Adaptive user multi-level and multi-interest preferences for sequential recommendation

doi:10.1007/s11280-025-01332-4

Adaptive user multi-level and multi-interest preferences for sequential recommendation

2025-02-10

1

PRE

AI

DOI：10.1007/s11280-025-01332-4

Original

Original request for help

Save

Abstract

En 中文

Sequential recommendation systems provide users with items of interest for the next moment by modeling their preferences in their interaction sequences. However, users' interests change over time, and mining multiple interests in their interaction sequences can provide more accurate recommendations. Mainstream multi-interest sequence recommendations utilize self-attention mechanisms and capsule networks to learn users' multi-interests. These approaches fail to decouple the capture of multi-interests within sequences and between users, leading to insufficient learning of user preferences. In addition, existing multi-interest methods cannot adaptively assign the number of interests to users. Therefore, this paper proposes an adaptive user multi-level and multi-interest preferences for sequential recommendation (AdaSR) approach that decouples the learning of users' local sequential and global collaborative preferences, and learns the corresponding number of interests for each user. Specifically, AdaSR first learns users' local sequential preferences and global collaborative preferences separately through a preference encoding module, which efficiently encodes the preference features of user interaction sequences and similar users; then, we decouple the user's local sequential preferences and global collaborative preferences to learn multi-level and multi-interest embeddings. Meanwhile, a gating mechanism is designed to remove the redundant embeddings between local sequential and global collaborative multi-interests, and adaptively learn the user's multi-interests. Finally, data augmentation and graph augmentation contrastive learning assistance tasks are performed during training to improve the quality of user preference embeddings. Experiments on both datasets show that our method significantly outperforms the baseline.

Keywords:

Sequential recommendation

Contrastive learning

Multi-interest learning

Data augmentation

Capsule networks

Adaptive user multi-level and multi-interest preferences for sequential recommendation

Abstract

Journal

Organization