International Journal of Computer Applications |
Foundation of Computer Science (FCS), NY, USA |
Volume 186 - Number 3 |
Year of Publication: 2024 |
Authors: Ashima Malik, S. Srinivasan, Piyush Prakash |
![]() |
Ashima Malik, S. Srinivasan, Piyush Prakash . Enhancing Recommendations of Items by Making Some Changes in Layers of BERT Model. International Journal of Computer Applications. 186, 3 ( Jan 2024), 28-36. DOI=10.5120/ijca2024923366
Introducing an innovative methodology for modeling user behavior sequences in recommendation systems, this paper proposes the use of a bidirectional self-attention network and Cloze task, drawing inspiration from Bidirectional Encoder Representations from Transformers (BERT) to enhance the recommendations of products on e-commerce websites. Traditional recommendation system models that are unidirectional have limitations, mainly in the power of hidden representations and rigid ordering of historical user interactions. Overcoming these limitations, the suggested BERT4Rec model is bidirectional, offering the context from both directions. The paper suggests utilizing the Cloze task to prevent data leakages from bidirectional conditioning. This includes masking random components within the input sequences and predicting them based on their nearby context. Comprehensive experiments are conducted, resulting in consistently better outcomes than state-of-the-art comparable options across four datasets. This exploration sets the groundwork by introducing the Cloze objective and deep bidirectional sequential modeling to the recommendation system field. Furthermore, the study is a foundation for future studies investigating explicit user modeling and incorporating item features.