Techniques and facilities for personalized search in folksonomy

: profiling, community mining and context modeling

Abstract

With the fast development and popularity of social media applications, user-generated data has been growing tremendously in Web 2.0 era. A typical example is the folksonomy (a.k.a collaborative tagging system), which not only permits users to share and upload their resources (e.g. photos, videos or URLs), but also allows users to label their interested ones with semantic-rich tags. Confronting such a big volume of resources, people need an effective and efficient method of exploring and indexing to find their desired data. Personalized search is an indispensable technique to achieve this goal. Therefore, our research focuses on techniques and facilities for personalized search in folksonomy. In this dissertation, we start with the research topic of user and resource profiling. As folksonomy (collaborative tags) can be viewed as semantic-rich information collection generated by "the wisdom of crowd". The core idea of profiling is to extract relevant pieces of information (tags) to depict users and resources. In particular, user and resource profiles are represented by vectors of relevant tags in folksonomy. Moreover, we propose a social-filtering resource profiling method to resolve the tag confliction raised by collective resource profiling in existing methods in order to facilitate personalized search. Next, we proceed to latent community mining in the folksonomy, in view that users with similar taste may share common or close views on resources and they form a latent user community. To fully explore the relationships among users, resources and tags, we propose a facility called Augmented Folksonomy Graph (AFG) by incorporating both resource contents and tag semantic similarities into folksonomy. Furthermore, we adopt a novel proximity to measure the similarity between any pair of users by random walk distance, and devise a prototype-based clustering method to discover the hidden user community from AFG. The mined communities not only facilitate the tag-based personalized search but also can be applied in content-based one. The last relevant issue addressed by this thesis is on how to model the search context (particularly verbal context) in a certain search task, since the contextual information is indispensable and valuable to understand user preferences and purposes. We propose a nested context model to contextualize user profiles for different queries, so as to tackle the shortcomings of uniform treatment of user profiles in conventional methods. Notably, the nested context model can be extended to verbal contextual graph, the latter can be utilized to find "dominating context" to further improve the performance of personalized search. To sum up, the techniques and facilities devised in this thesis can assist users to find their demanded resource effectively in a personalized manner. As a part of the research, we have conducted extensive experiments on public or/and collected data sets to evaluate our approaches. The empirical results validate the rationale and effectiveness of the proposed methods in terms of personalized search in folksonomy.

Date of Award	14 Feb 2014
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Qing LI (Supervisor)