User Session Recommendations Based On Deep Neural Networks

Recently, as the Internet is spreading widely. People are depending on recommender systems to help them make choices. Given a customer browsing e-commerce website, how can we recommend products for him/her interactively? To address this problem, in this dissertation, we first modelled the browsing process as a linear sequence of web pages. The basic idea of one possible solution is to made prediction on what the user will buy finally, while the user is pushing on to the next state of the page sequence. Unfortunately, the previous Collaborative Filtering algorithms are unable to recognize the pattern of the user’s browsing behavior.Thus, to address the modelling of the problem, we proposed a machine learning model called "DeepSession", which is based on the combination of Deep Recurrent Neural Networks and Deep Feed-forward Neural Networks. The Deep Recurrent Neural Networks perform well on extracting the pattern of sequential information, and make prediction on what product the user will buy. However, the number of states in Recurrent Neural Networks is always fixed, but the length of user sessions is increasing dynamically. Hence we proposed a history state as an extra state for the Recurrent Neural Networks, which aggregates all the stale pages. In addition, we also attempted to build a Deep Feed-forward Neural Networks simulating the traditional Collaborative Filtering algorithm, and modeling the history of what the user purchased. Combining both RNN and FNN models, the model gets a better prediction. On the other hand, it is well known that tuning Neural Networks is a tedious job. So we also proposed an automatic tuning framework with a heuristic algorithm, which improves the efficiency of the tuning process. And based on the DeepSession model, we designed a recommender system generalized for e-commerce websites.At the end of this dissertation, we conducted a series of experiments on a real world dataset sponsored by an e-commercial website. The results show that the DeepSession algorithm predicts better than collaborative filtering implemented by Spark ALS significantly and that the DeepSession algorithm is a good solution for the user session recommendation problem.