The impact of transparency and imitation over complex networks in strategic classification

Abstract

Classification algorithms are widely used in critical domains such as healthcare, bank loans, credit and fraud detection. These systems should be transparent, yet it remains unclear how individuals will use explanations to adjust their own features. Individuals often access multiple sources of information, from insights provided by institutions to experiences shared among peers. Based on the information received, individuals may decide to strategically adapt to obtain a favourable outcome, honestly improving or attempting to game the system. This paper studies the impact of transparency and social information on strategic classification. We assume that agents adapt based on best response and behavioural imitation along the edges of social networks. We observe that increasingly opaque decision rules can negatively impact the utility of institutions, especially in dense social networks. The number of False Positives is reduced in networks with a lower average degree, when users imitate the average behaviour, as opposed to the most extreme behaviours. When imitating the most extreme behaviour among their connections, users change their features to a large extent in networks with a higher average degree (i.e., higher density). This applies to both honest improving and gaming, with more pronounced impacts in the case of the latter, creating an additional source of risk for institutions. Our model and results reveal that behavioural imitation patterns and social network effects influence the downstream effects of algorithmic transparency.