Harmless users’ detection models.

We analyze and propose a solution for the exposure of vulnerable users to harmful content during their interaction with contact recommender systems in social platforms. Our approach is dedicated to maximizing the number of harmless accounts suggested to users at risk. For these users, the over-personalization of recommender systems can result in an exposure to triggering content. We consider anorexia nervosa as a use case. People with anorexia tend to seek accounts of peers that support their unhealthy habits. Contact recommender systems can unintentionally reinforce such behaviors. Our approach modifies the objective function of a content and topology-based recommendation algorithm to maximize the suggestion of harmless accounts for users at risk. This is done with data from Twitter of Spanish speaking users with anorexia. The design and evaluation of the proposal has involved the participation of clinicians and volunteers at the last stages of treatment. Results show that users with anorexia are willing to follow harmless accounts suggested in online platforms. There is a tradeoff in precision (Pr) when comparing our proposal (Pr = 0.41) with a regular recommendation approach (Pr = 0.58). However, results are promising as there is a 55% increase in the percentage of harmless accounts suggested.