A Position-Aware Self-Attention Neural Model for Personalized Session-Based Recommendation

Document Type : Original Article

Authors

Department of Computer Engineering, Yazd University, Yazd, Iran

Abstract

Personalized session-based recommendation seeks to predict the user's next click or interaction based on the user's previous interactions and behaviour in the present and historical sessions. Recent research has focused on self-attention networks (SANs) to capture users' global interests and behaviours. Thanks to their modeling of the global dependencies among session interactions, SANs have exhibited superior performance in session-based recommendation compared to other deep network approaches. These networks do not account for items' position and order in the session sequence. Nonetheless, the sequential data of the sessions and the item order in the session sequence can reflect the user's sequential interests and improve the recommendation's performance. This article proposes a Position-Aware Self-Attention Neural (PASAN) model for personalized session-based recommendations. The model uses a reverse-position encoding mechanism that assigns position embedding to items based on their order in the session sequence. The PASAN model jointly learns global interests through the self-attention network and sequential interests via positional encoding. Moreover, it employs users' historical and current session interactions to model their long-term preferences. First, PASAN is trained using anonymous sessions. Subsequently, personalized recommendations are provided for individual users by combining the current session with their historical sessions in a weighted manner. Experiments conducted on two real-world datasets show that the proposed model outperforms the current methods. The proposed model has improved by about 20% in terms of precision and about 8% in terms of mean reciprocal rank on the Reddit dataset.

Keywords

TABRIZ JOURNAL OF ELECTRICAL ENGINEERING

Articles in Press, Corrected Proof
Available Online from 28 September 2024

Files

Share

How to cite

Statistics