TDCF: A two-stage deep learning based recommendation model

Deep neural network (DNN) recently has been used in recommendation tasks and attained record performance. However, the input quality of the DNN model has a great influence on the recommendation performance. An efficient and effective initialization method is proposed in this work. Specifically, we propose a recommendation model based on two-stage deep learning. In the first stage, two separate marginalized stacked denoising auto-encoder (mSDA) models are applied to the user and item features to learn the latent factor vectors. In the second stage, the resulting latent factor vectors are combined and used as input vector for the DNN component to optimize the entire model. Extensive experiments with real-world datasets indicate that the proposed framework shows excellent recommendation performance compared with the state-of-the-art methods, even in the data sparse environment.