Transformer Based Sequential Recommender System

Abstract

Recommender systems (RS) aids in helping endusers by providing suggestions and predicting items of their interest in e-commerce and social media platforms. Sequence of user’s historical preferences are used by Sequential Recommendation system (SRS) to predict next user-item interaction. In recent literature, various deep learning methods like CNN and RNN have shown significant improvements in finding recommendations, however, anticipating future item pertaining to user’s past record history is still challenging. With the introduction of transformer architecture, SRS have gained major performance boost in generating precise recommendations. Recently proposed models based on transformer architecture predict next user-item by exploiting item identifiers only. Regardless of the efficacy of these models, we believe that performance of recommendation models can be improved by adding some additional descriptive item features along with the item identifiers. This paper proposes a transformer based SRS that models user behavior sequences, by incorporating auxiliary information along with item identifiers for producing more accurate recommendations. The proposed model extends the BERT4Rec model to incorporate auxiliary information by exploiting the ”Sentence Transformer model” to produce the sentence representations from the textual features of items. This dense vector representation is then merged with the item representations of user. Comprehensive experiments upon various benchmark datasets shows remarkable improvements when corelating with other similar state-of-the-art models.