Embedding metadata using deep collaborative filtering to address the cold start problem for the rating prediction task

In recent years, deep learning has yielded success in many research fields including machine translation, natural language processing, computer vision, and social network filtering. The area of deep learning in the recommender system is flourishing. Previous research has relied on incorporating metadata information in various application domains using deep learning techniques to achieve better recommendation accuracy. The use of metadata is desirable to address the cold start problem and better learning the user-item interaction, which is not captured by the user-item rating matrix. Existing methods rely on fixed user-item latent representation and ignore the metadata information. It restricts the model performance to correctly identify actual latent vectors, which results in high rating prediction error. To tackle these problems, we propose a generalized recommendation model named Meta Embedding Deep Collaborative Filtering (MEDCF), which inputs user demographics and item genre as metadata features together with the rating matrix. The proposed framework primarily comprises of Generalized Matrix Factorization (GMF), Multilayer Perceptron (MLP), and Neural Matrix Factorization (NeuMF) methods. GMF is applied to the rating matrix, whereas MLP is applied to metadata. Using NeuMF, the outputs for GMF and MLP are then concatenated and input to a neural network for rating prediction. To prove the effectiveness of proposed model, two metrics are used, Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE). The MEDCF model is experimented on MovieLens and Amazon Movies datasets showing a significant improvement over the baseline methods.