Classification-based deep neural network architecture for collaborative filtering recommender systems
J. Bobadilla, F. Ortega, S. Alonso, & R. Hurtado
Keywords: Deep Learning; Neural Classification; Neural Collaborative Filtering; Recommender Systems; Scalable Neural Architecture
This paper proposes a scalable and original classification-based deep neural architecture. Its collaborative filtering approach can be generalized to most of the existing recommender systems, since it just operates on the ratings dataset. The learning process is based on the binary relevant/non-relevant vote and the binary voted/non-voted item information. This data reduction provides a new level of abstraction and it makes possible to design the classification-based architecture. In addition to the original architecture, its prediction process has a novel approach: it does not need to make a large number of predictions to get recommendations. Instead to run forward the neural network for each prediction, our approach runs forward the neural network just once to get a set of probabilities in its categorical output layer. The proposed neural architecture has been tested by using the MovieLens and FilmTrust datasets. The state of art NCF neural model has been used as baseline; it outperforms KNN, MF, BPR and eALS, and it provides similar accuracy to deepFM. Results show a competitive recommendation quality and an interesting quality improvement on large number of recommendations, consistent with the architecture design. The architecture originality makes it possible to address a broad range of future works.