Hybrid (Content-based filtering and Collaborative filtering) recommender system based on BERT

Abstract

Recommender system (RS) is the system that recommends items to users based on user preference. There are two main methods that are used in RS, including content-based filtering (CBF) and collaborative filtering (CF). However, these two main methods are not effective enough to get the better recommendation. Hence, many people have explored other methods such as neural content-based filtering and neural collaborative filtering. But those methods do not focus on the interaction sequence of the target user. Recently, there are researches that proposed to model sequential recommendations based on users’ historical interaction. One of these is BERT4Rec which takes BERT, a language understanding model, to model a sequential recommendation. This method only considers the historical sequence (item sequence) of the target user and does not consider the interaction of other users in the system that is content-based filtering approach. For this reason, we proposed a new method called Hybrid recommender system based on BERT, which applied BERT on both CBF and CF. For CBF, we consider the item sequence of the target user as same as BERT4Rec, but the result of our proposed method is the target user profile. For CF, we consider other users who used to interact with the target item and call it as user sequence of the target item. The result of the CF side is the target item profile. Finally, after obtaining the results on both CBF and CF side, we use it to predict the rating score by the NCF approach. To evaluate our proposed method, we compared it with BERT4Rec on MovieLens-1M dataset in terms of accuracy by the NDCG approach. The experimental result shows that our proposed method outperforms BERT4Rec.