Data Science for Healthy Online Interactions

The web enables users to interact with one another and shape opinion at an unprecedented speed and scale. However, the prevalence of disinformation and malicious users makes the web unsafe and unreliable, for example, 40% of users have experienced online harassment and platforms have disabled user comments because of trolling. In this talk, I will present data science methods that help us to create a better and safer web ecosystem for everyone. In particular, I will present methods to extract knowledge from the social graph structure and augment with behavior signals to characterize, detect, and mitigate the damage of disinformation and malicious users. First, I will describe a graph mining collective classification algorithm to identify fake reviews on e-commerce platforms. The method learns trustworthiness scores from the user-to-product review network to identify sophisticated fraudsters. The method is currently being used in production at Flipkart, India’s largest e-commerce platform. Next, I will present the first web-scale characterization of multiple account abuse in online discussions and my method of statistical analysis of user interaction graphs to detect them. Finally, I will show how learning embeddings from the social network structure helps to predict online conflicts and to mitigate their damage. These methods power online tools that help administrators in Reddit and Wikipedia. I will conclude the talk by describing my future research directions that will enable us to proactively predict how malicious behavior will evolve in the future, both on web platforms and face-to-face interactions.