Highlights of Six KDD 2023 Papers on Personalized Recommendation and User Behavior Modeling

Paper 01: PEPNet: Parameter and Embedding Personalized Network for Infusing with Personalized Prior Information Download: https://arxiv.org/abs/2302.01115 Authors: 常健新, 张晨斌, 惠轶群, 冷德维, 牛亚男, 宋洋 (Kuaishou) Abstract: To address the challenges of multi‑task, multi‑domain recommendation in large‑scale services, PEPNet introduces a plug‑and‑play personalized network that injects personalized prior information via gate mechanisms to dynamically scale both low‑level embeddings and high‑level DNN hidden units. The embedding‑personalized sub‑network (EPNet) selects embeddings per domain, while the parameter‑personalized sub‑network (PPNet) adjusts DNN parameters per user. Deployed at Kuaishou, it serves over 300 million users daily and yields >1% online improvement across multiple tasks.

Paper 02: TWIN: Two‑stage Interest Network for Lifelong User Behavior Modeling in CTR Prediction at Kuaishou Download: https://arxiv.org/abs/2302.02352 Authors: 常健新, 张晨斌, 傅智毅, 臧晓雪, 关琳, 吕静, 惠轶群, 冷德维, 牛亚男, 宋洋 (Kuaishou) Abstract: Lifelong user behavior modeling aims to capture months‑long interaction histories containing tens of thousands of actions, which exceed the capacity of standard attention mechanisms. TWIN aligns the GSU (coarse filtering) and ESU (fine attention) stages by using identical distance metrics, ensuring consistent relevance scoring. It splits features into intrinsic video features and user‑video cross features, caches linear projections, and compresses cross features as attention biases, reducing computational bottlenecks by 99.3% and serving 340 million daily active users.

Paper 03: Tree based Progressive Regression Model for Watch‑Time Prediction in Short‑video Recommendation Download: https://arxiv.org/abs/2306.03392 Authors: 林肖, 陈潇凯, 宋林枫, 刘静伟, 李彪, 江鹏 (Kuaishou) Abstract: Watch‑time prediction is treated as an ordinal regression problem. TPM decomposes the task into a series of dependent classification sub‑tasks organized in a tree structure, where sibling nodes have ordinal relations and parent‑child nodes have conditional dependencies. The model outputs probabilities for leaf intervals, enabling variance estimation for uncertainty and bias adjustment via confounding factors. Deployed on Kuaishou’s main page, TPM achieves significant watch‑time gains.

Paper 04: Generative Flow Network for Listwise Recommendation (GFN4Rec) Download: https://arxiv.org/abs/2306.02239 Code: https://github.com/CharlieMat/GFN4Rec Authors: 刘殊畅, 蔡庆芃, 何占魁 (UCSD), 孙博文 (Peking University), Julian McAuley (UCSD), 郑东, 江鹏 (Kuaishou) Abstract: Listwise recommendation aims to generate an entire ranked list rather than optimizing individual items. GFN4Rec leverages the Generative Flow Network (GFlowNet) framework to model the list generation as a probabilistic flow graph, aligning the generation probability with the overall list reward. The approach maintains high recommendation quality while improving diversity, outperforming existing collaborative filtering, generative, and re‑ranking methods on multiple benchmarks.

Paper 05: PrefRec: Recommender Systems with Human Preferences for Reinforcing Long‑term User Engagement Download: https://arxiv.org/abs/2212.02779 Code: https://www.dropbox.com/sh/hgsqg5fabnvmp26/AABF-2dvarI_bdyygYEt5aw7a?dl=0 Authors: 薛万祺 (Nanyang Technological University), 蔡庆芃, 薛正海, 孙硕 (NTU), 刘殊畅, 郑东, 江鹏, 安波 (NTU) Abstract: PrefRec proposes a paradigm where reinforcement‑learning‑based recommenders learn a reward function directly from user preference data instead of hand‑crafted rewards, mitigating sparsity, delay, and randomness issues. It first trains an end‑to‑end reward model from preferences, then uses this learned reward to guide policy training, employing an auxiliary value function and expectation regression to improve critic accuracy. Experiments show PrefRec significantly outperforms SOTA on long‑term engagement metrics.

Paper 06: Graph Contrastive Learning with Generative Adversarial Network (GACN) Download: https://arxiv.org/abs/2308.00535 Code: https://www.dropbox.com/sh/hgsqg5fabnvmp26/AABF-2dvarI_bdyygYEt5aw7a?dl=0 Authors: 吴呈, 王朝坤, 徐劲草, 刘子扬 (Tsinghua University), 郑凯, 王晓伟, 宋洋 (Kuaishou) Abstract: Traditional graph contrastive learning generates views by random node/edge perturbations, which may ignore potential new edges or introduce noise. GACN introduces a generative adversarial network to produce contrastive views that respect the underlying graph distribution and evolution, generating realistic new edges. Joint training of graph generation and contrastive learning yields richer views and superior performance on downstream tasks.