A Framework for Modeling Influence, Opinions and Structure in Social Media

Akshay JavaAAAI 2007

Introduction

Traditional web modeling often treats the internet as a static collection of pages, but the "Blogosphere" is a dynamic, conversational ecosystem. Standard ranking algorithms like PageRank frequently fail to capture the nuances of social media influence, which is often topic-specific and emotionally charged. This research proposes a multi-layered framework to characterize information spread by modeling the interplay between social structure, user sentiment, and temporal changes.

Addressing the Nuances of Social Media

Social media presents unique challenges for data mining and influence modeling. Influence is rarely universal; a prominent technology blogger may have significant sway in their niche but zero influence on political discourse. Furthermore, links in social media are not always "votes" of confidence; they can represent rebuttals, disagreements, or "flame" posts. Finally, structural noise—such as blog-rolls, advertisements, and sidebars—can skew traditional data mining efforts by cluttering content with non-relevant links.

Methodology

The proposed framework consists of three primary components designed to address these challenges:

1. BlogVox: Opinion Extraction

BlogVox is an opinion retrieval system that utilizes an SVM-based classifier to estimate the likelihood that a post expresses an opinion on a specific topic. A critical innovation in this system is the use of heavy data-cleaning heuristics. By stripping away extraneous HTML elements like sidebars and ads before processing, the system focuses purely on the core content of the post.

Rather than treating every link as a positive edge, the framework assigns a vector to each edge representing both the topic and the sentiment (positive or negative). This allows for "trust propagation" models that can distinguish between communities of like-minded individuals and adversarial interactions, providing a more accurate map of social influence.

3. Generative Graph Modeling

To simulate realistic blog networks, the research extends traditional "preferential attachment" models. The proposed model assumes bloggers toggle between a Write Mode (creating content) and a Read Mode. In Read Mode, the blogger performs a modified random walk that includes a "recurrent state"—the blogger’s own homepage. This reflects the reality that social media users are "anchored" and frequently return to a familiar home base or set of favorite feeds.

Evaluation and Results

The framework was evaluated using several datasets, including the TREC 2006 Blog Track and real-world graph data. A key finding was the significant impact of data cleaning on retrieval performance. By filtering out the "noise" of blog templates, the system became substantially better at identifying both relevant topics and the opinions associated with them.

Performance comparison between unclean and cleaned indices for opinion and topic relevance retrieval.

As shown in the results above, the cleaned index provided a ~21% improvement in Mean Average Precision (MAP) for opinion retrieval and a ~17% improvement for topic relevance. This quantitative evidence underscores that cleaning is as vital as the modeling itself in social media analysis.

Discussion and Future Work

The research highlights that influence is often niche; in specialized communities, low-ranked blogs can serve as primary opinion leaders. However, several challenges remain. Natural language processing still struggles with linguistic nuances like sarcasm, contradictions, and the informal, often ungrammatical nature of blog text.

Future work will focus on moving toward ontological semantics to better understand the underlying reasons behind posts and developing more robust techniques for finding topic-specific sentiment in poorly structured text. Additionally, further modeling is required to capture how opinions and influence shift over time as specific "memes" or topics evolve.

Summary updated

BibTeX
@inproceedings{java2007framework,
  author = {Akshay Java},
  title = {{A Framework for Modeling Influence, Opinions and Structure in Social Media}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2007},
}
Figure