Developing Self-Advocacy Skills through Machine Learning Education: The Case of Ad Recommendation on Facebook

Abstract

Facebook users interact with algorithms every day. These algorithms can perpetuate harm via incongruent targeted ads, echo chambers, or "rabbit hole" recommendations. Education around the machine learning (ML) behind Facebook (FB) can help users to point out algorithmic bias and harm, and advocate for themselves effectively when things go wrong. One algorithm that FB users interact with regularly is User-Based Collaborative Filtering (UB-CF) which provides the basis for ad recommendation. We contribute a novel research approach for teaching users about a commonly used algorithm in machine learning in real-world context -- an instructive web application using real examples built from the user's own FB data on ad interests. The instruction also prompts users to reflect on their interactions with ML systems, specifically Facebook. In a between-subjects design, we tested both Data Science Novices and Experts on the efficacy of the UB-CF instruction. Taking care to highlight the voices of marginalized users, we use the application as a prompt for surfacing potential harms perpetuated by FB ad recommendations, and qualitatively analyze themes of harm and proposed solutions provided by users themselves. The instruction increased comprehension of UB-CF for both groups, and we show that comprehension is associated with mentioning the mechanisms of the algorithm more in advocacy statements, a crucial component of a successful argument. We provide recommendations for increased algorithmic transparency on social media and for including marginalized voices in the conversation of algorithmic harm that are of interest both to social media researchers and ML educators.