Mining Linguistic Styles in Bilateral Matching: A Contrastive Learning Approach to Reciprocal Recommendation

Reciprocal recommendation systems are crucial for online dating platforms to provide quality matches and reduce choice overload. However, the design of reciprocal recommendation systems grapples with the challenges of estimating interpersonal compatibility and predicting the likelihood that two prospective partners will accept each other. Furthermore, despite the crucial role of users’ linguistic styles in determining user match decision-making, the contemporary design of such recommendation systems has not yet effectively incorporated this information. To bridge these gaps, we develop an end-to-end personalized Linguistic Style Matching-based Reciprocal Recommendation System (LS-RRS). We propose cross-user and within-user contrastive learning strategies combined with random masking to extract users’ linguistic styles and further integrate visual and textual information using an efficient convolution block. LS-RRS further models the matching probability using a conditional probability function and introduces a preference inflation factor on the receiver side to account for the asymmetric roles of the bilateral sides. The proposed model addresses the challenge of incorporating users’ linguistic styles into reciprocal recommendation and details the modeling of the two-stage matching process. Extensive experiments show that LS-RRS outperforms state-of-the-art models in recommendation performance, with a 29.35% increase in NDCG@10 when incorporating linguistic styles. Our follow-up analyses further validate the importance and effectiveness of the linguistic style extraction design through word-level and sentence-level visualizations, as well as qualitative case studies. This research contributes to the literature on reciprocal recommendation and offers a viable solution for alleviating user choice overload on online dating platforms.