Photo courtesy of paulworthington

Congratulations to Dr. Sara Javanmardi on passing her Ph.D. thesis defense titled, “Measuring Content Quality in User Generated Content Systems: Towards a Machine Learning Approach”.

Abstract: User generated content systems (UGCs) such as wikis, blogs, and shared forums allow millions of users to share information. By their very nature, these systems contain information of different quality levels. The open nature of these systems makes it difficult for users to determine the quality of the information and the reputation of its providers. In this thesis I leverage machine learning techniques to model user behavior and content quality in Wikipedia. First, I parse and mine the entire English Wikipedia history pages in order to extract detailed user edit patterns and statistics. Then I use these patterns and statistics to derive several computational models of a user’s reputation. I evaluate the reputation predictions generated by these models on Wikipedia users and show that the models can be used to efficiently predict user behavior in Wikipedia. Secondly, based on these models, I generate several new user reputation features and show that they are strong predictors for locating low quality content in Wikipedia in the form of vandalism. To improve the accuracy of my approach, I extend the feature set by adding other features such as textual features, language model features, and meta data features. I describe a method for training classifiers for vandalism detection that are more accurate than other classifiers previously developed. Because of the high turnaround in user generated content systems, it is important for vandalism detection tools to be scalable and run in real–time. Therefore, based on a MapReduce paradigm, I use feature selection to find the minimal set of features that yield consistently high levels of accuracy. Because some features are more costly to compute than others, I use cost–sensitive feature selection to reduce the total computational cost of executing our models. I show that user reputation features contribute significantly to classifier performance. Finally, I use Amazon Mechanical Turk to classify vandalism into different categories and study how well I am predicting vandalism in each category. The methods I use in this thesis are general and can be applied to numerous other user generated content systems: the features we use are appropriate for use in other domains such as Facebook and Twitter, and the machine learning approaches we use are general and can be applied to other UGCs. This work can be used in several applications; one of them being search engines in which the measuring of content quality is a must for improving the quality of its ranking search results.

Congratulations Sara!