Attention-based causal representation learning for out-of-distribution recommendation

Abstract

Out-of-distribution (OOD) recommendations have emerged as a popular field in recommendation systems. Traditional causal OOD recommendation frameworks often overlook shifts in latent user features and the interrelations between different user preferences. To address these issues, this paper proposes an innovative framework called Attention-based Causal OOD Recommendation (ABCOR), which applies the attention mechanism in two distinct ways. For shifts in latent user features, variational attention is employed to analyze shift information and refine the interaction-generation process. Besides, ABCOR integrates a multi-head self-attention layer to infer the complex user preference relationship and enhance recommendation accuracy before calculating post-intervention interaction probabilities. The proposed method has been validated on two public real-world datasets, and the results demonstrate that the proposal significantly outperforms the current state-of-the-art COR methods. Codes are available at https://github.com/YaffaGan/ABCOR.