Deep autoencoders for feature learning with embeddings for recommendations: a novel recommender system solution

We propose “Deep Autoencoders for Feature Learning in Recommender Systems,” a novel discriminative model based on the incorporation of features from autoencoders in combination with embeddings into a deep neural network to predict ratings in recommender systems. The work has two major motivations. The first is to engineer features for recommender systems in a domain-agnostic way using autoencoders. The second is to develop a method that sets a benchmark for predictive accuracy. In our proposed solution, we build a user autoencoder and item autoencoder that extract latent features for the users and items, respectively. The additional features engineered are the latent features for the users and items, and these come from the bottleneck activations of the autoencoder. Our method of feature engineering is domain agnostic, as the inner-most activations would differ for domains without any additional effort required on part of the modeler. Next, we then use the activations of the inner-most layers of the autoencoders as features in a subsequent deep neural network to predict the ratings along-with user and item embeddings. Our method incorporates additional linear and nonlinear latent features from the autoencoders to improve predictive accuracy. This is different from the existing approaches that use autoencoders as full-fledged recommender systems or use autoencoders to generate features for a subsequent supervised learning algorithm or without using embeddings. We demonstrate the out performance of our solution on four different datasets of varying sizes and sparsity, namely MovieLens 100 K, MovieLens 1 M, FilmTrust and BookCrossing datasets, with strong experimental results. We have compared our DAFERec method against mDA-CF, TrustSVD, SVD variants, BiasedMF, ItemKNN and I-AutoRec methods. The results demonstrate that our proposed solution beats the benchmarks and is a highly flexible model that works on different datasets solving different business problems like book recommendations, movie recommendations and trust.