A Customized DeepICF+a with BiLSTM for Better Recommendation

Abstract

Recommendations are expected to help users make decisions when users are faced with a large amount of information. One technique for developing a recommendation system is item-based collaborative filtering (ICF), where this approach recommends items based on their similarity to the items with which users already interacted and comparable decisions made by other users. In recent years, many ICF approaches have made significant progress by using deep neural networks to learn similarities from data. Developing a recommender system based on ICF+attention approach has shown a significant output. In this research, we conducted experiments on MovieLens 1M dataset to build movie recommendation. A recent study demonstrates a good result by HR= 0.7084 and NDCG = 0.4380 of its performance. Previous work implemented MLP to predict the next watched movies. MLP performs poorly for prediction compared to BiLSTM performs better for prediction if the data in a historical model. In this work, we modify the architecture of the previous study (DeepICF+a with MLP) by replacing MLP model with BiLSTM Our work shows that the performance have a better result by 0.7121 and 0.4399 for HR and NDCG, respectively with configuration embedding size = 32, layers BiLSTM [64, 32, 16] and number negative = 8. The DeepICF+a with BiLSTM recommendation model provides a better optimization model for Train Loss with a score of 0.2064 and a Test Loss with a score of 0.1263 compared to MLP for train loss with a score of 0.2127 and a Test Loss with a score of 0.3167.