Unifying Generative and Dense Retrieval for Sequential Recommendation
Liu Yang, Fabian Paischer, Kaveh Hassani, Jiacheng Li, Shuai Shao,, Zhang Gabriel Li, Yun He, Xue Feng, Nima Noorshams, Sem Park, Bo Long, Robert, D Nowak, Xiaoli Gao, Hamid Eghbalzadeh
TL;DR
This paper compares generative and dense retrieval methods for sequential recommendation, introduces a hybrid model called LIGER that combines their strengths, and demonstrates improved performance and efficiency in benchmark tests.
Contribution
It provides a comprehensive comparison of generative and dense retrieval approaches and proposes LIGER, a hybrid model that enhances recommendation performance and cold-start item handling.
Findings
LIGER improves recommendation accuracy over individual methods.
Hybrid approach reduces memory and computation requirements.
Enhanced cold-start item recommendation in benchmarks.
Abstract
Sequential dense retrieval models utilize advanced sequence learning techniques to compute item and user representations, which are then used to rank relevant items for a user through inner product computation between the user and all item representations. However, this approach requires storing a unique representation for each item, resulting in significant memory requirements as the number of items grow. In contrast, the recently proposed generative retrieval paradigm offers a promising alternative by directly predicting item indices using a generative model trained on semantic IDs that encapsulate items' semantic information. Despite its potential for large-scale applications, a comprehensive comparison between generative retrieval and sequential dense retrieval under fair conditions is still lacking, leaving open questions regarding performance, and computation trade-offs. To…
Peer Reviews
Decision·Submitted to ICLR 2025
1. The idea of combining generative and discriminative approaches is promising. 2. The method presented by the authors is relatively simple and easy to understand and follow. 4. Experiments are conducted on four datasets, showing good performance.
1. There are issues with some claims or descriptions in the introduction. For instance, generative recommendation methods are not limited to codebook-based approaches like TIGER, yet the authors present it as the only form of generative retrieval; also, generative methods are not inherently incapable of handling cold start recommendations since some works can solve this problem. Additionally, the authors do not introduce the core idea of their method in the introduction. And few studies in the r
Rigorous research process: The paper conducts detailed experimental verification for each core hypothesis and key proposition. Through the corresponding experimental data, the advantages of the LIGER model in recommendation accuracy, cold start processing ability, and the quality of generated candidate sets are demonstrated. Research motivation : The research motivation of the paper is clear, and it studies two key issues in the recommendation system: how to strike a balance between performance
The chart design is not intuitive enough: The design of the chart (especially the Normalized Performance Gap related chart) has room for improvement in information presentation. The captions are too long and the content is dense, resulting in unclear information transmission and affecting readability. It is recommended to optimize the structure and presentation of the chart to more intuitively show the impact of different numbers of candidate items on the performance gap between generative retri
1. The paper presents a comprehensive comparison between generative retrieval and sequential dense retrieval models for sequential recommendation, identifying key limitations of generative retrieval, with clear illustrations and tabulates the computational cost. 2. The authors propose LIGER, a hybrid model that combines the strengths of dense and generative retrieval to address the limitations of generative retrieval, reducing the performance gap and improving the generation of cold-start items.
My main concern of the papers are as follows: 1. Novelty. Despite performance improvement, the paper is a simple extension of the TIGER method which the combination of dense retrieval, which is quite straightforward and hence the contribution in terms of the methodology is largely incremental. 2. Dataset. The choice of the datasets are not optimal. All four datasets are pretty old and relatively small, comparing to recent released datasets for recommendation tasks, such as KuaiRand, KuaiRec and
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsRecommender Systems and Techniques · Advanced Bandit Algorithms Research · Image Retrieval and Classification Techniques