Denoising Multi‐Level Preference Learning for Local Interest‐Oriented Session‐Based Recommendation

ABSTRACT Session‐based recommendation (SBR) is an important part of modern recommender systems. It can model user preferences without long‐term user profiles. This is useful in scenarios with anonymous users or fast‐changing user behaviors. In these cases, long‐term histories are missing or unreliable. However, SBR still faces several challenges. Graph‐based methods find it hard to capture users' multilayered and diverse interests. Interest shifts also introduce noise within sessions and across sessions. It is also difficult to model a user's immediate intent while keeping stable local interests. We propose the Denoising Multi‐Level Preference Learning for Local Interest‐oriented Graph Neural Network (DMPL‐GNN) to solve these problems. The model has three main parts. First, we design an adaptive graph module with convolutional residual networks. It learns fine‐grained local interests. Second, we add a target node and combine graph convolution with sparse attention. This builds a stable representation of global interests. Third, we design an intent fusion module. It treats local interest as the main factor for generating recommendations. Experiments on real‐world datasets show that DMPL‐GNN performs better than existing methods. Ablation studies also prove the usefulness of each module. Our work offers new ideas for improving session‐based recommendation. Our code is available at https://github.com/csgii/DMPL‐GNN/tree/main .