Discovering, assessing, and mitigating data bias in social media

Social media has generated a wealth of data. Billions of people tweet, sharing, post, and discuss everyday. Due to this increased activity, social media platforms provide new opportunities for research about human behavior, information diffusion, and influence propagation at a scale that is otherwise impossible. Social media data is a new treasure trove for data mining and predictive analytics. Since social media data differs from conventional data, it is imperative to study its unique characteristics. This work investigates data collection bias associated with social media. In particular, we propose computational methods to assess if there is bias due to the way a social media site makes its data available, to detect bias from data samples without access to the full data, and to mitigate bias by designing data collection strategies that maximize coverage to minimize bias. We also present a new kind of data bias stemming from API attacks with both algorithms, data, and validation results. This work demonstrates how some characteristics of social media data can be extensively studied and verified and how corresponding intervention mechanisms can be designed to overcome negative effects. The methods and findings of this work could be helpful in studying different characteristics of social media data.