End-to-End Learnable Item Tokenization for Generative Recommendation

Enze Liu,Bowen Zheng,Cheng Ling,Lantao Hu,Han Li,Wayne Xin Zhao

DOI: https://doi.org/10.48550/arXiv.2409.05546

2024-09-09

Abstract:Recently, generative recommendation has emerged as a promising new paradigm that directly generates item identifiers for recommendation. However, a key challenge lies in how to effectively construct item identifiers that are suitable for recommender systems. Existing methods typically decouple item tokenization from subsequent generative recommendation training, likely resulting in suboptimal performance. To address this limitation, we propose ETEGRec, a novel End-To-End Generative Recommender by seamlessly integrating item tokenization and generative recommendation. Our framework is developed based on the dual encoder-decoder architecture, which consists of an item tokenizer and a generative recommender. In order to achieve mutual enhancement between the two components, we propose a recommendation-oriented alignment approach by devising two specific optimization objectives: sequence-item alignment and preference-semantic alignment. These two alignment objectives can effectively couple the learning of item tokenizer and generative recommender, thereby fostering the mutual enhancement between the two components. Finally, we further devise an alternating optimization method, to facilitate stable and effective end-to-end learning of the entire framework. Extensive experiments demonstrate the effectiveness of our proposed framework compared to a series of traditional sequential recommendation models and generative recommendation baselines.

Information Retrieval

What problem does this paper attempt to address?

The problem this paper attempts to address is the effective construction of item identifiers in generative recommendation systems. Specifically, existing generative recommendation methods typically handle item tokenization separately from the subsequent generative recommendation training process, which may lead to suboptimal performance. To overcome this limitation, the paper proposes a new framework called ETEGRec, which achieves end-to-end learning by seamlessly integrating item tokenization and generative recommendation. ETEGRec is based on a dual encoder-decoder architecture, including an item tokenizer and a generative recommender. To facilitate mutual enhancement between the two components, the paper proposes two specific optimization objectives: sequence-item alignment and preference-semantic alignment. Additionally, the paper designs an alternating optimization method to ensure stable and effective end-to-end learning of the entire framework. Experimental results show that ETEGRec outperforms traditional sequential recommendation models and generative recommendation baseline models on multiple recommendation benchmark datasets.