UCLA

Various practitioners in building recommendation systems currently leverage deep learn- ing techniques. This thesis surveys several deep learning techniques to build recommender systems that serve personal recommendations. The paper begins with the traditional collaborative filtering model and builds several neural networks. The neural networks had similar structures but the main difference was the various feature sizes used to measure improvement on predictions. The neural networks were also fine-tuned by using dropout to prevent overfitting. Ensemble techniques are later applied in order to enhance recommendations. Finally, a bidirectional long short term memory (LSTM) was built to serve recommendations based on implicit feedback. The models are tested with an anime dataset from the web. The models were able to predict reasonable recommendations and both implicit and explicit recommendations can be made for users.