Understanding the mutual preferences between potential dating partners is core to the success of modern web-scale personalized recommendation systems that power online dating platforms. In contrast to classical user-item recommendation systems which model the unidirectional preferences of users to items, understanding the bidirectional preferences between people in a reciprocal recommendation system is more complex and challenging given the dynamic nature of interactions. In this paper, we describe a reciprocal recommendation system we built for one of the leading online dating applications in Japan. We also discuss the lessons learnt from designing, developing and deploying the reciprocal recommendation system in production.

Online dating platforms have changed the paradigm of how people seek potential relationships. In this context, reciprocal recommendation systems consider the mutual ’match’ potential between users, i.e users who are likely to interact and potentially ’like’ each other. We present our experiences on how we devised algorithms to overcome data specific nuances, built and deployed the system from scratch in a relatively short time-span for one of the prominent dating applications in Japan.