Three Selected Papers from WSDM 2024 on Recommendation Systems

WSDM 2024, the ACM International Conference on Web Search and Data Mining, accepted 109 papers; this article highlights three oral papers on recommendation systems.

Paper 1: Mixed Attention Network for Cross-domain Sequential Recommendation

Download: PDF | Code: GitHub

Authors: Guanyu Lin, Chen Gao, Yu Zheng, Jianxin Chang, Yanan Niu, Yang Song, Zhiheng Li, Depeng Jin, Yong Li, Meng Wang.

Abstract: Existing cross‑scene recommendation methods rely heavily on overlapping users, which is unrealistic; the paper identifies three challenges—different item features, diverse sequential behavior patterns, and non‑overlapping user interest transfer—and proposes a Mixed Attention Network (MAN) that combines item‑level, sequence‑level, and group‑level attention to achieve state‑of‑the‑art performance on multiple benchmarks.

Paper 2: Inverse Learning with Extremely Sparse Feedback for Recommendation

Download: PDF | Code: GitHub

Authors: Guanyu Lin, Chen Gao, Yu Zheng, Yinfeng Li, Jianxin Chang, Yanan Niu, Yang Song, Zhiheng Li, Depeng Jin, Yong Li.

Abstract: Traditional denoising recommendation focuses on either positive or negative samples; this work introduces a bidirectional denoising approach that simultaneously cleans both sides using an inverse‑weighted loss and meta‑learning‑based reverse gradients, achieving SOTA results on several denoising recommendation datasets.

Paper 3: LabelCraft: Empowering Short Video Recommendations with Automated Label Crafting

Download: PDF | Code: GitHub

Authors: Yimeng Bai, Yang Zhang, Jing Lu, Jianxin Chang, Xiaoxue Zang, Yanan Niu, Yang Song, Fuli Feng.

Abstract: Short‑video recommendation suffers from noisy user feedback; LabelCraft formulates label generation as a higher‑level optimization problem above the recommendation model, using a trainable label generator and meta‑learning to automatically craft labels that improve key operational metrics such as watch time, engagement, and retention, demonstrating superior performance on real‑world data.