Movie Recommendation System based on DeepFM

With the advent of the era of informatization, the information explosion effect in the field of audio and video continues to iterate at an exponential magnitude, which makes it increasingly difficult for human beings to make decisions. In this case, building an effective recommendation system will bring people suitable services and products as well as save people much time hesitating. By analyzing historical records and exploiting the inner interaction between behaviors of users, recommendation systems can provide professional advice. Movie recommendation systems are among one of the most widely used systems. Considering the popularity, further development is necessary. To predict the likeness of users to recommended movies, this work introduces a movie recommendation system that is built based on Deep Factorization Machines (DeepFM). DeepFM has the ability of memorization as well as generalization. In other words, the structure of DeepFM can explore the inner and outer correlation between different features due to the factorization machine component. In order to simplify the model, DeepFM shares it input to both parts with no need for feature engineering. This model structure leads to more precise recommendations by learning users’ behavior deeply. Besides, this work compares the network with logistic regression and Wide&Deep on three different metrics. In the experiments, DeepFM outperforms the other two models in all of the metrics. Moreover, DeepFM and Wide&Deep have significant improvements over machine learning methods, convincing the validity of the wide component plus deep component model structure. Confirmed by the experiments, this work demonstrates that DeepFM is an available and efficient model for movie recommendation.