Selected Papers from CIKM 2022 on Real‑Time Short Video Recommendation and Large‑Scale Datasets

CIKM (the 31st ACM International Conference on Information and Knowledge Management) is a prestigious B‑class conference recommended by the China Computer Federation (CCF) and was held from October 17 to 21, 2022.

Paper 01: Real‑time Short Video Recommendation on Mobile Devices Download: https://arxiv.org/abs/2208.09577 Authors: Xudong Gong, Qinlin Feng, Yuan Zhang, Jiangling Qin, Weijie Ding, Biao Li, Peng Jiang (all from Kuaishou). Abstract: In short‑video apps users generate rapid feedback, requiring recommender systems that can adjust rankings in real time. Traditional cloud‑based recommenders return a fixed ordered list and cannot incorporate fresh feedback before the next request, leading to latency and outdated rankings. The paper proposes a lightweight on‑device re‑ranking model that leverages immediate user signals and device‑specific features, combined with an adaptive beam‑search to improve ordering and user experience, achieving significant online gains.

Paper 02: KuaiRec – A Fully‑Observed Dataset and Insights for Evaluating Recommender Systems Download: https://arxiv.org/abs/2202.10842 Dataset website: https://kuairec.com Authors: Chongming Gao, Shijun Li, Wenqiang Lei, Jiawei Chen, Biao Li, Peng Jiang, Xiangnan He, Jiaxin Mao, Tat‑Seng Chua. Abstract: Most public recommendation datasets are sparse and biased, hindering research. KuaiRec collects fully observed interactions from 1,411 Kuaishou users on 3,327 videos, providing complete exposure data with millions of interactions. Experiments show that bias‑induced data sampling dramatically affects multi‑round conversational recommendation evaluation, underscoring the necessity of fully‑exposed datasets.

Paper 03: KuaiRand – An Unbiased Sequential Recommendation Dataset with Randomly Exposed Videos (Resource Track) Download: https://arxiv.org/abs/2208.08696 Dataset website: https://kuairand.com Authors: Chongming Gao, Shijun Li, Yuan Zhang, Jiawei Chen, Biao Li, Wenqiang Lei, Peng Jiang, Xiangnan He. Abstract: Exposure bias in commercial recommenders leads to biased interaction logs. KuaiRand mitigates this by randomly exposing millions of videos in the Kuaishou app, recording 12 feedback signals (clicks, likes, watch time, etc.) along with rich user and item features. The dataset supports research on interactive recommendation, reinforcement‑learning‑based recommendation, long‑sequence modeling, and multi‑task learning.

Paper 04: Billion‑User Customer Lifetime Value Prediction – An Industrial‑Scale Solution from Kuaishou Download: http://arxiv.org/abs/2208.13358 Authors: Kunpeng Li, Guangcui Shao, Naijun Yang, Xiao Fang, Yang Song. Abstract: Predicting Customer Lifetime Value (LTV) for billions of users is challenging due to complex, non‑stationary distributions. The authors introduce an Ordered‑Dependent Monotonic Network (ODMN) to capture sequential dependencies across time horizons, and a Multi‑Distribution Multi‑Expert (MDME) module that decomposes highly imbalanced distributions into balanced sub‑distributions. A new evaluation metric, Mutual Gini, better aligns estimated Lorenz curves with ground truth. The solution has been deployed in Kuaishou’s production environment with strong performance gains.