Modelling Group Recommender Systems

In this thesis, the goal is to systematically study how to model group for recommender systems with different kinds of auxiliary group information. 1) Social-aware recommendation approaches assume that the knowledge in social user-user connections can be shared and transferred to the domain of user-item interactions, whereby to help learn user preferences. However, most existing approaches merely adopt the first-order connections during transfer learning. We argue that better recommendation performance can also benefit from high-order social relations. We propose a novel Propagation-aware Transfer Learning Network (PTLN) based on the propagation of social relations to better mine the sharing knowledge hidden in social networks. Specifically, we explore social influence in two aspects: higher-order friends are utilized by order bias; different friends in the same order will have distinct importance for recommendation by an attention mechanism. Besides, we design a novel regularization to bridge the social relations and user-item interactions gap. 2) The existing social-aware recommendation works estimate the impact of user's friends in accordance with the similarity between the primary users and their friends while neglect the friend intimacy. Further, they suppose that the user criteria are the same. However, some users may give full ratings to moderate items while some users only have the same rating on high-quality items. Therefore, if we cannot distinguish these differences, the central node's representation will deviate when aggregating each order's neighboring information. We further point out that friends' impact in different orders should also be various while the existing methods compare the impacts of different orders' friends based on the aggregated representations at this order. Here we offer a novel Graph Convolution Network - MBGCN applying influence bias to distinguish neighbors' impacts on users and items. 3) Not limited to social networks, group relationships need to be expanded to a larger scope. It is unclear how we may obtain the information on the grouping effectiveness. We have no idea of the effectiveness of a formation/grouping if they never played together. Sub-groups may perform more superbly when they functioned together compared with other combinations. The ability to predict the effectiveness of any combination will bring about a tremendous advantage for recommendations. However, it is infeasible to examine the effectiveness of each sub-groups, as the number becomes prohibitively large. We proposed a novel Tensor Completion Cohesion model that can determine the effectiveness of any subgroup formation based on observing only a very limited number of them.

Author dc:creator

• Chang, Haodong

Statement dc:rights

• The author owns the copyright in this thesis including all reproduction and reuse rights for the work. The work may not be altered without the permission of the copyright owner. Attribution is essential when quoting or paraphrasing from this thesis.
• au.edu.uts.lib/ppc
• info:eu-repo/semantics/openAccess

Language dc:language.iso

en_US

Handle dc:identifier.uri

http://hdl.handle.net/10453/160671

OAI identifier oai:identifier

oai:opus.lib.uts.edu.au:10453/160671