Explainable online recommendation systems with self-identity theory and attribute learning method

Li, Zequn (2019) Explainable online recommendation systems with self-identity theory and attribute learning method. Doctoral thesis, Northumbria University.

Preview

Text (Doctoral Thesis) li.zequn_phd.pdf - Submitted Version Download (4MB) | Preview

Abstract

In recent years, Online Shopping plays an important role in daily life and how to improve the online shopping experience with Machine Learning and Recommender System has been discussed by a group of researchers. As a sub-field of Machine Learning, Computer vision has achieved significant developments during the last decade. The computer vision techniques can help machine to view images and extract useful information from images like human beings. However, the existing online recommender system has mostly used the labelled information and ignored the large amount of useful information extracted from images. This thesis proposed that the extracted information from images through computer vision techniques can be used in the current online recommender system for the improved online shopping experience. To do this, I firstly tackled the problem of insufficient data in the real online shopping environment. I proposed a pairwise constraint random forest algorithm with associating transfer learning strategy. This new algorithm can make use of weakly supervised labelled data which is relatively easy to collect in the real online shopping environment to train the attribute classification model. Secondly, I developed an explainable recommender system with self-identity theory. This new recommender framework is built based on the weakly learning algorithm proposed above to analyse human behaviours by self-identity theory from information system research. Compared with previous recommender system, my work concentrates on different customer behaviours distinguished by self-identity and result in an improved online shopping experience.
In summary, there are two major contributes for this thesis. Firstly, this thesis introduces a new weakly-supervised learning approach for semantic data classification in the online shopping environment. This new algorithm can work with noise partially labelled data to achieve better accuracy for attribute learning tasks. Secondly, by analysing the recommender system with self-identity theory, a new explainable Recommender System is proposed to improve online shopping experience. Besides, we also indicate the potential of further research in combining Computer Vision in Computer Science with online shopping experience in Information System research which can determine how Computer Vision can help to solve real world problems.

Actions (login required)

View Item

Downloads