Abstract/Summary

In the online world, especially the e-commerce websites, the interactions among various users and products can be naturally modelled and studied as complex networks. The network-based similarities, also known as the association rules, have thus found wide applications in examining the co-accessing patterns among products and providing recommendations for users accordingly. Focusing on two major forms of online recommendations, including personalised recommendations which are made for specific users, and recommendation networks which are hyperlinks connecting similar products as a networked system, this thesis explores the application of network-based similarity measures in recommendations, and examines the performances of them. For the personalised recommendation, the recommendation list for a specific user is shown to be changing vastly when the system evolves, due to the unstable quantification of object similarities, which is defined as the recommendation stability problem. To improve the recommendation stability is thus crucial for the user experience enhancement and the better understanding of user interests. By ranking the similarities in terms of stability and considering only the most stable ones, this thesis presents a top-n-stability method based on the Heat Conduction algorithm for tackle the stability problem as well as guarantee the accuracy and diversity of recommendations. Experiments on four benchmark datasets indicate that the proposed algorithm can significantly improve the recommendation stability and accuracy simultaneously and still retain the high-diversity nature of the Heat Conduction algorithm. Furthermore, we show that the dilemma among stability, accuracy and diversity is caused by the popularity bias of network-based similarity measures, that the popular objects tend to have more common neighbours with others and thus are considered more similar to others. Such popularity bias of similarity quantification will result in the biased recommendations, with either poor accuracy or poor diversity. Based on the bipartite network modelling of the user-object interactions, this thesis calculates the expected number of common neighbours of two objects with given popularities in random networks. A Balanced Common Neighbour similarity measure is accordingly developed by removing the random-driven common neighbours from the total number. Recommendation experiments in three data sets show that balancing the popularity bias in a certain degree can significantly improve the recommendations’ accuracy and diversity simultaneously. Objects such as products, news, articles in most online systems are connected to similar others through hyperlinks as recommendations for users. Recommendation networks of objects are thus resulted enabling users to explore the massive relevant online information by surfing from one to another. While it connects overwhelming online objects as networks and seems to be a good way for users to navigate the haphazard content-browsing systems, two outstanding questions still exist that, 1) can the users locate their interests by surfing on the network, and 2) is every object accessible in the network? By mining the co-accessing pattern among objects, we construct recommendation networks according to the object similarity matrix, and thereby theoretically explore its topology and dynamics. Modelling the users’ surfing behaviour as random walks, we examine how many history records of a target user can be retrieved during such process. Most measures are shown with limited accuracy and cannot help users to explore niche objects which may be not popular but fit some users’ interests. In order to achieve a good accuracy quickly in a short-term random walk, we show that the recommendation list should be short, where each object is expected to have generally 2~6 recommended objects. In terms of accessibility, the recommendation networks are shown to be unnavigable due to the emergence of traps, which are dense communities with few or even no links connecting outside. Such vicious cycles trapping surfing users constantly make a handful of objects dominating most of the web traffic. According to the local structure of the network, a simple measure entitled the local return rate is developed, which can be used to accurately and efficiently identify the significant traps in large-scale recommendation networks. To summarise, this thesis uncovers some fundamental challenges with network-based online recommendations including the stability problem and popularity bias for personalised recommendation and the information monopoly for recommendation networks. In addition to the proposed algorithm, measure and analytical methods, the results inform the needs of more careful system design in practice, and shed lights on the future studies on such challenges to develop better network-based similarity measures.