.png){kind=icon}
.png){kind=icon}
Return
Self-supervised session-based recommendation model based on global causal structure learning
.png){kind=icon}
PRE
AI
DOI：10.1007/s13042-025-02901-9.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Session-based recommendation can predict a user’s next interaction item through anonymous and short interaction sequences. The true preferences of users are crucial in providing accurate recommendations for session-based recommendation. However, most of the methods focus only on modeling the correlation dependencies between items, and there are limitations in modeling the multivariate dependencies between items. Causal relationships, on the other hand, can reveal the driving mechanism of user decisions and show the user’s decision path, assisting the model to capture user preferences more accurately. Therefore, we propose a causality and correlation-enhanced graph neural network framework to systematically model different dependencies among items by constructing the causal and correlation graphs, thereby capturing user preferences more accurately. Specifically, we introduce global causal structure learning into session-based recommendation for learning causal relationships between items. First, we employ a deep generative model and acyclic constraints to learn a causal graph, and the core of the generative model is a variational autoencoder parameterized by a graph neural network architecture. In addition, to capture more comprehensive information from causal and correlated perspectives, we deploy an attention mechanism to learn separate session representations for the two item embeddings. Simultaneously, we introduce self-supervised and contrastive learning to enhance session representations obtained from causal and correlated perspectives. Finally, we conduct extensive experiments on three real datasets to demonstrate the effectiveness of our design.
Keywords:
Session-based recommendation
Causal structure learning
Correlation
Self-supervised learning
Journal
IF:
2.7
Papers:
3.1K
Citations:
5.6K