each component.

VariantR@5R@10N@5N@10
no trigger2.888.272.083.81
w/o Calibrated Tail Injection4.1611.042.985.19
w/o Adaptive Dual-Path Fusion5.1114.513.616.34
w/o Position-Specific Trigger-PFFN4.2811.103.115.31
TriGen (full)5.5815.504.167.33

Table 2. Ablation study of TriGen on the industrial dataset. R@KR@K and N@KN@K denote Recall@KK and NDCG@KK (%), respectively. The best performance in each column is highlighted in bold.

Removing the trigger causes the largest accuracy drop. In trigger-induced scenarios, user behavior is strongly governed by the immediate interaction. Without the trigger, the model cannot identify the user’s instant intent, demonstrating that historical interaction data alone cannot yield accurate recommendations in such contexts.

Excluding the trigger from sequence encoding also degrades recommendation accuracy substantially. This indicates that relying solely on downstream trigger conditioning is insufficient. By appending the trigger to the sequence tail, the encoder exploits self-attention to contextualize past behaviors against the trigger, particularly the most recent interactions. Omitting this injection leaves historical representations uninformed by the trigger context.

Removing Adaptive Dual-Path Fusion while retaining Trigger-PFFN reduces both Recall and NDCG. In this variant, the decoder-side condition is derived directly from the original trigger embedding instead of from the fused trigger-guided and history-preserving representations. The performance gap supports the benefit of incorporating historical context through adaptive dual-path fusion, compared with using the trigger embedding alone.

Finally, removing Trigger-PFFN and eliminating all trigger conditioning from the decoding stage degrades recommendation performance. Compared with this variant, the full model achieves better results by providing position-specific trigger guidance throughout autoregressive generation. These results support the effectiveness of explicit decoder-side trigger conditioning.

Fine-grained Design Analysis

We further examine architectural choices within each component. Industrial-dataset results are reported in Table 3 and Figure 3.

Figure 3. Effect of the trigger repeat number on the industrial dataset. All evaluation metrics first improve and then decline, confirming that a calibrated repetition count achieves the best balance.

SettingR@5R@10N@5N@10
Encoder-side Designs
trigger at sequence head5.3114.583.886.79
w/o RoPE4.9614.243.586.56
w/o repeat5.2014.833.826.91
Dual-Path Fusion Designs
uniform path combination4.1511.872.985.47
Decoder-side Designs
start-only injection4.6212.933.365.99
shared projection layer4.9013.853.616.51
TriGen (full)5.5815.504.167.33

Table 3. Fine-grained design analysis on the industrial dataset. All variants share the same backbone architecture and differ only in the evaluated design choice. R@K and N@K denote Recall@K and NDCG@K (%), respectively. The best performance in each column is highlighted in bold.

Encoder-side Designs. We first examine the role of Calibrated Tail Injection. Placing the trigger at the sequence tail outperforms placing it at the sequence head. Tail placement makes the trigger positionally adjacent to recent behaviors, and the observed improvement supports this placement choice in our setting.

Removing rotary position embeddings leads to a performance decline, supporting the effectiveness of RoPE in the trigger-injected encoder. Together with tail placement, RoPE provides relative-position information for modeling interactions between the trigger and historical behaviors.

Finally, excluding repeated insertion reduces accuracy. Inserting the trigger only once fails to provide sufficient intervention strength within the sequence, while Figure 3 shows that performance first increases and then gradually declines as the repetition count grows. This bell-shaped trend demonstrates that trigger influence must be properly calibrated: an insufficient count leaves the trigger overshadowed by long history, whereas excessive repetition overwhelms the user’s historical context.

Dual-Path Fusion Designs. We further investigate the design of the gating module in Adaptive Dual-Path Fusion by comparing

it with a static uniform combination, where both paths are simply averaged without routing weights.

As shown in Table 3, replacing the adaptive router with uniform averaging results in a clear drop in performance. This degradation demonstrates that different requests require different balances between instant redirection and historical continuity. An incidental trigger click should rely more heavily on baseline historical tastes, while an exploratory session demands stronger trigger guidance. By dynamically predicting the routing weights from the global states of both pathways, the gating module effectively tailors the condition to individual user requests.

Decoder-side Designs. We evaluate two key design aspects of Position-Specific Trigger-PFFN on the decoder side: the injection position and the projection parameterization.

Comparing full position-wise injection against injecting the condition exclusively at the start token reveals that start-only injection is inadequate for autoregressive generation. In a multi-step decoding process, introducing the condition only at the initial position weakens its influence as the generated prefix grows longer. Providing condition guidance at every decoding step ensures persistent steering throughout the entire target item generation.

Moreover, replacing position-specific projections with a single shared feed-forward layer noticeably hurts generation quality. Because Semantic IDs represent items from broad categories down to specific attributes, each decoding position is responsible for a distinct semantic granularity. Dedicated projections allow early positions to focus on category-level steering while letting later positions specialize in attribute-level details, matching the coarse-to-fine structure of Semantic IDs.

Online A/B Test

To evaluate TriGen in production, we conducted an online A/B test on an e-commerce platform across six country markets: Indonesia

(ID), Malaysia (MY), Singapore (SG), the Philippines (PH), Thailand (TH), and Vietnam (VN). The experiment was deployed within the trigger-induced feed on product detail pages, a surface with exposure, click, and conversion signals. We adopted the deployed production model, which implements the competitive generative baseline GR+trigger appendGR + trigger\ append, as the control group, and deployed TriGen as the treatment group. The evaluation ran from June 28 to July 2, 2026, with each bucket allocated 10% of randomly sampled production traffic, while upstream candidate sources and downstream serving pipelines remained strictly identical.

The results are summarized in Figure 4. TriGen demonstrates robust performance improvements across heterogeneous international markets. In terms of platform monetization, TriGen achieves consistent revenue gains across all six countries, yielding relative lifts between +1.64% and +8.16%. Gross Merchandise Value (GMV) concurrently increases in four out of six countries, achieving up to a +20.54% improvement. On the traffic-weighted aggregate across Southeast Asia (SEA), TriGen delivers an overall lift of +3.27% in Revenue and +0.90% in GMV, with metrics demonstrating strong statistical significance (p<0.05p < 0.05, two-sided test).

Figure 4. Online A/B test results across six international markets. REV (Revenue) and GMV (Gross Merchandise Value) denote the relative lifts of TriGen over the deployed production baseline. SEA represents the traffic-weighted aggregate.

During serving, the additional computational overhead of TriGen is practically negligible. Because the tailored components rely solely on lightweight cross-attentions, a compact routing unit, and compact position-specific feed-forward layers, the online inference latency is approximately 10.5 ms, remaining virtually indistinguishable from that of the control model. These production results substantiate our offline findings, confirming that resolving the structural tension between trigger steering and preference continuity brings decisive commercial value under strict serving budgets. Following the sustained performance advantages observed during the trial, TriGen has been fully deployed to production since July 2026, serving primary online traffic across the platform.

Case Analysis

The offline and online results show that TriGen improves accuracy and business metrics, but they do not directly show how the trigger affects generation. We therefore inspect two real-world users in Figure 5 who clicked the same trigger, a quilted shoulder bag, but have different histories.

Two real-world user cases that share the same trigger

Figure 5. Two real-world cases that share the same trigger, a quilted shoulder bag. In the top case the history provides a travel-outfit context; in the bottom case it is unrelated to the trigger. Each row shows the history, the trigger, and the items TriGen generates.

In the first case, the history is informative. The user has browsed a dress, a sun hat, sunglasses, and a suitcase, which form a travel-outfit context. TriGen does not generate only bags. It returns summer tops and lace shorts that continue the outfit line, one bag close to the trigger, and a travel skincare set. This suggests that the history-preserving path keeps the travel scenario available, so the trigger is interpreted as part of the trip.

In the second case, the history provides little support for the click. The recent items are digital and desk accessories, which are unrelated to a shoulder bag. Here TriGen concentrates all slots on bags and varies style and material instead of category. Without useful history context, the model relies more on the trigger-conditioned path. Together, the two cases show that the fusion is adaptive. The same trigger is expanded into a broader scenario when the history supports it, and followed more closely when it does not.

Conclusion

In this work, we study generative recommendation for Trigger-Induced Recommendation. To model the user’s instant interest carried by the trigger, we propose TriGen, which combines encoder-side trigger injection, DualPathFusion, and TriggerPFFN. Experiments on both industrial and public datasets show that TriGen consistently outperforms strong baselines. Online A/B testing further validates its effectiveness in a real production environment, and TriGen has now been fully deployed online. Overall, our results show that explicit trigger modeling is crucial for generative recommendation in TIR.

Dataset and Trigger Construction

Trigger-induced recommendation is usually evaluated on the Alimama dataset, but its fields are anonymized so aggressively that no usable item text survives, and without text there is nothing for the Semantic ID tokenizer of Appendix § 9 to quantize. We therefore evaluate on Amazon Reviews [14], a widely used public benchmark in generative recommendation [16, 31], and simulate the trigger as described below.

The public dataset is the Industrial_and_Scientific category of Amazon Reviews 2018, in its 5-core version and restricted to the window from 2016-10 to 2018-11. Every user’s chronologically ordered interactions are expanded into (history, target) samples by a sliding window whose history keeps the 10 most recent items, and the samples of each user are split chronologically at a ratio of about 8:1:1, so the earliest windows form the training set and the latest ones the test set. Item text is used only by the tokenizer of Appendix § 9; the recommenders consume Semantic IDs only. Statistics of the industrial dataset are reported in Table 4.

Dataset#Users#Items#Samples
Industrial49,277,990158,533,2411,492,770,272

Table 4. Statistics of the industrial dataset.

The public data carries no trigger field, so the trigger has to be simulated. Given a sample with history (x1,…,xm)(x_1,\ldots,x_m) and target yy, we take the most recent history item as the trigger and keep the rest as the long-term history:

(x1,…,xm−1,xm), y⟹(x1,…,xm−1),xm,y,(9)(x_1,\ldots,x_{m-1},x_m),\ y \Longrightarrow(x_1,\ldots,x_{m-1}),\quad x_m,\quad y, \tag*{(9)}

which rests on the assumption that the most recent interaction explicitly guides the next one. This split only changes how the sequence is presented, not what it contains. Models that consume a trigger read (x1,…,xm−1)(x_1,\ldots,x_{m-1}) as history and xmx_m as the trigger, while every trigger-free baseline reads the whole sequence (x1,…,xm)(x_1,\ldots,x_m) as its history. Both therefore see exactly the same items, and no baseline is shortened by the rewriting. Samples whose original history has length one are dropped for all models alike, since they would leave the history-preserving path of DualPathFusion without input, so the evaluated sample set stays identical across models. Table 5 reports the statistics both before and after this step. Aggregated over the three splits, the data contain 3,686 items, 7,694 users and 49,371 interactions, i.e. a density of 0.174%.

One consequence of this protocol deserves an explicit statement, because it works against TriGen rather than for it. In production the trigger is a real event: the user has just clicked the item, and that click carries a short-term intent that the next interaction often follows. The simulated trigger has no such event behind it. It is an ordinary past interaction that merely happens to be the most recent one, so the intent signal TriGen is designed to exploit is weaker and noisier here than in deployment. Because the baselines keep the full sequence, the public-dataset margin is best read as a lower bound on what trigger modeling contributes: it is obtained with a degraded trigger, and the industrial dataset, where the trigger is a genuine field, is the setting in which the mechanism is measured at full strength.

Baselines

  • DIHN [17] is an early model designed for the trigger-induced setting, which it casts as a CTR task. A dedicated module scores how strongly the user is currently committed to the trigger, and that score modulates both a trigger–target fusion and a trigger-aware aggregation of the behaviour history.

  • DIAN [27] turns the balance between trigger and history into an explicit decision. The network judges whether the request is driven by the trigger or by longer-term preference, and the trigger-based and trigger-free predictions are then merged through a probability decomposition over these two cases rather than through a learned gate.

  • DEI2N [28] argues that the instant interest revealed by the trigger does not stay constant. Alongside an intent estimator in the spirit of DIHN, it adds a time-aware component and an explicit trigger–target interaction unit, so that the weight given to the trigger can drift as the sequence progresses.

  • GRU4Rec [6] encodes the user behaviour sequence with gated recurrent units and predicts the next item from the final hidden state. It is the standard recurrent baseline for session-based and sequential recommendation.

  • SASRec [9] replaces recurrence with a causal self-attention stack, which lets every position attend to all earlier behaviours, and is trained with next-item prediction over the item ID vocabulary.

  • TIGER [16] is the representative Semantic ID generative recommender. Items are quantized into hierarchical discrete codes, and an encoder–decoder Transformer autoregressively generates the code sequence of the next item, which is then decoded back to an item by beam search over the code space.

  • HSTU [30] formulates recommendation as sequential transduction over a unified stream of user actions, with an attention design tailored to long, non-stationary behaviour sequences.

  • LC-Rec [31] adapts a large language model to Semantic ID recommendation. On top of the quantized indices it introduces a set of alignment tuning tasks that bridge the language space of the backbone and the newly introduced item index tokens, and then recommends by generating index tokens.

  • GR + trigger append is a simple generative baseline that appends the trigger item as a single terminal token to the historical sequence within the standard encoder self-attention mechanism, evaluating whether directly exposing the trigger to the encoder is sufficient.

  • GR + trigger conditioning keeps the historical sequence encoder unchanged and projects the trigger representation via a shared feed-forward layer, adding it directly to the token embeddings at every autoregressive decoding step to assess uniform condition steering.

  • GR + trigger-history attention uses the trigger representation as a cross-attention query over the pristine historical sequence to construct an aggregated context vector for generation, verifying whether conventional attention pooling without hierarchical decoding adaptation is adequate.

Semantic ID Construction

The title and description of each item are concatenated and encoded by a pretrained text embedding model, whose last hidden states are mean pooled into one embedding vector xi∈R256x_i \in\mathbb{R}^{256}. A residual-quantized autoencoder [10] then maps it to zi=Enc⁡(xi)∈R32z_i = \operatorname{Enc}(x_i) \in\mathbb{R}^{32} and quantizes it with L=3L = 3 codebooks E(l)∈RK×32E^{(l)} \in\mathbb{R}^{K \times32} of size K=256K = 256. Starting from r1=zir_1 = z_i, every level selects its nearest code and passes the residual on:

ci(l)=arg⁡min⁡k∥rl−Ek(l)∥2,rl+1=rl−Eci(l)(l)(10)c_i^{(l)} = \underset{k}{\arg\min}\lVert r_l - E_k^{(l)}\rVert_2,\qquad r_{l+1} = r_l - E_{c_i^{(l)}}^{(l)} \tag*{(10)}

so that item ii is represented by the tuple (ci(1),ci(2),ci(3))(c_i^{(1)}, c_i^{(2)}, c_i^{(3)}) and reconstructed from z^i=∑lEci(l)(l)\hat{z}_i = \sum_l E_{c_i^{(l)}}^{(l)}. The tokenizer is trained with a reconstruction term and the usual quantization terms,

L=∥xi−Dec⁡(z^i)∥2+∑l(∥sg⁡[rl]−Eci(l)(l)∥2+β∥rl−sg⁡[Eci(l)(l)]∥2),(11)\mathcal{L} = \lVert x_i - \operatorname{Dec}(\hat{z}_i)\rVert^2 + \sum_l \left(\lVert\operatorname{sg}[r_l] - E_{c_i^{(l)}}^{(l)}\rVert^2 + \beta\lVert r_l - \operatorname{sg}[E_{c_i^{(l)}}^{(l)}]\rVert^2\right), \tag*{(11)}

with β=0.25\beta= 0.25, sg⁡[⋅]\operatorname{sg}[\cdot] the stop-gradient operator and codebooks initialized by kk-means; a Sinkhorn-based assignment [3] on the last level spreads items that would otherwise share an identical tuple. Semantic IDs are produced once and then frozen: during recommender training they receive no gradient and act purely as a discrete input and supervision target.

Quantization is not injective, so a few items still collide; 3,686 items map to 3,670 distinct tuples on our data. We therefore keep a one-to-many index, and at evaluation time every generated tuple is expanded into all of its items, which occupy consecutive slots of the ranked list. A one-to-one mapping would make the losing items unreachable and understate Recall; discarding the groups would inflate it.

Implementation Details

Table 6 lists the configuration used on the public dataset; all variants share it and differ only in the component under study. The public-data construction retains at most 10 history items; extracting the trigger leaves at most 9, which are left-padded to 12 positions. The encoder uses left padding, so the most recent behaviour sits at the right end of the sequence and the tail injection is adjacent to it in RoPE terms. Since one leftmost padding position is dropped for the trigger, the encoder length stays at M=12M = 12 for every variant, so trigger-aware and trigger-free models never differ in sequence budget. On the industrial dataset, the trigger is repeated three times at the tail to strengthen its influence amid long histories containing many valid behavioural items; on the public dataset, the histories are short, so the trigger is inserted only once without repetition. Decoding is constrained by a trie of all valid Semantic IDs. On the industrial dataset, Semantic IDs consist of 8 levels with 300 codes each (vs. 3×2563 \times256 on the public dataset). Note that the industrial dataset is evaluated under a multi-positive recall setting, as a single trigger-induced session may contain multiple positive target items. Every reported number is the mean over five seeds, with standard deviations reported in Appendix § 11. Public-dataset experiments run on a single NVIDIA A100 GPU with PyTorch 2.4.0 and CUDA 12.1; the industrial models are trained on eight NVIDIA L20 GPUs.

ItemValue
Semantic ID3 levels ×\times 256 codes, token dim. 128
hidden / FFN128 / 320
layers2 encoder, 2 decoder, 4 attention heads
position enc.RoPE, θ=104\theta= 10^4, dropout 0.1
history length12, trigger at the tail
output headcosine softmax, τ=20\tau= 20, tied weights
fusionconvex, router bias 2.0
parameters1.51M
optimizerAdamW, lr 10−310^{-3}, weight decay 10−410^{-4}
schedule1000 warmup steps, cosine decay
batch / epochs256 / 15, gradient clipping 1.0
selectionbest validation Recall@20
inferencebeam search, size 20, trie-constrained

Table 6. Hyper-parameters of TriGen on the public dataset.

Statistical Significance and Standard Deviations

To assess reproducibility and statistical stability, all models are trained and evaluated across five independent runs with distinct random seeds. The main text reports the mean metrics across these runs in Table 1. Table 7 reports the corresponding standard deviations on both the industrial dataset and the Amazon benchmark.

CategoryMethodIndustrial DatasetAmazon Industrial_and_Scientific
R@5R@5R@10R@10N@5N@5N@10N@10R@5R@5R@10R@10N@5N@5N@10N@10
TIR-specificDIHN0.010.020.010.020.180.240.180.21
TIR-specificDIAN0.010.020.010.020.110.340.160.13
TIR-specificDEI2N0.020.030.010.020.140.250.080.11
SequentialGRU4Rec0.010.020.010.020.160.190.130.12
SequentialSASRec0.010.020.010.020.180.220.140.14
Generative (Trigger-free)HSTU0.020.020.010.020.130.180.120.14
Generative (Trigger-free)TIGER0.020.020.010.020.210.380.190.18
Generative (Trigger-free)LC-Rec0.010.020.010.020.290.150.130.09
Simple TIR-GRGR + trigger append0.020.030.010.020.350.260.220.16
Simple TIR-GRGR + trigger conditioning0.020.030.010.020.170.080.160.12
Simple TIR-GRGR + trigger-history attention0.020.030.020.030.390.350.320.28
OursTriGen0.020.030.010.020.240.310.180.15

Table 7. Standard deviations across 5 independent runs on the industrial dataset and the Amazon Industrial_and_Scientific benchmark. The corresponding means are reported in Table 1. R@KR@K and N@KN@K denote Recall@KK and NDCG@KK, respectively.

Given that the industrial evaluation is conducted over millions of user interactions, the empirical variances remain marginal (within ±0.01\pm0.01 to ±0.03\pm0.03), providing strong statistical confidence in the measured improvements. In contrast, standard deviations on the public dataset exhibit wider ranges due to its smaller sample size, which aligns with expected statistical sampling behavior. Across all runs, TriGen sustains consistent gains over both discriminative and generative baselines.

Limitations

The public trigger is simulated. No public dataset with Semantic-ID-ready item text exposes a real trigger field, so on Amazon the trigger is the last history item rather than an observed click. Baselines keep the full sequence, so this does not favour TriGen; it instead caps the trigger signal available to it, and the public gap should be read as conservative. Only the industrial dataset, where the trigger is an independent field, measures the mechanism at full strength.

Only a single trigger is modelled. TriGen assumes exactly one trigger per request, which matches the detail-page scenario but not surfaces where several recent signals arrive at once. Extending both injection points to a set of triggers, and how the fusion arbitrates among them, is left to future work.

References

  1. [1]Michele Bevilacqua, Giuseppe Ottaviano, Patrick Lewis, Wen-tau Yih, Sebastian Riedel, and Fabio Petroni. 2022. Autoregressive search engines: generating sub-strings as document identifiers. In Proceedings of the 36th International Conference on Neural Information Processing Systems (New Orleans, LA, USA) (NIPS ’22). Curran Associates Inc., Red Hook, NY, USA, Article 2296, 16 pages.
  2. [2]Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep Neural Networks for YouTube Recommendations. In Proceedings of the 10th ACM Conference on Recommender Systems (Boston, Massachusetts, USA) (RecSys ’16). Association for Computing Machinery, New York, NY, USA, 191–198. doi:10.1145/2959100.2959190DOI
  3. [3]Marco Cuturi. 2013. Sinkhorn distances: lightspeed computation of optimal transport. In Proceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2 (Lake Tahoe, Nevada) (NIPS’13). Curran Associates Inc., Red Hook, NY, USA, 2292–2300.
  4. [4]Jianxiong Deng, Shiyao Wang, Kuo Cai, Leijian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, and Guorui Zhou. 2025. OneRec: Unifying Retrieve and Rank with Generative Recommendation and Iterative Preference Alignment. arXiv:2502.18965 [cs.IR] https://arxiv.org/abs/2502.18965
  5. [5]Shijie Geng, Shuyang Wang, Luo Zihui, Fu Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5). In Proceedings of the 16th ACM Conference on Recommender Systems (Seattle, WA, USA) (RecSys ’22). Association for Computing Machinery, New York, NY, USA, 291–315. doi:10.1145/3523227.3546767
  6. [6]Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. 2016. Session-based Recommendations with Recurrent Neural Networks. arXiv:1511.06939 [cs.LG] https://arxiv.org/abs/1511.06939
  7. [7]Yupeng Hou, Jiangcheng Li, Ashley Shin, Jinsung Jeon, Abhishek Santhanam, Wei Shao, Kaveh Hassani, Ning Yao, and Julian McAuley. 2025. Generating Long Semantic IDs in Parallel for Recommendation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (Toronto ON, Canada) (KDD ’25). Association for Computing Machinery, New York, NY, USA, 956–966. doi:10.1145/3711896.3736979
  8. [8]Alexander Keller and Janina Kellein. 2002. Cumulated gain-based evaluation of IR techniques. ACM Trans. Inf. Syst. 20, 4 (Oct. 2002), 422–446. doi:10.1145/582415.582418
  9. [9]Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Recommendation. In 2018 IEEE International Conference on Data Mining (ICDM). 197–206. doi:10.1109/ICDM.2018.00035DOI
  10. [10]Doyup Lee, Chihyeon Kim, Saehon Kim, Minsu Cho, and Wook-Shin Han. 2022. Autoregressive Image Generation using Residual Quantization. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 11513–11522. doi:10.1109/CVPR52688.2022.01123DOI
  11. [11]Enze Liu, Bowen Zheng, Cheng Ling, Lantao Hu, Han Li, and Wayne Xin Zhao. 2025. Generative Recommender with End-to-End Learnable Item Tokenization. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (Padua, Italy) (SIGIR ’25). Association for Computing Machinery, New York, NY, USA, 729–739. doi:10.1145/3726302.3729989DOI
  12. [12]Han Liu, Yinwei Wei, Xuemeeng Song, Weili Guan-Fang Li, and Liqiang Nie. 2026. MMGRec: Multimodal Generative Recommendation with Transformer Model. ACM Trans. Multimedia Comput. Commun. Appl. 22, 10, Article 278 (Oct. 2026), 21 pages. doi:10.1145/3838594DOI
  13. [13]Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. In International Conference on Learning Representations. https://openreview.net/ forum?id=Bkg6RiCqY7
  14. [14]Jianmo Ni, Jiasheng Li, and Julian McAuley. 2019. Justifying Recommendations using Distantly-Labeled Reviews and Fine-Grained Aspects. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan (Eds.). Association for Computational Linguistics, Hong Kong, China, 188–197. doi:10.18653/v1/D19-1018DOI
  15. [15]Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res. 21, 1, Article 140 (Jan. 2020), 67 pages.arxiv.org/abs/1910.10683
  16. [16]Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Maheshwaran Sathiamoorthy. 2023. Recommender Systems with Generative Retrieval. In Thirty-seventh Conference on Neural Information Processing Systems. https://openreview.net/forum?id=BJQfQUU32warxiv.org/abs/2305.05065
  17. [17]Qijie Shen, Hong Wen, Wanjie Tao, Jing Zhang, Fuyu Lv, Zulong Chen, and Zhao Li. 2022. Deep Interest Highlight Network for Click-Through Rate Prediction in Trigger-Induced Recommendation. In Proceedings of the ACM Web Conference 2022 (Virtual Event, Lyon, France) (WWW ’22). Association for Computing Machinery, New York, NY, USA, 422–430. doi:10.1145/3485447.3511970DOI
  18. [18]Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. Roformer: Enhanced transformer with Rotary Position Embedding. Neurocomput. 568, C (Feb. 2024), 12 pages. doi:10.1016/j.neucom.2023.127063
  19. [19]Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (Beijing, China) (CIKM ’19). Association for Computing Machinery, New York, NY, USA, 1441–1450. doi:10.1145/3357384.3357895
  20. [20]Weiwei Sun, Lingyong Yan, Zheng Chen, Shuaixiang Wang, Haichao Zhu, Pengjie Ren, Zhumin Chen, Dawei Yin, Maarten de Rijke, and Zhaochun Ren. 2023. Learning to Tokenize for Generative Retrieval. arXiv:2304.04171 [cs.IR] https://arxiv.org/abs/2304.04171
  21. [21]Yi Tay, Vinh Q. Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W. Cohen, and Donald Metzler. 2022. Transformer memory as a differentiable search index. In Proceedings of the 36th International Conference on Neural Information Processing Systems (New Orleans, LA, USA) (NIPS ’22). Curran Associates Inc., Red Hook, NY, USA, Article 1587, 13 pages.
  22. [22]Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. 2017. Neural discrete representation learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems (Long Beach, California, USA) (NIPS ’17). Curran Associates Inc., Red Hook, NY, USA, 6309–6318.
  23. [23]Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. In Neural Information Processing Systems. https://api.semanticscholar.org/CorpusID:13756489
  24. [24]Wenjie Wang, Hongchao Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See-Kiong Ng, and Tat-Seng Chua. 2024. Learnable Item Tokenization for Generative Recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management (Boise, ID, USA) (CIKM ’24). Association for Computing Machinery, New York, NY, USA, 2400–2409. doi:10.1145/3627673.3679569DOI
  25. [25]Yujing Wang, Yingyan Hou, Haonan Wang, Ziming Miao, Shibin Wu, Hao Sun, Qi Chen, Yuqing Xia, Chengmin Chi, Guoshuai Zhao, Zheng Liu, Xing Xie, Hao Allen Sun, Weiwei Deng, Qi Zhang, and Mao Yang. 2022. A neural corpus indexer for document retrieval. In Proceedings of the 36th International Conference on Neural Information Processing Systems (New Orleans, LA, USA) (NIPS ’22). Curran Associates Inc., Red Hook, NY, USA, Article 1856, 15 pages.arxiv.org/abs/2206.02743
  26. [26]Ye Wang, Jiahao Xun, Minjie Hong, Jieming Zhu, Tao Jin, Wang Lin, Haoyuan Li, Linjun Li, Yan Xia, Zhou Zhao, and Zhenhua Dong. 2024. EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Barcelona, Spain) (KDD ’24). Association for Computing Machinery, New York, NY, USA, 3245–3254. doi:10.1145/3637528.3671775DOI
  27. [27]Yaxian Xia, Yi Cao, Shinao Hu, Tong Liu, and Lingling Lu. 2023. Deep Intention-Aware Network for Click-Through Rate Prediction. In Companion Proceedings of the ACM Web Conference 2023 (Austin, TX, USA) (WWW ’23 Companion). Association for Computing Machinery, New York, NY, USA, 533–537. doi:10.1145/3543873.3584661
  28. [28]Zhibo Xiao, Luwei Yang, Tao Zhang, Wen Jiang, Wei Ning, and Yujiu Yang. 2024. Deep Evolutional Instant Interest Network for CTR Prediction in Trigger-Induced Recommendation. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining (Merida, Mexico) (WSDM ’24). Association for Computing Machinery, New York, NY, USA, 846–854. doi:10.1145/3616855.3635829
  29. [29]Jun Yin, Zhengxin Zeng, Mingzheng Li, Hao Yan, Chaozhu Li, Weihao Han, Jianjin Zhang, Ruohen Liu, Hao Sun, Weiwei Deng, Feng Sun, Qi Zhang, Shirui Pan, and Senzhang Wang. 2025. Unleash LLMs Potential for Sequential Recommendation by Coordinating Dual Dynamic Index Mechanism. In Proceedings of the ACM on Web Conference 2025 (Sydney NSW, Australia) (WWW ’25). Association for Computing Machinery, New York, NY, USA, 216–227. doi:10.1145/3696410.3714866DOI
  30. [30]Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Jiayuan He, Yinghai Lu, and Yu Shi. 2024. Actions speak louder than words: trillion-parameter sequential transducers for generative recommendations. In Proceedings of the 41st International Conference on Machine Learning (Vienna, Austria) (ICML ’24). JMLR.org, Article 2414, 26 pages.
  31. [31]Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation. In 2024 IEEE 40th International International Conference on Data Engineering (ICDE), 1435–1448. doi:10.1109/ICDE60146.2024.00118DOI
  32. [32]Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Deep interest evolution network for click-through rate prediction. In Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence and Thirty-First Innovative Applications of Artificial Intelligence Conference and Ninth AAAI Symposium on Educational Advances in Artificial Intelligence (Honolulu, Hawaii, USA) (AAAI’19/IAAI’19/EAAI’19). AAAI Press, Article 729, 8 pages. doi:10.1609/aaai.v33i01.33015941
  33. [33]Guorui Zhou, Xiaoqiang Zhu, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep Interest Network for Click-Through Rate Prediction. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (London, United Kingdom) (KDD ’18). Association for Computing Machinery, New York, NY, USA, 1059–1068. doi:10.1145/3219819.3219823

Paper details

vlnc:W_8q4msrny
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

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