Interests Discovery in Social Networks Based on a Semantically Enriched Bayesian Network Model

Abstract

Online Social Networks have significantly consolidated the role of web users from content consumers to content producers as well. In these platforms, the textual form of user-generated content emerged as the most engaged and convenient form for the users to communicate their online community about their activities and interests. Discovering the interests of users based on their user-generated textual contents can enrich the user profile with implicit and dynamic information that are semantically related in order to enhance the User Model. This enhancement in the User Model can elevate the performance of personalized web applications and services such as recommender systems. Discovering the topic of interests of users from their user-generated textual contents is still an open research challenge. Contrarily to traditional text documents, user-generated textual contents are highly focused, not domain specific, short in length, informal, multilingual, unstructured and grammatical error prone text messages. These lingual characteristics make it inapplicable to use the standard information retrieval techniques such as text classification and Natural Language Processing to discover the interests of users in an efficient way. These techniques require structured and grammatically correct text to be able to catch the implicit syntactic relations (grammatical links) between terms. To be able to automatically catch the explicit semantic (occurrence associations within text) of a term efficiently, these techniques require high terms frequency in the text. In addition, user-generated textual contents could have implicit semantic relations between entities (semantically related entities) that play a big role in discovering the topic of interest. Usually, the user submits social-posts at different points in time. Knowing the temporal factor can dramatically affect in catching the semantic relations between the contents of user-generated textual content. The implicit syntactic, explicit and implicit semantic and temporal relations are factors that yield to uncertainty in inferring the right topic of interests of a user based on his user-generated textual contents. Based on the causal implicit relation between the components of user-generated textual contents, we introduce a framework to discover the topic of interests of users in Online Social Networks based on user-generated textual content. The proposed framework able to extract the proper content, extract the proper features, semantically enrich these features, and represent them in a Bayesian Network model that can catch the explicit, implicit and temporal relations and can infer topic of interests efficiently. This dissertation develops novel methods and algorithms for the interests discovery based on user-generated textual content. Of particular interest in this scenario is the identification of the topic of interest using Bayesian Network model, which helps to inference under uncertainty. However, the sparsity of user-generated textual content and the ambiguity of natural languages and vocabularies used to generate them present problems that go beyond the capabilities of typical text classification techniques. The algorithms and models introduced in this dissertation are especially tailored to reduce the effects of the lingual characteristics of user-generated textual content in order to derive more reliable interests discovery model. Experimental evaluation for different settings and various datasets show that the proposed framework introduce solutions that outperform the current state of the art.