Algorithmic Fairness for Networked Algorithms

Ana-Andreea Stoica

Abstract: Recent evidence points to the detrimental effects of algorithmic deployment on human datasets, as often times such algorithms mirror and exacerbate existing inequalities in the input data. This work focuses on understanding the disparate effects of algorithms on social inequality and building theory and applications for graph algorithms with ramifications in the way we learn information online and offline. We show that in the case of recommendation algorithms, the most common heuristics that learn connections for providing social recommendations exacerbate disparity between different communities in a bi-populated network by reinforcing certain patterns in the network, such as homophilic behavior. Similar results occur for content recommendation, where we show that minority viewpoints are being further diminished by algorithms that learn relational data and over-recommend a majority viewpoint. On the other hand, algorithms may leverage community affiliation to disperse information in a network in a more effective manner while being more equitable in terms of the demographics reached in certain conditions. For such studies, we find closed-form conditions of the results using graph theoretical models that replicate inequality in social networks and use them to develop a set of algorithms that use network statistics to diffuse information in a feature-aware way, effectively reaching more communities than the status quo heuristics that are blind to sensitive features. Through validation on real-world data, we show that such learning algorithms benefit from being feature-aware in learning relational data in order to mitigate bias.