An Algorithmic Audit of Online Matrimonial Platforms in India: Biases in Matchmaking Recommendation Systems

Online matrimonial platforms enable people to discover prospective partners who mutually meet each other’s preference criteria. Matrimonial relations in the Indian context have traditionally been considered as a union of two families, where a variety of attributes like age and income, and family preferences such as religious beliefs and class and caste hierarchies are matched, making it a complex process. Many such preferences may, however, be considered regressive and arising from entrenched biased social norms, and online matrimonial platforms may reinforce such biases through rule-based and machine-learning based algorithms. Some such norms for opposite sex matching preferences are for males to match with females having an income lower than them, and similarly to match with females having an age lower than them. Such norms are arguably biased and may reinforce gender-based inequalities, for example, where females may marry young while males build their careers. Online matrimonial platforms may present multiple opportunities to counter such biased social norms, ranging from setting up dedicated platforms that do not factor such variables and relationships, or deliberately constrain search and recommendation results by adhering to certain fairness criteria, or provide transparency and algorithmic control to users to govern their selection criteria. In this article, we begin with conducting an algorithmic audit of recommendations generated on two popular online platforms in India about whether they echo and reinforce these norms. We find that indeed they do, likely by not attempting to correct for them, and we also propose and evaluate alternative algorithms that can yield fairness aware recommendations while simultaneously satisfying user preferences. Our study is an attempt to highlight that designers and managers of such platforms can be transparent about ideological choices they make of norms they support or counter, so that platforms built and managed by them can offer agency to users to receive recommendations based on their preferences. At the same time, we acknowledge that being able to represent in data various societal norms and individual preferences in all their complexity may always remain incomplete, and consequently any attempts to counter biases will be limited in their impact as well. However, since datafication continues to encompass more and more domains, and digital platforms continue to algorithmize and intermediate many personal and group processes, we feel our study is relevant to highlight a new application domain that has significant societal implications on gender and related equalities but has not been studied extensively so far.