Influence diffusion detection between linked bloggers

The easy access and availability of blogs have encouraged web-users to change from consumers to providers of information. Providers of such content exert a certain level of influence on the receivers, giving rise to the interest in influence detection within the blogosphere. Previous works have focused on simple blog features to detect blogosphere influence which may not yield accurate results as influence itself is a complex concept that requires more in-depth analysis, where other intrinsic blog features, such as sentiments and agreement expressed in the blog posts or bloggers’ influence styles, that provide further knowledge on the blogs’ influence could be used to improve influence diffusion detection performance. This study aims to develop a model that uses sentiments on common topics, agreement expressions, and influence styles as possible features to detect influence diffusion between linked bloggers. Influence is defined and limited to the scope in the capacity of the linked blogger to exert on the linking blogger to agree with or have similar sentiments on the discussed topics.The objectives addressed in this dissertation study are: 1) Identify the relevant blog features that are useful in detecting influence diffusion; 2) Establish an automatic sentiment analysis approach that takes into consideration the complex relationships between words in the blog posts for influence diffusion detection; and 3) Develop an influence diffusion detection model using similar sentiments between the linking and linked bloggers, agreement towards the linked bloggers, and bloggers’ influence styles as features to improve performance. The first phase of the study involved statistical analysis on the blog dataset to identify the various blog features that could indicate influence. Results from the initial study show that similar sentiments on common concepts between linked blog posts give the clearest indication of influence as compared with other blog features. Recent sentiment analysis studies focused on the functional relationships between words using typed dependency parsing, which provides a refined analysis on the grammar and semantics of textual data. Consequently, the second phase of the study evaluated and established an automatic sentiment analysis process based on a linguistic and semantic analysis approach, which further considered the complex relationships between words in the context of the target terms. Context in this case refers to the neighboring terms of a current term within the blog post. In addition, influence as a concept could be correlated with other factors such as the different influence styles. Bloggers are not restricted to a monolithic description of influence based on link existence, and may differ in the manner in which they exert influence. Knowing the bloggers’ influence styles can better describe how influence is propagated in the network. Based on the findings in phase one and two, the third phase analyzed bloggers’ influence styles and developed the Influence Diffusion Detection (IDD) model to automatically detect influence diffusion between linked bloggers using the sentiments on common topics between the linking and linked bloggers, agreement towards the linked bloggers, and bloggers’ influence styles as features. Influence style refers to the manner in which a blogger exerts influence through the blog postings, and can be described through the engagement, persuasion, and persona styles of the blogger. The results show that The IDD model performed well (average 76% F-score) compared to the in-degree and sentiment values baseline approaches as the model closely maps the influence characteristics of the bloggers to the links between them through providing a clearer description of the manner in which influence was exerted. Though this dissertation study focused on the analysis of influence diffusion between bloggers and their immediate neighbors, the influence diffusion across a blogosphere network could be inferred through analyzing the influence diffusion through subsequent linked bloggers. The main contributions of this study are the use of sentiment analysis to extract the opinions of the linking and linked blog posts on the topics discussed to detect influence diffusion, and the introduction of influence styles as features in an influence diffusion detection model approach. As a practical implication, the IDD model could automatically detect influence diffusion between linked bloggers with higher accuracy and further describe the detected influence in relation to the influence styles of the linked bloggers. In addition, the influence style of each blogger along the influence path could be plotted to visualize the chain of influence styles pattern in the bloggers’ network.DOCTOR OF PHILOSOPHY (WKWSCI