Improving the classification of newsgroup messages through social network analysis

Newsgroup participants interact with their communities through conversation threads. They may respond to a message to answer a question, debate a topic, support or disagree with another person's point, or digress and write about a different subject. Understanding the structure of threads and the sentiment of the participants' interaction is valuable for search and moderation of newsgroups. In this paper, we focus on automatic classification of message replies into several types. For representing messages we consider rich feature sets that combine the standard author reply-to network properties with features derived from four additional structures identified in the data: 1) a network of authors who participate in the same threads, 2) network of authors who post similar content, 3) network of threads sharing common authors, and 4) network of content-related threads. For selected newsgroups we train linear SVM classifiers to identify agreement and disagreement with the original message, and question and answer patterns in the threads. We show that the use of newly defined features substantially improves classification of messages in comparison with the SVM model based only on the standard reply-to network.