Personalised recommendation : challenges and experimental issues

Abstract

With the shift towards an increasingly digital lifestyle, recommender systems play a critical role in helping consumers to find the best product or service amongst a variety of options. Unsurprisingly, personalised recommendations have become part and parcel of our daily lives. For instance, recommender systems are widely adopted across various domains, including e-commerce platforms (e.g. Amazon, eBay, Taobao), location-based social networks (e.g. Yelp, Foursquare), and social media (e.g. Facebook, Instagram, Twitter). Arguably, both the importance and practicability of recommender systems have been a key driving force behind the sustained interest from both academia and industry. Nevertheless, there are various challenges and experimental issues which affect the predictive performance and/or robustness of a recommendation system. In this dissertation, we propose novel hybrid models to overcome a long-standing challenge for personalised recommendation, i.e. the cold-start problem, by leveraging different types of content information in conjunction with recent advances in deep learning. Furthermore, we identify and examine challenges, as well as experimental issues, that persist in personalised recommendation.