EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration **Authors:** Ye Wang, Jiahao Xun, Minjie Hong, Jieming Zhu, Tao Jin, Wang Lin, Haoyuan Li, Linjun Li, Yan Xia, Zhou Zhao, Zhenhua Dong **Affiliation:** Zhejiang University and Huawei Noah’s Ark Lab **Abstract:** Generative retrieval has emerged as a promising approach to sequential recommendation, framing candidate item retrieval as an autoregressive sequence generation problem. However, existing methods often focus solely on either behavioral or semantic aspects of item information, neglecting their complementary nature. To address this limitation, EAGER introduces a novel generative recommendation framework that integrates both behavioral and semantic information. EAGER addresses three key challenges: 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. EAGER employs a two-stream generation architecture with a shared encoder and two separate decoders for behavior and semantic tokens, a global contrastive task with a summary token, and a semantic-guided transfer task designed to promote cross-interactions. Extensive experiments on four public benchmarks demonstrate EAGER's superior performance compared to existing methods. **Keywords:** Generative Recommendation, Autoregressive Generation, Semantic Tokenization, Behavior-Semantic Collaboration **Introduction:** Recommender systems use deep sequential models to capture user-item patterns. EAGER proposes a novel two-stream generative architecture that integrates behavior and semantic information. It includes a shared encoder and two separate decoders for behavior and semantic codes, a global contrastive task with a summary token, and a semantic-guided transfer task. EAGER enhances the integration of behavioral and semantic information, improving the overall effectiveness of recommender systems. **Method:** EAGER's framework consists of: 1. **Two-stream Generation Architecture:** A shared encoder models user interaction history, and two separate decoders predict behavior and semantic codes. 2. **Global Contrastive Task:** A summary token captures global knowledge to improve auto-regressive generation quality. 3. **Semantic-guided Transfer Task:** An auxiliary transformer promotes cross-information and cross-decoder interaction. **Experiments:** EAGER is evaluated on four datasets, showing superior performance compared to existing methods. Ablation studies and hyper-parameter analysis further validate the effectiveness of EAGER's components. **Conclusion:** EAGER integrates behavioral and semantic information for unified generative recommendation, enhancing the effectiveness of recommender systems. Future work will explore incorporating large language models and multimodal AI techniques.EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration **Authors:** Ye Wang, Jiahao Xun, Minjie Hong, Jieming Zhu, Tao Jin, Wang Lin, Haoyuan Li, Linjun Li, Yan Xia, Zhou Zhao, Zhenhua Dong **Affiliation:** Zhejiang University and Huawei Noah’s Ark Lab **Abstract:** Generative retrieval has emerged as a promising approach to sequential recommendation, framing candidate item retrieval as an autoregressive sequence generation problem. However, existing methods often focus solely on either behavioral or semantic aspects of item information, neglecting their complementary nature. To address this limitation, EAGER introduces a novel generative recommendation framework that integrates both behavioral and semantic information. EAGER addresses three key challenges: 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. EAGER employs a two-stream generation architecture with a shared encoder and two separate decoders for behavior and semantic tokens, a global contrastive task with a summary token, and a semantic-guided transfer task designed to promote cross-interactions. Extensive experiments on four public benchmarks demonstrate EAGER's superior performance compared to existing methods. **Keywords:** Generative Recommendation, Autoregressive Generation, Semantic Tokenization, Behavior-Semantic Collaboration **Introduction:** Recommender systems use deep sequential models to capture user-item patterns. EAGER proposes a novel two-stream generative architecture that integrates behavior and semantic information. It includes a shared encoder and two separate decoders for behavior and semantic codes, a global contrastive task with a summary token, and a semantic-guided transfer task. EAGER enhances the integration of behavioral and semantic information, improving the overall effectiveness of recommender systems. **Method:** EAGER's framework consists of: 1. **Two-stream Generation Architecture:** A shared encoder models user interaction history, and two separate decoders predict behavior and semantic codes. 2. **Global Contrastive Task:** A summary token captures global knowledge to improve auto-regressive generation quality. 3. **Semantic-guided Transfer Task:** An auxiliary transformer promotes cross-information and cross-decoder interaction. **Experiments:** EAGER is evaluated on four datasets, showing superior performance compared to existing methods. Ablation studies and hyper-parameter analysis further validate the effectiveness of EAGER's components. **Conclusion:** EAGER integrates behavioral and semantic information for unified generative recommendation, enhancing the effectiveness of recommender systems. Future work will explore incorporating large language models and multimodal AI techniques.