Multi‐model deep learning approach for collaborative filtering recommendation system

Abstract

As a result of a huge volume of implicit feedback such as browsing and clicks, many researchers are involving in designing recommender systems (RSs) based on implicit feedback. Though implicit feedback is too challenging, it is highly applicable to use in building recommendation systems. Conventional collaborative filtering techniques such as matrix decomposition, which consider user preferences as a linear combination of user and item latent features, have limited learning capacities, hence suffer from a cold start and data sparsity problems. To tackle these problems, the research direction towards considering the integration of conventional collaborative filtering with deep neural networks to maps user and item features. Conversely, the scalability and the sparsity of the data affect the performance of the methods and limit the worthiness of the results of the recommendations. Therefore, the authors proposed a multi-model deep learning (MMDL) approach by integrating user and item functions to construct a hybrid RS and significant improvement. The MMDL approach combines deep autoencoder with a one-dimensional convolution neural network model that learns user and item features to predict user preferences. From detail experimentation on two real-world datasets, the proposed work exhibits substantial performance when compared to the existing methods.