FORGE: Forming Semantic Identifiers for Generative Retrieval in Industrial Datasets

Kairui Fu, Tao Zhang, Shuwen Xiao, Ziyang Wang, Xinming Zhang, Chenchi Zhang, Yuliang Yan, Junjun Zheng, Yu Li, Zhihong Chen, Jian Wu, Xiangheng Kong, Shengyu Zhang, Kun Kuang, Yuning Jiang, Bo Zheng

Huggingface Github 知乎 Wechat

• The first industrial dataset about generative retrieval with semantic identifiers, which contains 14 billion user interactions and multimodal features of 250 million items sampled from Taobao.
• Subsequent proposed optimizations of data modality and ID collisions are validated with both offline (15\% improvements on HitRate) and online (0.35% improvements on transaction count) experiments in the “Guess You Like” Section of Taobao.

RankGR: Rank-Enhanced Generative Retrieval with Listwise Direct Preference Optimization in Recommendation

Kairui Fu*, Changfa Wu*, Kun Yuan, Binbin Cao, Dunxian Huang, Yuliang Yan, Junjun Zheng, Jianning Zhang, Silu Zhou, Jian Wu, Kun Kuang

• A rank-enhanced generative retrieval framework for recommendation. Equipped with the listwise direct preference optimization in IAP and the extra scoring module in RSP, RankGR breaks the limitations of existing methods, particularly their inability to capture hierarchy preference and rich interactions between SIDs and user sequences
• Deployed on Taobao’s “Guess You Like” section, RankGR achieved a 1.08% increase in online item page views and captured 49.88% of total exposures.

PI2I: A Personalized Item-Based Collaborative Filtering Retrieval Framework

Shaoqing Wang*, Yingcai Ma*, Kairui Fu*, Ziyang Wang, Dunxian Huang, YuliangYan, Jian Wu

Huggingface

• A novel two-stage retrieval framework that enhances the personalization capabilities of traditional collaborative filtering.
• Deployed on Taobao’s “Guess You Like” section, PI2I achieved a 1.05% increase in online transaction rates.

ThinkRec: Thinking-based recommendation via LLM

Qihang Yu*, Kairui Fu*, Shengyu Zhang, Zheqi Lv, Fan Wu, Fei Wu

Project

• Almost the first emphasizes the importance of activating the thinking of LLMs to make recommendations more interpretable and effective.

RASTP: Representation-Aware Semantic Token Pruning for Generative Recommendation with Semantic Identifiers

Tianyu Zhan*, Kairui Fu*, Zheqi Lv, Shengyu Zhang

Project

• An effective strategy to selectively prunes less informative tokens in the input sequence for semantic identifiers based recommendation.

MALLOC: Benchmarking the Memory-aware Long Sequence Compression for Large Sequential Recommendation

Qihang Yu*, Kairui Fu*, Zhaocheng Du*, Yuxuan Si, Kaiyuan Li, Weihao Zhao, Zhicheng Zhang, Jieming Zhu, Quanyu Dai, Zhenhua Dong, Shengyu Zhang, Kun Kuang, Fei Wu

Project

• A benchmark that establishes a rigorous multi-dimensional evaluation protocol that couples standard ranking metrics with system-level constraints for long-sequence compression in large recommender systems.

CHORD: Customizing Hybrid-precision On-device Model for Sequential Recommendation with Device-cloud Collaboration

Tianqi Liu*, Kairui Fu*, Shengyu Zhang, Wenyan Fan, Zhaocheng Du, Jieming Zhu, Fan Wu, Fei Wu

• A framework for device-cloud collaborative personalized mixed-precision quantization that generates lightweight networks for heterogeneous mobile devices.

Tackling Device Data Distribution Real-time Shift via Prototype-based Parameter Editing

Zheqi Lv, Wenqiao Zhang, Kairui Fu, Qi Tian, Shengyu Zhang, Jiajie Su, Jingyuan Chen, Kun Kuang, Fei Wu

• The composition of coarse and fine-grained intersts for tackling the on-device continuous data distribution shift in both vision and recommendation tasks.

Forward Once for All: Structural Parameterized Adaptation for Efficient Cloud-coordinated On-device Recommendation

Kairui Fu, Zheqi Lv, Shengyu Zhang, Fan Wu, Kun Kuang

Project

• An early attempt to investigate the joint customization of both structure and parameters, analyzing the challenges of interest heterogeneity, network transmission, and on-device inference simultaneously.

MergeNet: Knowledge Migration across Heterogeneous Models, Tasks, and Modalities

Kunxi Li*, Tianyu Zhan*, Kairui Fu*, Shengyu Zhang, Kun Kuang, Jiwei Li, Zhou Zhao, Fan Wu, Fei Wu

Project 知乎

• Leverage parameters as the medium to achieve knowledge transfer between heterogeneous models, tasks, and modalities.

DIET: Customized Slimming for Incompatible Networks in Sequential Recommendation

Kairui Fu, Shengyu Zhang, Zheqi Lv, Jingyuan Chen, Jiwei Li

• Tackle both the parameter personalization and the communication efficiency under strict device constraints in device-cloud collaborative recommendation.

End-to-End Optimization of Quantization-Based Structure Learning and Interventional Next-Item Recommendation

Kairui Fu, Qiaowei Miao, Shengyu Zhang, Kun Kuang, Fei Wu

• Investigate the inconsistent distribution of users in recommender system and the difficulty in causal structure learning accompanied by the intervention of recommender system.