Скачать или смотреть IP2: Entity-Guided Interest Probing for Personalized News Recommendation

The speaker investigates what triggers clicks in news consumption and argues that entities are key proxies for user preferences. IP2 models interest at two levels: intra-news (within an article) and inter-news (across articles). A Signature Entity Encoder, built on bidirectional transformers, produces a focus-weighted entity representation and is pretrained with a title–entity contrastive objective plus an auxiliary denoiser. A dual-tower user–item architecture with cross-attention links semantic and entity signals to capture short- and long-term preferences and maintain recommendation diversity. Experiments on MIND and Addressa show consistent gains over semantic-only and entity-aware baselines. Ablations confirm the contribution of entity utilization and pretraining choices. The method reduces reliance on lagging knowledge graphs and remains effective for emerging entities.