Cross domain recommender systems using matrix and tensor factorization

Today, the amount and importance of available data on the internet are growing exponentially. These digital data has become a primary source of information and the people’s life bonded to them tightly. The data comes in diverse shapes and from various resources and users utilize them in almost all their personal or social activities. However, selecting a desirable option from the huge list of available options can be really frustrating and time-consuming. Recommender systems aim to ease this process by finding the proper items which are more likely to be interested by users. Undoubtedly, there is not even one social media or online service which can continue its’ work properly without using recommender systems. On the other hand, almost all available recommendation techniques suffer from some common issues: the data sparsity, the cold-start, and the new-user problems. This thesis tackles the mentioned problems using different methods. While, most of the recommender methods rely on using single domain information, in this thesis, the main focus is on using multi-domain information to create cross-domain recommender systems. A cross-domain recommender system is not only able to handle the cold-start and new-user situations much better, but it also helps to incorporate different features exposed in diverse domains together and capture a better understanding of the users’ preferences which means producing more accurate recommendations. In this thesis, a pre-clustering stage is proposed to reduce the data sparsity as well. Various cross-domain knowledge-based recommender systems are suggested to recommend items in two popular social media, the Twitter and LinkedIn, by using different information available in both domains. The state of art techniques in this field, namely matrix factorization and tensor decomposition, are implemented to develop cross-domain recommender systems. The presented recommender systems based on the coupled nonnegative matrix factorization and PARAFAC-style tensor decomposition are evaluated using real-world datasets and it is shown that they superior to the baseline matrix factorization collaborative filtering. In addition, network analysis is performed on the extracted data from Twitter and LinkedIn