betaFeedback

WorksComputer Science

Keep it Simple! A Multiple-Choice Task can Benefit Generative Recommendation

Abstract

Sequential recommendation has recently witnessed growing interest in generative retrieval. One of the most popular approaches is to autoregressively generate the semantic identifiers (SIDs) of items as recommendations. Existing efforts mainly rely on semantic alignment, sophisticated architectural designs, and extensive hyperparameter tuning. However, the conventional learning framework for generative models could have foundational limitations. Specifically, most generative models directly predict target items without explicitly learning negative compositions and relative ranking signals. This is not well aligned with the philosophy of Bayesian personalized ranking (BPR), which emphasizes the importance of relative rankings between positive and negative samples. In this paper, we show that adding a simple multiple-choice auxiliary task into the learning objective can effectively address this issue. With minimal adaptation, we introduce Task Identifiers and Candidate Constructions to synthesize training data for the multiple-choice auxiliary task. In addition, we provide theoretical analysis of the fundamental benefits of the multiple-choice auxiliary task from the perspectives of information theory and compositional generalization. Extensive experiments on four publicly available benchmark datasets show that our approach can improve TIGER, a conventional generative recommendation model, by 4.9% to 79% in NDCG@10.

CCS Concepts

• Information systems → Recommender systems.

Introduction

Sequential recommendation [1, 4, 13, 17, 25, 27, 31, 32, 39] has long been a critical domain within recommendation systems. In general, these models recommend the next item by capturing the underlying intents within temporal sequences of user interactions. Due to the sequential nature, several recent studies [1, 4, 17, 25, 31, 32] focus on the generative paradigm. Within this space, a significant focus [6, 20, 25, 31, 38] has emerged on the use of Semantic Identifiers (SIDs). This usually includes two steps of development, SID Generation and Generative Retrieval. First, SIDs are generated by encoding item semantics with hierarchical coding, such as RQ-VAE [25, 35] or RK-Means [2, 12, 21]. Subsequently, the chronological items interaction is mapped into a sequence of item SIDs and fed into the generative recommender. The next item SID is then generated via the autoregressive process.

Based on this architecture, existing efforts focus on integrating and aligning collaborative information into language models [6, 16, 38], or maintaining more effective SID representations [6, 20, 31] to better align with recommenders. Despite their effectiveness, these models often rely on multiple alignment tasks, complex architectural designs, or extensive hyperparameter tuning. These issues could stem from the original design of the learning framework for generative models. First, the retrieval task involves substantial ambiguity when models are learned to generate the next item directly from the entire item corpus. In other words, models during training would only be asked to memorize fixed patterns. In particular, these models could suffer from out-of-distribution patterns, such as unseen compositions, as these patterns are rarely observed during training. Moreover, recent studies [31, 34] point out that the collaborative signals are beneficial to be added into the constructions of SIDs or model architectures, thereby bridging the gap between semantics and behavioral interaction patterns. However, it is arguable that the collaborative information can be naturally captured when training generative retrieval models on sequential data. From this aspect, retrieval models are expected to retain sufficient collaborative knowledge, without the need of introducing external models and extensive hyperparameter tuning.

These observations motivate us to revisit the design of the sequential training objective itself. In particular, we note that the current approach only defines the objective to align sequences with ground-truth target items, while leaving negative compositions unspecific. This idea is not well-aligned with Bayesian personalized ranking (BPR) [26], which emphasizes the importance of relative rankings between positive and negative samples. To mitigate these challenges, we propose to add a simple multiple-choice auxiliary task to the learning objective as a natural and effective remedy. Specifically, the multiple-choice auxiliary task can explicitly present a subset of candidate items to encourage models to learn both positive and negative signals. This additional task can be seamlessly fit into the original sequential models, such as TIGER [25], even without relying on pretrained language weights. The whole framework remains simple and elegant without introducing any complicated training strategy.

In this paper, we propose to integrate the multiple-choice task into generative sequential recommendation. Extensive experiments validate the effectiveness and investigate the underlying principles about how this auxiliary task addresses the aforementioned limitations. Our contributions can be summarized as follows:

  • We provide a simple manner to integrate the multiple-choice auxiliary task into generative recommendation models, along with theoretical analysis of its advantages.

  • We show that the multiple-choice auxiliary task can successfully mitigate the limitations and achieve better recommendation performance.

  • We conduct a series of in-depth analysis, including the exploration of the underlying factors behind the performance, the relative importance of the main sequential task and the auxiliary task, and the impacts of different candidate sampling strategies.

Related Work

Generative Recommendation System

Generative recommendation has been a widely studied area [4, 10, 17, 20, 25, 31, 32, 38]. In recent years, research has increasingly focused on representing items using semantic identifiers (SIDs), which are constructed through vector quantization (e.g., RQ-VAE), as first proposed by TIGER [25]. Following this line of work, LC-Rec [38] integrates the ideas of SID on LLMs, and designs tasks to bridge the gap between language and recommendation information. LETTER [31] formulates better SID representations by considering collaborative signals and controlling the code embeddings diversity. ETEGRec [20] learns SIDs in an end-to-end manner, allowing the item tokenizer and recommender to be jointly optimized through the alignment between sequences and items, as well as preferences and semantics.

These approaches have been shown to improve the generative recommendation system. However, they often rely on multiple alignment objectives between language, collaborative signals, and SID representations, which introduce additional hyperparameters and therefore increase optimization complexity. Instead, our work directly utilizes the natural property of sequence-to-sequence model architectures to bridge the gaps between pretrained SIDs and collaborative knowledge.

Auxiliary Tasks on Sequential Recommendation

Recently, auxiliary tasks have been widely adopted in sequential recommendation systems to enhance model performance and representation learning. S3S^3-Rec [39] and UniSRec [9] leverage auxiliary tasks, such as cloze, attributes matching, and sequence-sequence contrastive learning, for model pretraining. Several LLM-based recommenders focus on multi-task learning (MTL) [4, 17], or design auxiliary tasks to align recommendation knowledge with language semantics [1, 16, 32, 38]. Among them, P5 [4] expresses items as their naive IDs and includes the multiple-choice task as one of the primary objectives. [1] extends P5 by introducing additional auxiliary tasks, such as the BPR task, and represents items using both naive IDs and titles. On the other hand, [16] focuses on SID-based item representations and constructs the negative sampling task to learn from items that users are unlikely to interact with.

Despite their effectiveness, prior studies either rely on language-based item identifiers or introduce negative signals only through the training labels without explicit comparison between positive and negative items. In our work, we identify the lack of comparative signals between items in recommendation sequences and investigate how the multiple-choice task can benefit the SID-based generative recommendation system. Notably, unlike previous works, our proposed framework does not rely on pretrained language weights during generative process.

we compare our work with several related studies that contain the idea of multiple-choice task or negative sampling in Table 1.

MethodSID-basedExplicit ComparisonLLM Free
P5 [4]✗✓✗
Cao et al. [1]✗✓✗
Li et al. [16]✓✗✗
Ours✓✓✓

Table 1. Comparison of our work with related studies that contain the multiple-choice task or negative sampling, where “Explicit Comparison” means the model explicitly contrasts positive items against negative ones, and “LLM Free” indicates the proposed method does not rely on pretrained language weights during generative process.

Compositional Generalization

In various domains, such as Natural Language Processing (NLP) and Computer Vision (CV), compositional generalization is increasingly recognized as a fundamental research problem [3, 11, 14, 15, 19, 24, 33, 40]. Generally speaking, these studies focus on generalizing the learned model to out-of-distribution (OOD) scenario from a sufficient number of compositions in the training distribution. For example, [11] proposes mutual exclusive training to augment the number of unique primitives to mitigate compositional challenges. [15] introduces a retrieval-augmented framework aimed at improving the multi-sourced compositional generalization in the field of visual question answering. [33] formally defines the compositional generalization, and provides a theorem of when a function can be generalized.

Despite its significant advancements in various domains, we observe that research into compositional generalization remains underdeveloped in sequential recommendation systems. In our study, we bridge this gap by proposing a simple multiple-choice auxiliary task can serve as a signal to facilitate compositional generalization.

Proposed Framework

In this section, we formally define the problem scope in this work and then present our proposed framework along with detailed analysis. Figure 1 illustrates the overall framework.

The overall framework integrating the multiple-choice task as an auxiliary

Figure 1. The overall framework of our approach to integrate the multiple-choice task as an auxiliary. In this illustration, we assume a candidate size of nc=3n_c = 3.

Problem Formulation & Preliminary

In this paper, we focus on the sequential recommendation task. Formally, let u∈Uu \in\mathcal{U} denote a user and iu,t∈Ii_{u,t} \in\mathcal{I} represent the target next item with which the user uu interacts at time tt. Let g(⋅)g(\cdot) be a representation function that maps the user’s interaction history sequence, su=[iu,1,iu,2,…,iu,t−1]∈Sus_u = [i_{u,1}, i_{u,2}, \ldots, i_{u,t-1}] \in\mathcal{S}_u, to a latent representation hu,t∈Rdh_{u,t} \in\mathbb{R}^d. The objective of the recommendation system is to learn a well-behaved representation hu,t=g(su)h_{u,t} = g(s_u) and a scoring function s(h,i)s(h,i) such that the ground-truth next item iu,ti_{u,t} is ranked as high as possible.

For generative models [20, 25, 31], SIDs are constructed from a sequence of semantic tokens based on their semantic meaning. In this paper, we focus on using RQ-VAE [25, 35] to obtain semantic tokens:

RQ-VAE⁡(iut)=(qut1,qut2,…,qutm),(1)\operatorname{RQ\text{-}VAE}(i_{u_t}) = (q_{u_t}^{1}, q_{u_t}^{2}, \ldots, q_{u_t}^{m}), \tag*{(1)}

where qutlq_{u_t}^{l} denotes the token at level ll. Each item is represented by the concatenation of its corresponding semantic tokens to form its SID, and the user sequence suts_{u_t} is then transformed into the concatenation of the SIDs of interacted items. The learning objective Lmain\mathcal{L}_{\mathrm{main}} is formulated as the negative log-likelihood:

Lmain=−∑sut∈Su∑l=1mlog⁡P(qutl∣qut<l,sut),(2)\mathcal{L}_{\mathrm{main}} = - \sum_{s_{u_t} \in\mathcal{S}_u} \sum_{l=1}^{m} \log P(q_{u_t}^{l} \mid q_{u_t}^{<l}, s_{u_t}), \tag*{(2)}

where qut<lq_{u_t}^{<l} denotes the semantic tokens before qutlq_{u_t}^{l}. The score function s(hut,i)s(h_{u_t}, i) is then modeled as the beam score in the autoregressive generation process with beam search.

Multiple-Choice Auxiliary Task

We design the multiple-choice auxiliary task by extending the representation function gg to additionally condition on the candidate set cutc_{u_t}.

g(sut)→g(sut,cut),iut∈cut.(3)g(s_{u_t}) \rightarrow g(s_{u_t}, c_{u_t}), \quad i_{u_t} \in c_{u_t}. \tag*{(3)}

The multiple-choice task is introduced as an auxiliary to the main sequential recommendation task where both tasks share the same representation function gg. For the main task, the candidate set is not used, and the model reduces to the original formulation g(sut)g(s_{u_t}).

Task Identifiers

To distinguish between the two tasks, we introduce task identifiers along with a separator token. We incorporate a mechanism similar to prefix-tuning [18], by prepending a learnable header token, p1p_1 or p2p_2, in front of the user history sequence, where p1p_1 and p2p_2 denote the main task and the auxiliary task, respectively. Also, a separating token p3p_3 is introduced to mark the boundary between the history sequence and the candidate set.

Inputs={[p1 sut],for the main task[p2 sut p3 cut],for the auxiliary task.(4)\mathit{Inputs} = \begin{cases} [p_1\ s_{u_t}], & \text{for the main task} \\ [p_2\ s_{u_t}\ p_3\ c_{u_t}], & \text{for the auxiliary task}. \end{cases} \tag*{(4)}

The three special tokens, p1p_1, p2p_2 and p3p_3, possess their own independent token embeddings, and the models are expected to learn suitable representations that help distinguish the tasks and separate the history sequence from the candidates.

Candidate Construction for Auxiliary Task

The candidate set cutc_{u_t} with ncn_c items is formed by combining the ground truth item iuti_{u_t} and nc−1n_c-1 negative examples uniformly sampled from any unseen items as:

cut=negut∪{iut}c_{u_t} = neg_{u_t} \cup\{i_{u_t}\}
negut∼I∖{iut}.(5)neg_{u_t} \sim\mathcal{I} \setminus\{i_{u_t}\}. \tag*{(5)}

After sampling, the candidate items are randomly shuffled and then arranged in a sequential order, one item at a time, cut=(c1c2…cnc)1c_{u_t} = (c_1 c_2 \ldots c_{n_c})^1. The goal of the auxiliary task is to select the positive item iuti_{u_t} from the candidate set cutc_{u_t}. The candidate set is dynamically refreshed for each sequence suts_{u_t} in each epoch. In a special case when nc=1n_c = 1, negut=∅neg_{u_t} = \emptyset is an empty set. Therefore, the objective can be reinterpreted as an answer-memorization task.

Training Objective

The multiple-choice auxiliary task can be naturally formulated as sequence-to-sequence problems, conditioned both on the historical sequence and the candidate set. Therefore, we can extend the training objective in Equation (2) as:

Laux=−∑sut∈Su∑l=1mlog⁡P(qutl∣qut<l,sut,cut).(6)\mathcal{L}_{\mathrm{aux}} = - \sum_{s_{u_t} \in\mathcal{S}_u} \sum_{l=1}^{m} \log P(q_{u_t}^{l} \mid q_{u_t}^{<l}, s_{u_t}, c_{u_t}). \tag*{(6)}

It is noteworthy that the design of the auxiliary task shares the same decoding targets (i.e., the ground truth next item) as the main task, which helps minimize potential objective conflicts.

Combining the original main task objective Lmain\mathcal{L}_{\mathrm{main}}, the overall training loss L\mathcal{L} is the sum of both:

L=Lmain+Laux.(7)\mathcal{L} = \mathcal{L}_{\mathrm{main}} + \mathcal{L}_{\mathrm{aux}}. \tag*{(7)}

In a training epoch, each training sequence (sut)(s_{u_t}) is used once for the main task and once for the auxiliary task. Each batch is then randomly constructed from the combined training samples, regardless of the task type.

Theoretical Analysis

Here we examine the advantages and fundamental benefits brought by the multiple-choice auxiliary task.

The Viewpoint of Information Theory

First, we start to analyze the advantage from the perspective of information theory. Before integrating the auxiliary task, the model is trained to maximize the mutual information I(G;I+)I(G; I^{+}) between its latent representations G=g(S)G = g(S) and the target next-item I+I^{+}. Here, we denote SS and I+I^{+} as the random variables of the historical sequence and the target positive item, respectively.

When incorporating the candidate set CC, the latent representation is extended to G=g(S,C)G = g(S,C). For nc>1n_c > 1, the candidate set contains negative items, C={I+,I−}={I+,I1−,…,Inc−1−}C = \{I^{+}, I^{-}\} = \{I^{+}, I^{-}_{1}, \ldots, I^{-}_{n_c-1}\}. The model is asked to select I+I^{+} instead of all other negative items in I−I^{-}. That is, we implicitly impose an information bottleneck [29] to maximize the mutual information with I+I^{+} while minimizing the mutual information with I−I^{-} as:

max⁡g[I(G;I+∣C)−βI(G;I−∣C)],(8)\max_g [ I(G; I^{+} \mid C) - \beta I(G; I^{-} \mid C) ], \tag*{(8)}

where β\beta is an implicit Lagrange multiplier. Equation 8 aligns with the philosophy of BPR [26] ranking, as the latent representation is learned to prefer positive item to negative items. Since we still preserve the main task, whose target is to maximize I(G;I+)I(G; I^{+}), we may further assume that the model will also converge to maximize I(G;I+∣C)I(G; I^{+} \mid C). In this case, we naturally have

βI(G;I−∣C)≪I(G;I+∣C),(9)\beta I(G; I^{-} \mid C) \ll I(G; I^{+} \mid C), \tag*{(9)}

as the model starts to converge.

From (9), we can regard negative items as small noise perturbation. To remain consistent predictions across different negative items, gg should be robust. This perspective is particularly useful in recommendation, where the user sequence may contain items that do not perfectly match the user’s current intent.

Facilitating Compositional Generalization

Beyond the perspective of information theory, incorporating candidate items as inputs introduces a natural structure that facilitates compositional generalization.

In the generative retrieval process, the training distribution of the main task is highly sparse, where many item compositions rarely or never co-occur. Consequently, the model struggles to capture the underlying behaviors of unseen compositions. For instance, users interested in guitars are more likely to purchase relevant accessories rather than items related to orchestra. This lack of joint observations for compositions between guitars and orchestra could lead to weak generalization on such unseen combinations during testing. The situation can potentially be even worse, when items share similar semantic information but exhibit different behavioral patterns. The ill-defined representation function gg may induce unintended correlations. Since the training distribution lacks the composition for these unrelated items, predefined semantics force them to share a similar representation, leading to incorrect recommendations.

The multiple-choice task then augments the training distribution by providing candidate items, and asking the model to pick the correct one. Through this process, the model starts to capture the compositions between items driven by different intents, which facilitates the learning of a more well-defined representation function to generalize from the training distribution to the test distribution. In particular, since iu∈cu,ti_u \in c_{u,t}, the auxiliary task forces the model to memorize the positive signal between history su,ts_{u,t} and target iu,ti_{u,t}. This is crucial because the training labels could become part of the input sequence. While the previous sequential task constructions often overlook these positive compositions, the multiple-choice task can also effectively bridge the gap.

Dataset# Users# ItemsHistory Length per userSparsity
Scientific50,98525,8488.10699.97 %
Instruments57,43924,5878.91799.96 %
Yelp30,43120,03310.40799.95 %
MovieLens 1M6,0403,416165.509695.16 %

Table 2. Dataset statistics for the four real-world benchmarks.

Experiments

In this section, we present our experimental setup and results on four real-world datasets. The experiments are designed to answer the following research questions (RQs):

  • RQ1: How does the multiple-choice auxiliary task affect the performance compared to standard baselines, and how does it vary across different values of ncn_c?

  • RQ2: How does the model perform across various scenarios, and what key factors drive its behavior?

  • RQ3: What is the relative importance of the two tasks during the training process, and what is the impact of different candidate sampling methods?

  • RQ4: How does the multiple-choice task behave in our setting?

Experimental Settings

Datasets

We consider four real-world datasets across various kinds of sequential recommendation scenarios.

  • Amazon Industrial and Scientific and Musical Instruments are from the Amazon’23 review dataset [8], including interactions ranging from 1996 to 2023. This dataset is notable for its high sparsity. For simplicity, we refer to these two datasets as Scientific and Instruments, respectively.

  • Yelp2 is a widely used dataset for business recommendation. We adopt the item metadata and interaction sequences as preprocessed by LETTER [31] for our experiments.

  • MovieLens1M [5] is the 1-million-interaction dataset derived from user interactions with the MovieLens online recommender system. The user-item interactions are denser compared to Amazon and Yelp datasets.

For preprocessing, following previous works [31, 38], standard 5-core filtering and leave-one-out strategy are applied. Table 2 summarizes the statistics of each dataset.

Implementation Details

To derive SIDs, we apply pretrained Sentence-T5 [23] (sentence-t5-xl) to generate semantic embeddings for RQ-VAE. The layer number of RQ-VAE codes is set to 4, with codebook size in each level being 256, across all datasets. The three special tokens p1p_1, p2p_2 and p3p_3 are introduced as three independent new tokens. The learning rate is set to 5×10−45 \times10^{-4}, and we set the maximum history length to 20, following previous generative recommendation models [31, 38]. The number of epochs is set to 200, with early stopping patience of 20 based only on the main task validation loss. During generation, constrained beam search is applied with beam size 20.

Baseline Methods

We compare our framework with two categories of conventional baseline methods as follows:

  1. Traditional Baselines:

  • Caser [28] applies horizontal and vertical convolution layers to capture union-level and point-level sequential patterns.

  • Bert4Rec [27] performs sequential recommendation on a bidirectional transformer architecture with mask cloze task as its objective.

  • GRU4Rec [7] uses RNN to capture sequential patterns.

  • SASRec [13] utilizes self-attention mechanism to model sequences and predict the next item.

  • HGN [22] applies a hierarchical gating network to capture both the user long-term and short-term interests.

  1. Generative Baselines:

  • TIGER [25] uses text embedding from metadata information to construct SIDs via RQ-VAE, and generates the next item via an encoder-decoder architecture. We consider three variants with different SID constructions. TIGER and TIGER-SAS utilize pretrained Sentence-T5 [23] embeddings and fine-tuned embeddings learned from SASRec [13] to derive SIDs. LETTER-TIGER [31] integrates collaborative and diversity regularization to learn the SID tokenizer.

  • ETEGRec [20] integrates the item tokenizer and generative retrieval into an end-to-end framework, based on the approach of recommendation-oriented alignment.

For traditional baselines, we adopt the open-source recommendation framework RecBole [36, 37] to implement all methods. RecBole adopts cross-entropy loss for these baseline methods, leading to stronger and more competitive baselines. For generative baselines, TIGER and TIGER-SAS follow the same SID construction method and implementation setup as described in Section 4.1.2 for a fair comparison.3 For LETTER-TIGER [31], we use LLaMA-2-7B [30] to construct the semantic embeddings, while other details follow the official implementation. For ETEGRec, we obtain the SASRec semantic embeddings using RecBole and subsequently follow the official implementation.

Overall Performance (RQ1)

In this section, we evaluate the overall performance of our approach and baselines to investigate the effectiveness of the auxiliary task to answer RQ1.

We first evaluate the performance differences against baselines. We consider two settings: one with a single negative and one positive item (TIGER+nc=2n_c = 2), and the other with 10 candidate items (TIGER+nc=10n_c = 10). We use common recommendation metrics, Hit@K and NDCG@K, abbreviated as H@K and N@K, with K∈{5,10}K \in\{5, 10\}. The results are reported in Table 3. Detailed comparison with multiple runs across different seeds are shown in Table 9 in Appendix A.

MethodsScientific H@5Scientific N@5Scientific H@10Scientific N@10Instruments H@5Instruments N@5Instruments H@10Instruments N@10Yelp H@5Yelp N@5Yelp H@10Yelp N@10MovieLens H@5MovieLens N@5MovieLens H@10MovieLens N@10
Caser0.01290.00820.02180.01100.02160.01340.03600.01800.02910.02080.04030.02440.13540.08730.21230.1120
Bert4Rec0.01440.00920.02390.01220.02570.01640.04070.02120.01670.01050.02780.01410.15350.09880.24490.1283
GRU4Rec0.01990.01270.03190.01660.03010.01930.04750.02500.02250.01440.03840.01950.16950.11240.25150.1386
SASRec0.02500.01470.03910.01930.03350.02100.05260.02720.04350.03250.06130.03830.14300.08980.22620.1164
HGN0.02070.01300.03460.01740.03050.01950.04840.02520.03670.02530.05400.03080.07580.04810.12800.0649
TIGER0.02550.01670.03980.02130.03530.02280.05540.02930.02430.01580.03900.02050.18110.12580.26270.1520
TIGER-SAS0.02270.01480.03630.01920.03350.02220.05230.02820.02560.01610.04410.02200.19090.13300.27430.1598
LETTER-TIGER0.02550.01640.04010.02110.03540.02300.05550.02940.02890.01860.04660.02430.19340.13260.27780.1597
ETEGRec0.02690.01720.04250.02220.03850.02500.05980.03180.03360.02140.05400.02790.19590.13250.27900.1594
TIGER+ncn_c = 20.0274*0.0179*0.0426*0.0228*0.0385*0.0253*0.0587*0.0318*0.0320*0.0225*0.0489*0.0279*0.1950*0.1338*0.2742*0.1594*
TIGER+ncn_c = 100.0284*0.0184*0.0442*0.0235*0.0385*0.0256*0.0594*0.0323*0.0444*0.0310*0.0621*0.0367*0.1921*0.1314*0.2770*0.1586*
Gain (v.s TIGER)11.4 %10.2 %11.1 %10.3 %9.1 %12.3 %7.2 %10.2 %82.7 %96.2 %59.2 %79.0 %7.7 %6.4 %5.4 %4.9 %

Table 3. Table 3: Overall Performance Comparison. We denote our approach as “TIGER+ncn_c”, where ncn_c is the number of candidates. Bold and underlined font are used to denote the best and second-best results. * denotes a significant improvement (p<0.05p < 0.05) compared to TIGER via paired t-test on per-user metrics.

ModelDatasetsHit@10NDCG@10
TIGERScientific0.0400 ± 0.00080.0213 ± 0.0004
Instruments0.0536 ± 0.00130.0284 ± 0.0006
Yelp0.0352 ± 0.00290.0182 ± 0.0016
MovieLens0.2630 ± 0.00500.1517 ± 0.0028
TIGER+nc=2n_c = 2Scientific0.0442 ± 0.00120.0238 ± 0.0007
Instruments0.0581 ± 0.00060.0311 ± 0.0004
Yelp0.0498 ± 0.00380.0282 ± 0.0024
MovieLens0.2738 ± 0.00410.1582 ± 0.0026

Table 9. The mean and standard deviation of the performance across different dataset for TIGER and TIGER+nc=2n_c = 2. We computed 5 runs using different random seeds. The improvements from TIGER+nc=2n_c = 2 to TIGER are all significant (p<0.05p < 0.05) via Welch’s t-test on per-run performance.

From Table 3, we observe that incorporating the multiple-choice task as an auxiliary objective yields significant performance improvements over the base model TIGER across all datasets. Additionally, the performance is competitive with mainstream generative models, even achieving the best ranking among all baselines on the Scientific datasets. Except for MovieLens, “TIGER+nc=10n_c = 10” consistently ranks among the best or second-best methods. Notably, our approach remains simple, interpretable, and does not rely on extensive hyperparameter tuning. We attribute this improvement to a more robust (Section 3.3.1) and well-defined (Section 3.3.2) representation qq, resulting from the integration of the multiple-choice auxiliary task.

Impact of Candidate Number. Subsequently, we conduct the performance study across different values of ncn_c. As shown in Figure 2, we observe a trend of performance improvement as ncn_c increases. When ncn_c becomes excessively large, the performance begins to saturate. This suggests that the model may have already captured sufficient compositional patterns among items. Additional negative signals could become redundant, and a larger number of candidates may increase the likelihood of including false negatives, defined as items unseen by the users, but aligned with their intents, thus resulting in a performance drop.

Figure 2. Performance (Hit@10 and NDCG@10) vs. Number of Candidates (ncn_c). Here, nc=0n_c = 0 indicates the performance of base model TIGER specifically.

In particular, when nc=1n_c = 1, we still observe improvements on several datasets. Although the gains are generally less significant, this suggests that explicitly learning compositional patterns between the training label iui_u and its corresponding sequence sus_u can already benefit the models at test time.

In-depth Analysis and Discussion

Model Confidence on Ground-Truth Labels (RQ2)

To understand how confidently our models perform on all testing samples, we start the in-depth analysis by probing into its likelihood of generating ground-truth next items. As shown in Figure 3, when ncn_c increases, we observe an overall improvement of median likelihood values across different datasets. As ncn_c continues to grow, the likelihood values eventually begin to saturate. The growth trends align with the performance observed earlier (Figure 2). The results suggest that the model develops a higher confidence regarding the ground-truth items in the majority of testing samples.

Figure 3. Log likelihood vs Number of Candidates (ncn_c) across different datasets. The log-likelihood is computed via the teacher forcing decoding. We aggregate the values from the first semantic token to EOS. Here, nc=0n_c = 0 means the base model TIGER specifically.

Predicting Items with Different Frequency (RQ2)

Next, we examine the performance across items with different interaction frequency. Specifically, we evaluate on three test subsets, “popular,” “medium,” and “long-tail,” where the ground-truth items (iuti_{u_t}) have a training interaction frequency that falls within the top, middle and bottom one-third among all items. The data statistics are documented in Table 10 in Appendix A.

DatasetScientific InstrumentsYelpMovieLens
popular2983539632183364518
medium9456715652871198
long-tail11694106516808324

Table 10. The number of users in different splits based on the frequency of ground-truth items (§ 4.5.2).

The evaluation results are shown in Table 4. Incorporating the multiple-choice auxiliary task consistently improves performance in popular settings, while causing no significant degradation in medium and long-tail subsets. This suggests that the auxiliary task enhances the model’s discriminative ability to distinguish easier ground-truth items from other negative samples, without confusing harder cases. Even on Yelp and MovieLens, we observe performance improvements in the medium and long-tail subsets. For Scientific and Instruments, the performance gains are more modest. This could be due to a greater difficulty of the medium and long-tail scenarios in these datasets.

DatasetsSplitsTIGER H@10TIGER N@10TIGER + nc=10n_c = 10 H@10TIGER + nc=10n_c = 10 N@10
Scientificpopular0.06710.03590.07410.0396
medium0.00250.00100.00370.0016
long-tail0.00040.00020.00040.0001
Instrumentspopular0.07960.04210.08550.0466
medium0.00280.00110.00240.0012
long-tail0.00070.00030.00090.0004
Yelppopular0.06130.03190.07680.0420
medium0.00700.00410.03540.0245
long-tail0.00380.00230.04320.0317
MovieLenspopular0.30430.17940.32090.1866
medium0.16780.08550.17700.0908
long-tail0.03400.01550.03400.0194

Table 4. Performance under different sequence splits based on the popularity of ground-truth items. The popular, medium, and long-tail splits are categorized according to whether the ground-truth next item’s interaction frequency in the training set falls into the top, middle, or bottom one-third of all items.

Performance on Different Sequence Length (RQ2)

In this section, we further examine the performance difference across users with different sequence lengths. As shown in Figure 4, we partition user sequences into short, medium, and long categories by dividing them into three bins based on sequence length, and compare their performance against the base model TIGER. Among all datasets, incorporating the candidate set benefits performance for users with different sequence lengths, without exhibiting a clear bias toward any specific group.

Figure 4. Performance difference across user sequence lengths. “Short”, “Medium”, and “Long” correspond to users whose sequence lengths fall in the bottom third, middle third, and top third in the dataset, respectively. NDCG@10 is reported.

From Figure 4, we also observe that the performance consistently declines as the sequence length increases for both TIGER and “TIGER+nc=10n_c = 10”. This is possibly driven by the increasing complexity and diversity of user intents in longer interaction histories. Notably, our framework still achieves the same degree of improvement over TIGER in these challenging settings, suggesting its robustness to long sequence scenarios.

Illustrative Examples (RQ2)

Apart from the overall performance comparison, we further present two illustrative examples to provide qualitative insights into how real-world datasets benefit from the auxiliary task, shown in Table 5 and Table 6.

Interaction History1. Dangerous Liaisons (Drama, Romance)<br>2. Babe (Children’s, Comedy, Drama)<br>3. Trainspotting (Drama)<br>4. Color Purple (Drama)<br>5. Papillon (Drama)
TIGER(1) Rain Man (Drama)<br>(2) Apollo 13 (Drama)<br>(3) Field of Dreams (Drama)
TIGER + nc=10n_c = 10(1) E.T. the Extra-Terrestrial (Children’s, Drama, Fantasy, Sci-Fi)<br>(2) Apollo 13 (Drama)<br>(3) It’s a Wonderful Life (Drama)

Table 5. An example of a correct prediction in our approach from MovieLens dataset. The interaction history is presented sequentially, with only the most recent five items displayed. We present the top-3 predictions for this sequence, where the ground-truth item is highlighted in bold.

1. Yamaha YPG-235 76-Key Portable Grand Piano
Interaction History2. Yamaha OEM PKBB1 Adjustable Padded Keyboard X-Style Bench, Black, 19.5 inches
3. On-Stage KS7191 Classic Double-X Keyboard Stand
4. LyxPro HAS-30 Closed Back Over-Ear Professional Recording Headphones for Studio Monitoring, DJ and Home Entertainment,Black
TIGER(1) Yamaha OEM PKBB1 Adjustable Padded Keyboard X-Style Bench, Black, 19.5 inches
(2) M-Audio SP-2 - Universal Sustain Pedal with Piano Style Action For MIDI Keyboards, Digital Pianos & More
(3) RockJam Xfinity Heavy-Duty, Double-X, Pre- Assembled, Infinitely Adjustable Piano Keyboard Stand with Locking Straps
TIGER + nc=10n_c = 10(1) Blue Yeti USB Microphone for PC, Mac, Gaming, Recording, Streaming, Podcasting, Studio and Computer Condenser Mic with Blue VO!CE effects, 4 Pickup Patterns, Plug and Play – Blackout
(2) Enabled Dragonpad USA- 6" Microphone Studio Pop Filter with Clamp - BLK/WHT
(3) M-Audio SP-2 - Universal Sustain Pedal with Piano Style Action For MIDI Keyboards, Digital Pianos & More

Table 6. An example of a correct prediction in our approach from Instruments dataset. The interaction history is presented sequentially. We present the top-3 predictions for this sequence, where the ground-truth item is highlighted in bold.

In Table 5, the user has watched movies that are mostly associated with the Drama genre. As a result, TIGER produces three recommendations that are purely within the Drama category. However, since a Children’s movie appears in the recent interaction history, our approach captures this signal in the user intent and

additionally recommends a movie related to both the Children’s and Drama categories.

Another example is shown in Table 6, where the user interaction history consists of a piano, a keyboard, a keyboard stand, and recording headphones. Based on the sequence, original TIGER chooses to recommend keyboard-related accessories, which are semantically similar to the observed history. In contrast, TIGER with the auxiliary task predicts microphone and its accessory as top-2 choices, which are absent from the interaction history, but consistent with the user intent.

These two examples suggest that the auxiliary task helps the model capture a higher level behavior intent, instead of relying solely on metadata semantic similarity between items.

Relative Importance of the Two Tasks (RQ3)

In our main experimental setup, each training sequence participates in both the main and auxiliary tasks within each training epoch, resulting in an identical proportion for both tasks. To further investigate the relative importance of the two tasks, we consider that each sample is assigned to either the main or auxiliary task via Bernoulli sampling, where the main task is selected with probability pp, which corresponds to the main task proportion. A larger value of pp indicates the model has greater emphasis on the main task.

As shown in Figure 5, we observe that the performance decreases across the three examined datasets when the main task proportion is disproportionately higher than that of the auxiliary task. Moreover, on the Scientific and Yelp datasets, the performance also decreases when the main task proportion is substantially lower. These observations indicate that the main task proportion should not be too high, while an excessively low proportion may also be harmful for certain domains. Although the optimal value may vary across different models and datasets, our findings suggest that balancing the two tasks during the learning process remains the most effective approach for achieving good performance on new datasets.

Figure 5. The performance of different task proportion between the main task and the auxiliary task during training. The number of candidate set ncn_c is fixed as 10. Our default deterministic setting corresponds to an equal proportion between the two tasks, and is indicated by the star markers as reference points (blue as Hit@10 and red as NDCG@10).

Different Candidate Sampling Strategies (RQ3)

In our main setting, we adopt uniform sampling to construct the candidate set. To better understand how different sampling strategies affect model behavior, we consider four different candidate construction settings for the in-depth analysis:

  • Uniform: Negative candidates are derived by the original uniform sampling strategy.

  • Long-tail: Negative candidates consist entirely of long-tail items with the same definition as in Section 4.3.2.

  • Warm-set: Negative candidates consist entirely of warm-set items, defined as non-long-tail items.

  • Dynamic: Negative candidates are sampled using a dynamic strategy that gradually shifts from long-tail items to warm-set items. Specifically, the Bernoulli probability of sampling from long-tail items is p=e−ep/Tp=e^{-e_p/T}, with the corresponding probability of sampling from warm-set items given by 1−p1-p, where epe_p represents the current epoch and TT is a hyperparameter, set to 100.

As shown in Table 7, except for the dense MovieLens dataset, we observe that long-tail sampling achieves comparable performance to uniform or warm-set sampling. This indicates that incorporating the multiple-choice auxiliary task helps the model establish better comparative relationships between sequences and long-tail items, and that the learned compositions may further generalize to other negative signals. On the other hand, on the MovieLens dataset, warm-set or uniform sampling yields better performance than long-tail sampling. This could be because the main task training sequences already contain sufficient long-tail items, allowing the model to learn the comparative signals between positive items and long-tail negatives. However, the model is more likely to confuse positive items with warm-set negatives when it is trained only on the main task. Therefore, separating these warm-set negatives becomes more crucial.

MethodsScientific Hit@10Scientific NDCG@10Instruments Hit@10Instruments NDCG@10Yelp Hit@10Yelp NDCG@10MovieLens Hit@10MovieLens NDCG@10
Uniform0.04420.02350.05940.03230.06210.03670.27700.1586
Long-tail0.04450.02390.05940.03210.06290.03680.26640.1517
Warm-set0.04520.02430.05880.03180.06040.03530.27370.1588
Dynamic0.04580.02470.05900.03200.06230.03640.27400.1602

Table 7. The overall performance for different candidate sampling strategies. (nc=10n_c = 10)

In our experiments, we also examine the dynamic setting by gradually shifting from long-tail sampling to warm-set sampling. Under this strategy, we observe that introducing warm-set negatives in the later stages can successfully improve the performance over pure long-tail sampling on the MovieLens dataset. However, across all datasets, the resulting performance does not yield significant gains compared to the uniform setting. This finding suggests that the benefits from dynamic scheduling are marginal, and a static uniform sampling strategy may already be sufficient.

Investigation on the Multiple-Choice Task (RQ4)

Since we treat the multiple-choice task as an auxiliary objective and do not include it in the validation phase (§ 4.2), it is interesting to investigate whether the model develops the ability to perform the multiple-choice task, and how it balances information between the interaction history and the candidate items. In this section, we further analyze these questions through the model’s selection capability and its cross-attention behavior in the auxiliary task. To this end, we construct candidate sets of size nc=2n_c=2 and nc=10n_c=10 for each test instance.

For the selection capability, we measure whether the top-1 predicted item lies within the candidate set. As shown in Table 8, we observe that the prediction accuracy across all datasets for both

DatasetsScientific InstrumentsYelpMovieLens
TIGER+nc=2n_c = 21.000001.000000.999970.99983
TIGER+nc=10n_c = 101.000001.000000.999841.00000

Table 8. Accuracy of whether the top-1 predicted item lies within the candidate set.

nc=2n_c = 2 and nc=10n_c = 10 are equal or close to 1.0. This indicates that the model has effectively learned to conduct the multiple-choice selection when needed.

For the cross-attention analysis, we analyze how the decoder starting token attends to different parts of the encoder sequence. We split each encoder sequence into two segments, the history part (from p2p_2 to p3p_3) and the candidate part (after p3p_3). We then compute the average attention across all heads and tokens in each segment. Afterwards, we calculate the attention proportion of the candidate part via:

cand_attention_prop=cand_attentioncand_attention+history_attention(10)\mathrm{cand\_attention\_prop} = \frac{\mathrm{cand\_attention}}{\mathrm{cand\_attention} + \mathrm{history\_attention}} \tag*{(10)}

As shown in Figure 6, we find that in both nc=2n_c = 2 and nc=10n_c = 10, models generally extract more knowledge from the history in the early layers, and then become focused on the candidate knowledge in the later layers. For Yelp with nc=10n_c = 10, however, the candidate attention proportion significantly declines in later layers. This potentially stems from candidate information already being integrated into historical representations through encoder self-attention.

Figure 6. The candidate attention proportion (Equation 10) for nc=2n_c = 2 and nc=10n_c = 10. The attention score is averaged across all heads and tokens in each segment.

Conclusion

In this paper, we integrate the multiple-choice task as an auxiliary into generative recommendation models to alleviate the lack of comparative signals. Through analysis from the perspectives of information bottleneck and compositional generalization, we argue that the multiple-choice auxiliary task can improve the recommendation model by encouraging a more robust representation function and promoting richer compositional patterns. We conduct extensive experiments on four real-world datasets and demonstrate that our simple and elegant learning framework achieves competitive performance against mainstream strong baselines.

Our work identifies the fundamental challenges of generative retrieval with SIDs, which provides an alternative direction for improving generative recommenders. Future work can explore extending the proposed framework to different generative architectures, as well as developing advanced domain dependent candidate construction mechanisms.

A More Details on Experiments

We further report the mean and standard deviation of the overall performance for TIGER and “TIGER+nc=2n_c = 2” across 5 different random seeds in Table 9, as well as the testing subset statistics for the analysis of predicting items with different frequency (§ 4.5.2) in Table 10.

References

References

  1. [1]Yuwei Cao, Nikhil Mehta, Xinyang Yi, Raghunandan Hulikal Keshavan, Lukasz Heldt, Lichan Hong, Ed Chi, and Maheswaran Sathiamoorthy. 2024. Aligning Large Language Models with Recommendation Knowledge. In Findings of the Association for Computational Linguistics: NAACL 2024, Kevin Duh, Helena Gomez, and Steven Bethard (Eds.). Association for Computational Linguistics, Mexico City, Mexico, 1051–1066. https://doi.org/10.18653/v1/2024.findings-naacl.67
  2. [2]Jiaxin Deng, Shiyao Wang, Kuo Cai, Leijian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, and Guorui Zhou. 2025. OneRec: Unifying retrieval and rank with generative recommender and iterative preference alignment. arXiv preprint arXiv:2502.18965 (2025).
  3. [3]Zhou Du, Zhaoqiang Yuan, Xiao Wu, and Changsheng Xu. 2025. Disentanglement Based Equivariant Learning for Compositional VQA. IEEE Transactions on Multimedia 27 (2025), 8160–8173. https://doi.org/10.1109/TMM.2025.3604897
  4. [4]Shijie Geng, Chuhanq Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5). In Proceedings of the Sixteenth ACM Conference on Recommender Systems.arxiv.org/abs/2203.13366
  5. [5]F. Maxwell Harper and Joseph A. Konstan. 2015. The MovieLens Datasets: History and Context. ACM Trans. Interact. Intell. Syst. 5, 4, Article 19 (Dec. 2015), 19 pages. https://doi.org/10.1145/2827872
  6. [6]Ruining He, Lukasz Heldt, Lichan Hong, Raghunandan Keshavan, Shifan Mao, Nikhil Mehta, Zhengyang Su, Alicia Tsai, Yueqi Wang, Shao-Chuan Wang, Xinyang Yi, Lexi Baugher, Baykal Cakici, Ed Chi, Cristos Goodrow, Ningren Han, He Ma, Romer Rosales, Abby Van Soest, Devansh Tandon, Su-Lin Wu, Weilong Yang, and Yilin Zheng. 2026. PLUM: Adapting Pre-trained Language Models for Industrial-scale Generative Recommendations. In Proceedings of the ACM Web Conference 2026 (United Arab Emirates) (WWW ’26). Association for Computing Machinery, New York, NY, USA, 8093–8104. https://doi.org/10.1145/3774904.3792802
  7. [7]Barlas Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. 2016. Session-based Recommendations with Recurrent Neural Networks. In 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2–4, 2016, Conference Track Proceedings, Yoshua Bengio and Yann LeCun (Eds.). http://arxiv.org/abs/1511.06939
  8. [8]Yupeng Hou, Jiacheng Li, Zhankui He, An Yan, Xiusi Chen, and Julian McAuley. 2024. Bridging Language and Items for Retrieval and Recommendation. arXiv preprint arXiv:2403.03952 (2024).arxiv.org/abs/2403.03952
  9. [9]Yupeng Hou, Shanlei Mu, Wayne Xin Zhao, Yaliang Li, Bolin Ding, and Ji-Rong Wen. 2022. Towards Universal Sequence Representation Learning for Recommender Systems. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (WashingtON DC, USA) (KDD ’22). Association for Computing Machinery, New York, NY, USA, 585–593. https://doi.org/10.1145/3534678.3539381
  10. [10]Wenyue Hua, Shuyuan Xu, Yingqiang Ge, and Yongfeng Zhang. 2023. How to Index Item IDs for Recommendation Foundation Models. In Proceedings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region (Beijing, China) (SIGIR-AP ’23). Association for Computing Machinery, New York, NY, USA, 195–204. https://doi.org/10.1145/3624918.3625339
  11. [11]Yichen Jiang, Xiang Zhou, and Mohit Bansal. 2022. Mutual Exclusivity Training and Primitive Augmentation to Induce Compositionality. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang (Eds.). Association for Computational Linguistics, Abu Dhabi, United Arab Emirates, 11778–11793. https://doi.org/10.18653/v1/2022.emnlp-main.808
  12. [12]Clark Mingxuan Ju, Liam Collins, Leonardo Neves, Bhuvesh Kumar, Louis Yufeng Wang, Tong Zhao, and Neil Shah. 2025. Generative Recommendation with Semantic IDs: A Practitioner’s Handbook. In Proceedings of the 34th ACM International Conference on Information and Knowledge Management (Seoul, Republic of Korea) (CIKM ’25). Association for Computing Machinery, New York, NY, USA, 6420–6425. https://doi.org/10.1145/3746252.3761612
  13. [13]Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Recommendation. In 2018 IEEE International Conference on Data Mining (ICDM), 197–206. https://doi.org/10.1109/ICDM.2018.00035
  14. [14]Chuanhao Li, Zhen Li, Chenchen Jing, Yunde Jia, and Yuwei Wu. 2023. Exploring the effect of primitives for compositional generalization in vision-and-language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19092–19101.DOI
  15. [15]Chuanhao Li, Wenbo Ye, Zhen Li, Yuwei Wu, and Yunde Jia. 2025. Multi-sourced Compositional generalization in visual question answering. In Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence (Montreal, Canada) (IJCAI ’25). Article 150, 9 pages. https://doi.org/10.24963/ijcai.2025/150
  16. [16]Guanghan Li, Xun Zhang, Yufei Zhang, Yifan Yin, Guojun Yin, and Wei Lin. 2025. Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization. Proceedings of the AAAI Conference on Artificial Intelligence 39, 11 (Apr. 2025), 12040–12048. https://doi.org/10.1609/aaai.v39i11.33311
  17. [17]Lei Li, Yongfeng Zhang, and Li Chen. 2023. Prompt Distillation for Efficient LLM-based Recommendation. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management (Birmingham, United Kingdom) (CIKM ’23). Association for Computing Machinery, New York, NY, USA, 1348–1357. https://doi.org/10.1145/3583780.3615017
  18. [18]Xiang Lisa Li and Percy Liang. 2021. Prefix-Tuning: Optimizing Continuous Prompts for Generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Navigli (Eds.). Association for Computational Linguistics, Online, 4582–4597. https://doi.org/10.18653/v1/2021.acl-long.353
  19. [19]Zhaoyi Li, Ying Wei, and Defu Lian. 2023. Learning to Substitute Spans towards Improving Compositional Generalization. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (Eds.). Association for Computational Linguistics, Toronto, Canada, 2791–2811. https://doi.org/10.18653/v1/2023.acl-long.157
  20. [20]Enze Liu, Bowen Zheng, Cheng Leng, 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. https://doi.org/10.1145/3726302.3729989
  21. [21]Xinchen Luo, Jiangxia Cao, Tianyun Sun, Jinkai Yu, Rui Huang, Wei Yuan, Hezheng Lin, Yichen Zheng, Shiyao Wang, Qigen Hu, Changqing Qiu, Jiaqi Zhang, Xu Zhang, Zhiheng Yan, Jingming Zhang, Simin Zhang, Mingxing Wen, Zhaojie Liu, and Guorui Zhou. 2025. QARM: Quantitative Alignment Multi-Modal Recommendation at Quaishou. In Proceedings of the 34th ACM International Conference on Information and Knowledge Management (Seoul, Republic of Korea) (CIKM ’25). Association for Computing Machinery, New York, NY, USA, 5915–5922. https://doi.org/10.1145/3746252.3761502
  22. [22]Chen Ma, Peng Kang, and Xiao Liu. 2019. Hierarchical gating networks for sequential recommendation. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 825–833.arxiv.org/abs/1906.09217
  23. [23]Jianmo Ni, Gustavo Hernandez Abrego, Noah Constant, Ji Ma, Keith Hall, Daniel Cer, and Yinfei Yang. 2022. Sentence-T5: Scalable sentence encoders from pre-trained text-to-text models. In Findings of the association for computational linguistics: ACL 2022, 1864–1874.arxiv.org/abs/2108.08877
  24. [24]Arkil Patel, Satwik Bhattamishra, Phil Blunsom, and Navin Goyal. 2022. Re-visiting the Compositional Generalization Abilities of Neural Sequence Models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), Smaranda Muresan, Preslav Nakov, and Aline Villavicencio (Eds.). Association for Computational Linguistics, Dublin, Ireland, 424–434. https://doi.org/10.18653/v1/2022.acl-short.46
  25. [25]Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, Maciej Kula, Ed Chi, and Maheswaran Sathiamoorthy. 2023. Recommender Systems with Generative Retrieval. In Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (Eds.), Vol. 36. Curran Associates, Inc., 10299–10315. https://proceedings.neurips.cc/paper_files/paper/2023/file/20dcab0f14046a56cb02b61daf91229-Paper-Conference.pdf
  26. [26]Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. 2009. BPR: Bayesian personalized ranking from implicit feedback. In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence (Montreal, Quebec, Canada) (UAI ’09). AUAI Press, Arlington, Virginia, USA, 452–461.
  27. [27]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. https://doi.org/10.1145/3357384.3357895
  28. [28]Jiaxi Tang and Ke Wang. 2018. Personalized top-n sequential recommendation via convolutional sequence embedding. In Proceedings of the eleventh ACM international conference on web search and data mining, 565–573.
  29. [29]Naftali Tishby, Fernando C. Pereira, and William Bialek. 2000. The information bottleneck method. arXiv:physics/0004057 [physics.data-an] https://arxiv.org/abs/physics/0004057
  30. [30]Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023).arxiv.org/abs/2307.09288
  31. [31]Guanghan Li, Xiao Bin, 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. https://doi.org/10.1145/3627673.3679569
  32. [32]Xinfeng Wang, Jin Cui, Yoshimi Suzuki, and Fumiyo Fukumoto. 2024. RDRRec: Rationale Distillation for LLM-based Recommendation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 65–74. https://doi.org/10.18653/v1/2024.acl-short.6
  33. [33]Thaddäus Wiedemer, Prasanna Mayilvahanan, Matthias Bethge, and Wieland Brendel. 2023. Compositional Generalization from First Principles. In Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (Eds.), Vol. 36. Curran Associates, Inc., 6941–6960. https://proceedings.neurips.cc/paper_files/paper/2023/file/15f6a10899f55cce53fe39939af6f930-Paper-Conference.pdfarxiv.org/abs/2307.05596
  34. [34]Longtao Xiao, Haozhao Wang, Cheng Wang, Linfei Ji, Yifan Wang, Jieming Zhu, Zhenhua Dong, Rui Zhang, and Ruixuan Li. 2025. UNGER: Generative Recommendation with A Unified Code via Semantic and Collaborative Integration. ACM Trans. Inf. Syst. 44, 2, Article 31 (Dec. 2025), 31 pages. https://doi.org/10.1145/3773771
  35. [35]Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. 2021. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing 30 (2021), 495–507.arxiv.org/abs/2107.03312
  36. [36]Wayne Xin Zhao, Yupeng Hou, Xingyu Pan, Chen Yang, Zeyu Zhang, Zihan Lin, Jingsen Zhang, Shuqing Bian, Jiakai Tang, Wenqi Sun, et al. 2022. RecBole 2.0: Towards a more up-to-date recommendation library. In Proceedings of the 31st ACM International conference on information & knowledge management. 4722–4726.arxiv.org/abs/2206.07351
  37. [37]Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan Li, Yujie Lu, Hui Wang, Changxin Tian, et al. 2021. Recbole: Towards a unified, comprehensive and efficient framework for recommendation algorithms. In proceedings of the 30th acm international conference on information & knowledge management. 4653–4664.arxiv.org/abs/2011.01731
  38. [38]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 Conference on Data Engineering (ICDE). 1435–1448. https://doi.org/10.1109/ICDE60146.2024.00118
  39. [39]Kun Zhou, Hui Wang, Wayne Xin Zhao, Yutao Zhu, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. 2020. S3-Rec: Self-Supervised Learning for Sequential Recommendation with Mutual Information Maximization. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (Virtual Event, Ireland) (CIKM ’20). Association for Computing Machinery, New York, NY, USA, 1893–1902. https://doi.org/10.1145/3340531.3411954
  40. [40]Xiang Zhou, Yichen Jiang, and Mohit Bansal. 2023. Data Factors for Better Compositional Generalization. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, Singapore, 14549–14566. https://doi.org/10.18653/v1/2023.emnlp-main.898

Paper details

vlnc:W_pt6qtpaa
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