A Reciprocal Embedding Framework For Modelling Mutual Preferences

R. Ramanathan, Nicolas K. Shinada, Michinobu Shimatani, Yuhei Yamaguchi, Junichi Tanaka, Yuta Iizuka, Sucheendra K. PalaniappanAAAI 2021 · DOI

Abstract

Reciprocal recommendation is a critical task in online dating where a successful outcome requires mutual interest from both parties. This paper introduces a flexible embedding framework that decouples bidirectional preferences into separate unidirectional models, which are then fused and refined through a two-stage production pipeline. By addressing challenges such as preference drift and extreme data sparsity, the proposed system achieved a 16.9% increase in matches and a 60% improvement in conversion rates on Tapple, one of Japan's largest dating platforms.

Introduction

Most modern recommendation systems, such as those used by Netflix or Amazon, are unidirectional; the system optimizes for a user's interest in an item without requiring the item to "reciprocate." In contrast, online dating platforms operate on a reciprocal basis: a recommendation is only successful if it results in a mutual "like" or match. This introduces unique challenges, including preference drift (where user tastes change based on recent interactions), the need to model bidirectional interest, and extreme data sparsity (often as low as 0.00001 density).

Methodology

The proposed framework treats reciprocal recommendation as a fusion of two distinct unidirectional problems: Men-to-Women and Women-to-Men preferences.

Unidirectional Modeling

The system employs three primary algorithms to learn latent user preferences from interaction subsets:

  • Matrix Factorization (MF): Decomposes interaction matrices into low-rank user and item embeddings.
  • Bayesian Personalized Ranking (BPR): Optimizes for the relative order of user preferences.
  • Neural Networks: A feed-forward architecture that incorporates side information (e.g., age, interests) to enhance embedding quality.

The Reciprocal Pipeline

To handle the scale of millions of users, the system uses a two-stage architecture. First, Candidate Generation uses the learned embeddings to retrieve a few hundred high-potential candidates. Second, Online Re-ranking filters these candidates in real-time based on dynamic factors like recent activity, location, and diversity to ensure recommendation freshness.

Score Aggregation

To produce a final mutual preference score, the framework experiments with various fusion functions—including arithmetic mean, harmonic mean, and product—to combine the scores from the two unidirectional models.

Data Analysis and Insights

The study utilized data from Tapple, a major Japanese dating app with over 5 million users. Analysis revealed that the interaction data follows a classic "long-tail" distribution, where a small minority of highly active users accounts for the majority of likes and matches.

Figure 1: Log-log distribution of user interactions (Likes, Dislikes, and Matches) showing the characteristic long-tail behavior of the dataset.

Furthermore, the researchers found that "Location is King." Most interactions occur within the same prefecture, suggesting that geographic constraints are a primary driver of user behavior. This is reflected in the interaction density across different regions.

Figure 6: Heatmap of interactions across Japan's 47 prefectures, illustrating the high concentration of matches within local geographic boundaries.

Key Results

The Reciprocal Recommender System (RRS) was benchmarked against a baseline model (Match-RS) that only learns from successful matches.

  • Offline Evaluation: The RRS framework demonstrated a 16.9% increase in matches and a 26.74% increase in likes over the baseline.
  • Online Deployment: The addition of the online re-ranking stage led to a 60% boost in conversion rates (the rate at which likes turn into mutual matches), highlighting the importance of balancing model relevance with real-time user activity.

Conclusion and Future Work

By decoupling reciprocal preferences and utilizing a robust two-stage pipeline, this framework significantly improves matching efficiency in sparse, large-scale environments. Future research will explore joint embedding spaces to encode mutual interest directly into single vector representations and the use of attention mechanisms to better navigate sparse interaction networks.

Summary updated

BibTeX
@inproceedings{ramanathan2021reciprocal,
  author = {R. Ramanathan and Nicolas K. Shinada and Michinobu Shimatani and Yuhei Yamaguchi and Junichi Tanaka and Yuta Iizuka and Sucheendra K. Palaniappan},
  title = {{A Reciprocal Embedding Framework For Modelling Mutual Preferences}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i17.17807},
}
Figure