A recommendation system is an information filtering system to predict users' rating to items. Large-scale recommendation systems typically involve more than one billion users and items. Collaborative filtering based recommendation systems which were commonly used until recently suffer from the cold-start problem and the scalability problem making them unsuitable for large-scale recommendation systems. Model-based recommendation systems have the advantage of better performance and scalability in large-scale systems. Contextual signals, such as user browsing histories, check-ins, latest posts, page views, etc., provide great potential to improve model-based recommendation systems.

This thesis presents studies aimed at improving the performance of model-based recommendation systems with contextual signals. In particular, we propose: (i) a new linear embedding method for sparse features, thus, allowing us to quickly convert sparse features to latent vector space for further training while still preserving the comparable embedding quality to neural network embedding models. (ii) a new offline embedding refinement model that can refine learned embedding vectors using various contextual signals. (ii) an ensemble based online serving model for page recommendation systems. We demonstrate the robustness and effectiveness of our approaches by conducting various offline and online evaluations. More specifically, we apply t-sne, a nonlinear dimension reduction technique and visualize the embedding vectors generated by embedding models; the result shows that our embedding models successfully capture the location similarity from our training set. Also, when introduced as ranking features, our embedding vectors can efficiently improve offline AUC. Online A/B testing with T-statistic also prove that our models are suitable for a real-life application.

Concluding this thesis, we summarize our contributions as follows. (i) We find out that linear embedding model can have comparable embedding quality with neural network models. (ii) We discover that with a considerably sparse dataset, lock-free stochastic gradient descent can effectively improve training speed with almost no influence on model quality. (iii) We studied the importance of contextual signals in a large-scale recommendation system. (iv) We discover that learning representations from contextual data sets can play a critical role in improving the accuracy of the pairwise scoring model. (v) We conduct experiments on a real-life dataset.