Source-linked AI summary
EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration
Ye Wang, Jiahao Xun, Minjie Hong, Jieming Zhu, Tao Jin, Wang Lin, Haoyuan Li, Linjun Li, Yan Xia, Zhou Zhao, Zhenhua Dong
TL;DR
Existing generative recommendation methods often emphasize behavioral or semantic information alone, despite their complementary strengths for user preferences and item characterization. EAGER uses a two-stream architecture with a shared encoder and separate decoders to predict behavior and semantic codes, merging stream results using prediction entropy for item ranking. Across four public recommendation benchmarks, EAGER outperforms existing generative and traditional methods.
Problem
Existing generative recommendation methods often emphasize behavioral or semantic information alone, despite their complementary strengths for user preferences and item characterization.
Source: §1 · p. 2Method
EAGER uses a two-stream architecture with a shared encoder and separate decoders to predict behavior and semantic codes, merging stream results using prediction entropy for item ranking.
Source: §1 · p. 2Results
Across four public recommendation benchmarks, EAGER outperforms existing generative and traditional methods.
Source: §1 · p. 3Takeaways & Limitations
Abstract
from arXiv · showhide
Generative retrieval has recently emerged as a promising approach to sequential recommendation, framing candidate item retrieval as an autoregressive sequence generation problem. However, existing generative methods typically focus solely on either behavioral or semantic aspects of item information, neglecting their complementary nature and thus resulting in limited effectiveness. To address this limitation, we introduce EAGER, a novel generative recommendation framework that seamlessly integrates both behavioral and semantic information. Specifically, we identify three key challenges in combining these two types of information: a unified generative architecture capable of handling two feature types, ensuring sufficient and independent learning for each type, and fostering subtle interactions that enhance collaborative information utilization. To achieve these goals, we propose (1) a two-stream generation architecture leveraging a shared encoder and two separate decoders to decode behavior tokens and semantic tokens with a confidence-based ranking strategy; (2) a global contrastive task with summary tokens to achieve discriminative decoding for each type of information; and (3) a semantic-guided transfer task designed to implicitly promote cross-interactions through reconstruction and estimation objectives. We validate the effectiveness of EAGER on four public benchmarks, demonstrating its superior performance compared to existing methods.
1 INTRODUCTION
EAGER addresses generative recommendation’s reliance on either behavioral or semantic information by combining the complementary signals in a two-stream framework. It pairs distinct learning objectives with semantic-guided transfer and reports superiority on four public benchmarks.
1 INTRODUCTION
- Generative retrieval frames next-item retrieval as autoregressive item-code prediction, while conventional recommendation separates representation learning from index construction.The separated phases challenge end-to-end optimization.Source: §1 · pp. 1–2
- Existing generative methods often focus on behavior or semantics alone, despite behavior’s user-specific preferences and semantics’ broader item insights.Source: §1 · p. 2
- EAGER uses a shared encoder and separate behavior and semantic decoders, ranking their predictions with stream entropy as a confidence measure.Source: §1 · p. 2
- A summary token and global contrastive task give each decoder global discriminatory learning alongside autoregressive generation.The token summarizes accumulated sequence knowledge.Source: §1 · p. 2
- Semantic-guided transfer uses global semantic features to reconstruct masked behavior tokens and recognize their alignment, promoting implicit cross-stream exchange.Source: §1 · p. 2
- EAGER outperforms existing generative and traditional recommendation methods across four public benchmarks.Source: §1 · p. 3
2 RELATED WORK
Related work spans sequential recommenders that commonly rely on ANN retrieval, tree-based matching, and generative retrieval using discrete identifiers and autoregressive generation.
2 RELATED WORK
- Sequential recommenders include GRU4REC, self-attentive SASRec [14], BERT4Rec [27], and self-supervised S3-Rec, with many relying on ANN indexes.Source: §2 · p. 3
- Tree-based methods [41] [42] offer another retrieval approach, while TIGER [23] predicts semantic identifiers derived from item side information.Source: §2 · p. 3
- Generative retrieval combines discrete semantic tokenization and autoregressive generation; document methods use varied identifiers, while pretrained models pose resource challenges for large-scale recommendation retrieval.Source: §2 · p. 3
3 METHOD
EAGER generates item codes from user interaction histories using separate behavior and semantic streams. Global contrastive learning and semantic-guided transfer complement generation, and confidence-based ranking merges the stream predictions.
3.1 Problem Formulation
- Given a user's interaction history, sequential recommendation returns next-item candidates; generative recommendation maps that history to serialized item codes and uses beam search to select candidates.The model generates item codes autoregressively from the interaction sequence.Source: §3.1 · p. 3
3.2 Overall Pipeline
- EAGER combines two-stream generation for behavior and semantic information, a summary-token global contrastive task, and semantic-guided transfer for cross-information and cross-decoder interaction.After training, confidence-based ranking merges predictions from the two streams.Source: §3.2 · p. 3
3.3 Two-stream Generation Architecture
- A shared encoder models interaction history, while separate decoders generate behavior and semantic codes, allowing each decoder to specialize in one code.Separate decoders mitigate supervision differences and enable parallel generation compared with one shared decoder.Source: §3.3 · p. 3
3.4 Global Contrastive Task
- A learnable end-of-sequence summary token distills global item-embedding knowledge through contrastive learning to strengthen each decoder's discriminative capability.Its gradient reaches preceding code tokens; the contrastive objective distills behavior and semantic embeddings from their pretrained encoders.Source: §3.4 · p. 4
3.5 Semantic-guided Transfer Task
- Semantic-guided transfer uses an auxiliary bidirectional Transformer with cross-attention to let behavior-code tokens attend to the semantic global feature.Training reconstructs randomly masked behavior tokens and recognizes whether behavior codes are relevant to the semantic feature, using randomly replaced tokens as negatives.Source: §3.5 · pp. 4–5
3.6 Training and Inference
- Training combines generation, contrastive, reconstruction, and recognition losses; inference beam-searches each stream and ranks the combined predictions by log-probability confidence.Lower log-probability values indicate greater confidence in the ranking strategy.Source: §3.6 · p. 5
4 EXPERIMENTS
The experiments address performance, component contributions, and design choices across four open-source sequential-recommendation datasets, using 5-core filtering and leave-one-out evaluation. They compare EAGER with existing methods and analyze its modules and hyperparameters.
4.1 Experimental Setting
- Experiments use four open-source 5-core datasets, leave-one-out evaluation, Recall and NDCG at top 5/10/20, and histories capped at 20 items.The final item is held out for testing, the preceding item for validation, and earlier items for training.Source: §4.1 · p. 5
4.2 Performance Comparison (RQ1)
- Across datasets, EAGER almost achieves better results than base models; on Beauty, it improves Recall@5 by up to 31.49% and NDCG@5 by 32.26% over TIGER, the second-best baseline.Source: §4.2 · p. 7
4.3 Ablation Study (RQ2)
- Removing any of TSG, GCT, or STT degrades performance, and removing all three yields the worst results across datasets.Removing TSG causes the largest decline; removing GCT causes greater drops than removing STT.Source: Table 3 · p. 6, §4.3 · p. 7
4.4 Model Analysis (RQ3)
- The Tail summary-token position performs best, while positive-only Cosine and Smooth ℓ1 outperform InfoNCE in the global contrastive task.The authors attribute the metric result partly to frozen pretrained features and noisy negatives in category-limited datasets.Source: Table 5 · p. 7, §4.4 · p. 8, Table 7 · p. 8
- Textual semantics with behavior outperform visual semantics with behavior, three-stream integration improves performance, and STT recognition matters more than reconstruction.Swapping guidance to behavior-guided transfer performs worse than semantic guidance.Source: §4.4 · p. 8, Figure 4 · p. 8
4.5 Hyper-Parameter Analysis (RQ3)
- Increasing decoder depth continuously improves Recall@10 and NDCG@10 across two datasets, but deeper models infer more slowly.The encoder-layer count has negligible performance effect, whereas decoder-layer count has a more significant influence.Source: §4.5 · p. 8, Figure 5 · p. 9
- As branch number k rises from 64 to 512, performance increases monotonically on Toys but declines from 256 to 512 on Beauty; larger k shortens identifiers and can increase inference time.The authors report EAGER outperforming the base model across the tested settings.Source: §4.5 · pp. 8–9, Figure 6 · p. 9
5 CONCLUSION AND FUTURE WORK
EAGER integrates behavioral and semantic information for generative recommendation through two-stream generation, global contrastive learning, and semantic-guided transfer. The authors report effectiveness and robustness in comparisons and analyses, and identify large language models and multimodal AI as future directions.
5 CONCLUSION AND FUTURE WORK
- EAGER is a unified generative recommendation framework designed to integrate behavioral and semantic information.Source: §5 · p. 9
- Its three components are two-stream generation, a summary-token global contrastive task, and semantic-guided transfer between decoders and features.Source: §5 · p. 9
- The paper reports that extensive comparisons and detailed analyses demonstrate EAGER’s effectiveness and robustness.Source: §5 · p. 9
- Future work will explore incorporating large language models and multimodal AI techniques.Source: §5 · p. 9