Towards a new generation of deep learning based recommender systems

Nowadays, modern online services that have huge amount of consumable data often rely on recommender systems to distribute suitable and engaging information to their users. Only invented twenty years ago, recommender systems now have become an essential part in many products of big technology companies. From the inception, recommendation models are relatively simple algorithms such as item-item collaborative filtering and textual content-based recommendation. However, as the application's functional complexity, customer's demands and volume of data keep increasing over the years, the design and modelling of recommender systems also evolve rapidly. One clear trend we are seeing is the intertwine of deep learning methods into main design principles of the recommender system such as collaborative filtering or content-based principles. Not only this approach improves performance and scalability of recommendation engines, but also opens door for new and advanced recommender systems applications, and we address this new type of recommender systems as deep recommender systems.

In this thesis, we investigate this new area of deep recommender systems and have the following contributions. We first analyse how deep learning model's hyperparameters affect the performance of a recommender systems. Extensive analysis allows us to understand the learning capability of deep learning models when comparing to traditional recommender systems. Our second contribution is the exploration of graph convolutional neural network (GCN) methods in handling recommendation tasks. In this work, we develop the HeteGraph framework bases on GCN principles, and evaluate HeteGraph on popular recommendation tasks. Our in-depth experimentation shows high-performance results of the popular recommendation tasks which are item-rating prediction and diversified-item recommendation. The third contribution in this thesis project is our research summary on a new type of deep recommender system application, which is deep conversational recommender system. Deep conversational recommender system has a unique composition of a dialogue system and an integrated recommendation module. Our research explains important concepts, and effective approaches to design such a complex system, which we believe are fairly useful to readers who are interested in this new application domain. Our final contribution is the development of a news recommendation model called CUPMAR. News recommendation task has gained lots of attention lately due to the advances in natural language processing (NLP) techniques thanks to deep neural networks. Thus, by integrating the state-of-the-art NLP deep learning methods to recommender systems, we create CUPMAR model and achieve the state-of-the-art performance on several evaluation metrics. We conclude our thesis with the discussion about future directions of deep recommender systems. We hope this thesis will be a valuable resource for researchers and practitioners who want to learn and contribute to this vibrant research area of deep recommender systems.