Content-based recommender system for an academic social network / Vala Ali Rohani

Vala Ali, Rohani (2014) Content-based recommender system for an academic social network / Vala Ali Rohani. PhD thesis, University of Malaya.

Preview

PDF (Full Text) Download (5Mb) | Preview

Abstract

The rapid growth of Web 2.0 applications, such as blogs and social networks creates rich online information and provides various new sources of knowledge. The situation, however, leads to a great challenge in terms of information overload among social network users. Recommender systems (RSs) alleviate this problem with a technique that suggests relevant information from the abundance of Web data by considering the user’s previous preference. Collaborative and content-based are the recommendation techniques typically used in existing RSs. The content-based method is employed more widely though. Similar to the collaborative, the content-based technique suffers from the cold-start dilemma that is caused by the incapability of RSs to make reliable recommendations in situations when new items or new users are involved. Such issues have an impact on prediction accuracy in existing algorithms, and hence, a better approach is required. In this study, a new algorithm is proposed to represent an enhanced version of content-based recommender systems by utilizing social networking features. In its formulation, the algorithm considers the interests and preferences of users’ friends and faculty mates in addition to users’ own preferences. The algorithm exploits all interests and preferences in a hierarchy tree structure. Since no offline data on Academic Social Networks (ASNs) exists and concerning the advantages of online study benefits, a real runtime environment of ASN called MyExpert was built in order to conduct an online study to assess the four recommender algorithms. Each recommender system algorithm, including the enhanced version of the content-based recommender systems using social networking (ECSN), is later incorporated into MyExpert to propose to members of this online society the most relevant academic items including jobs, news, scholarships and conferences. By using MyExpert, the online study was carried out to collect real feedback from live interactions between iv users and the system. The assessment ran for 14 consecutive weeks from 7th September to 26th December, 2012. MyExpert had 920 members from 10 universities in Malaysia at the time of evaluation. Four metrics, namely precision, recall, fallout, and F1 were employed to measure the prediction accuracy of each algorithm. Although the experiment conducted presented some threats, the results indicated that the ECSN algorithm not only improves the prediction accuracy of recommendations but also resolves the cold start problem in the existing recommender systems algorithms.

Actions (For repository staff only : Login required)