Mixture of Sequence Framework for Long-Sequence Recommendation

A new framework called Mixture of Sequence (MoS) addresses the challenge of modeling long user sequences in sequential recommendation systems. The key issue is that users exhibit significant interest shifts over long sequences, introducing irrelevant information. Empirical analysis reveals a recurring pattern termed 'session hopping,' where interests remain stable within short sessions but shift drastically across sessions and may reappear. MoS is a model-agnostic Mixture-of-Experts approach that extracts theme-specific and multi-scale subsequences from noisy raw sequences. It uses a theme-aware routing mechanism to adaptively learn latent themes and organize sequences, improving click-through rate prediction accuracy. The framework is described in a paper on arXiv (2604.20858v1).