Self-supervised Contrastive Enhancement with Symmetric Few-shot Learning Towers for Cold-start News Recommendation

Nowadays, news spreads faster than it is consumed. This, alongside the rapid news cycle and delayed updates, has led to a challenging news cold-start issue. Likewise, the user cold-start problem, due to limited user engagement, has long hindered recommendations. To tackle both of them, we introduce the Symmetric Few-shot Learning framework for Cold-start News Recommendation (SFCNR), built upon self-supervised contrastive enhancement. Our approach employs symmetric few-shot learning towers (SFTs) to transform warm user/news attributes into their behavior/content features during training. We design two innovative feature alignment strategies to enhance towers training. Subsequently, this tower generates virtual features for cold users/news during inference, leveraging tower-stored prior knowledge through a personalized gating network. We assess the SFCNR on four quality news recommendation models, conducting comprehensive experiments on two kinds of News dataset. Results showcase significant performance boosts for both warm and cold-start scenarios compared to baseline models.