Tree Similarity Measure-based Recommender Systems

Abstract

The rapid growth of web information provides excellent opportunities for developing e-services in many applications but also caused increasingly severe information overload problems whereby users are not able to locate relevant information to exactly meet their needs efficiently by using the current Internet search functions. A personalised recommender system aims to handle this issue.A big challenge in current recommender system research is: the items and user profiles in many recommender system applications nowadays, such as the e-business and e-learning recommender systems, are so complex that they can only be described in complicated tree structures. Therefore, the item or user similarity measure, as the core technique of the recommendation approach, becomes a tree similarity measure, which existing recommender systems cannot provide. Another challenge is that in many real life situations, online recommendations to customers in selecting the most suitable products/services are often made under incomplete and uncertain information, which needs fuzzy set theory and techniques to deal with. Thus, how to use fuzzy set techniques to handle data uncertainty issues in tree-structured items or user profiles needs to be investigated.This research aims to handle these two challenges in both theoretical and practical aspects. It first defines a tree-structured data model, which can be used to model treestructured items, user profiles and user preferences. A comprehensive similarity measure on tree-structured data considering all the information on tree structures, nodes' concepts, weights and values is then developed, which can be used to compute the semantic similarity between tree-structured items or users, and the matching degree of items to tree-structured user requests. Based on the tree-structured data model, the tree-structured iv items and user requirements are modelled as item trees and user request trees respectively. An item tree and user request tree-based hybrid recommendation approach is then developed. To model users' fuzzy tree-structured preferences, a fuzzy preference tree model is proposed. A fuzzy preference tree-based recommendation approach is then developed. Experimental results on an Australian business dataset and the Movielens dataset show that the proposed recommendation approaches have good performance and are well-suited in dealing with tree-structured data in recommender systems. By use of the proposed tree similarity measure and recommendation approaches based on that, two real world applications, a business partner recommender system, Smart BizSeeker, and an e-learning recommender system, ELRS, are designed and implemented, which demonstrate the applicability and effectiveness of the proposed approaches. v