Advanced Collaborative Filtering and Image-based Recommender Systems

Abstract

Due to burst of growth of information available all over the world, it has been of great necessity to retrieve most suitable data from the entire warehouse for each unique person. E-commerce can be considered as an example which possess vast amount of products that need to be personalized and then recommended to customers. In this case, Recommender Systems (RS) have been frequently studied in recent years. In this thesis, we firstly focus on improving performance of Collaborative Filtering (CF) and then develop new approaches to alleviate the cold-start problem.

We firstly improve the SGD-base Matrix Factorization method, which is one of the most effective CF approaches, by taking into consideration item attributes. The developed MFA model treat categorical information of items as virtual users and then insert ratings of maximum value into the original rating matrix if an item belongs to the category. Then k-nearest-neighbour (KNN) method is combined with MFA model to make further improvement. A threshold is then set during SGD to filter out the most misleading ratings before the SGD is applied again to train the factors.

The second section of our research is to use other side information with user's past purchase record to generate future recommendations. At first, helpfulness information (i.e. helpful votes provided by other members in the community) of each review is used to produce the Weighted Matrix Factorization (WMF) model. Then timestamp of each rating is substituted into calculation of factors. In this section, WMF and TWMF are implemented on chronologically sorted datasets, so that it could simulate more real situation.

In the last section, we assume that all products to be recommended are completely new to the system, meaning that none of their information such as colour, price, categories can be substituted into calculation. Therefore, classic CF models will fail to learn factors of these new products. In this case, we develop image-based RS to recommend visually similar products, in which SSIM/CW-SSIM, CNN with KNN, CNN with ridge regression and CNN with SGD are used to achieve the target, and all of them are proved to be effective.