Reciprocal Recommender Systems

Updated 24 March 2026

Reciprocal recommender systems are specialized algorithms for two-sided matching that require mutual acceptance, differing from traditional one-sided recommendations.
They integrate bilateral scoring functions with methods like collaborative filtering, neural architectures, and causal inference to accurately model mutual preferences.
Applications span online dating, job matching, and peer group formation, addressing challenges in fairness, computational efficiency, and balanced exposure.

A reciprocal recommender system (RRS) is a class of algorithms that address two-sided matching problems where successful outcomes require mutual agreement by both parties. Unlike traditional recommenders that recommend items to users, RRSs recommend users to each other—examples include online dating, friendship recommendation, peer matching, and two-sided job marketplaces. The operational requirement of bilateral (reciprocal) acceptance introduces new modeling, fairness, and computational challenges, motivating a diverse literature spanning collaborative filtering, matching theory, neural architectures, fairness-aware allocation, and causal inference frameworks.

1. Formal Problem Definitions and Core Concepts

The key distinction of RRSs is the requirement that both user $x$ and candidate $y$ (where $x, y$ are elements of two possibly distinct sets, e.g., $X, Y$ ) must mutually accept or express interest for a match to be realized. This changes the objective relative to classic user–item recommenders in several ways:

Reciprocity condition: A recommendation $(x, y)$ is only successful if $x$ is recommended $y$ and $y$ is recommended $x$ (typically appearing in each other's top- $K$ recommendation lists) (Prabhakar et al., 2017).
Preference estimation: Each side's preference must be modeled, resulting in two scoring functions $p_{x,y}$ and $p_{y,x}$ , which are then fused using symmetric aggregation (e.g., harmonic mean, geometric mean) or more sophisticated approaches (Tomita et al., 2022, Zhao et al., 2013, Xia et al., 2015).
Matching as two-sided intervention: Reciprocal recommendation is naturally cast as a function of bilateral treatments or interventions. In the potential-outcome framework, recommendations to both $x$ and $y$ can be viewed as independent interventions whose interactions jointly influence the probability of matching (Yang et al., 2024).
Objective functions: Aggregate objectives include maximizing the total expected number of matches (social welfare), optimizing fairness with respect to matched opportunities (e.g., envy-freeness), and balancing between efficiency and equity (Tomita et al., 20 Jan 2026, Tomita et al., 2024).

This two-sided nature introduces constraints and dependencies absent from traditional recommender paradigms, requiring new models for both scoring and allocation.

2. Algorithmic Advances in Reciprocal Recommendation

Algorithmic methods in RRSs can be organized along several axes: collaborative filtering models, content-based approaches, matching-theoretic and equilibrium methods, neural architectures, and causality-oriented techniques.

2.1 Neighborhood and Graph-based Collaborative Filtering

Early RRSs for online dating and peer matching often extended user-based CF by defining directional similarities for both "taste" (who a user prefers) and "attractiveness" (who prefers the user), combining them with symmetric fusion (typically harmonic mean). Jaccard-based interest and attractiveness similarities induce projection graphs capturing these dual dimensions (Xia et al., 2015, Zhao et al., 2013). The recommendation score $R(x, y)$ is systematically constructed from bilateral similarities and normalized to correct for user activity disparities.

2.2 Content-based and Attribute Matching

Content-based RRSs rely on user profiles (demographics, preferences, explicit interests). Multi-attribute distance functions with optional user-defined priority reweighting—followed by symmetry-enforcing procedures such as harmonic mean fusion and intersection of top- $K$ lists—deliver matchable candidate sets (Prabhakar et al., 2017, Voigt et al., 2021).

2.3 Matching Theory: Stable and Transferable-Utility Equilibria

Several modern RRS architectures draw on economic matching theory, instantiating the system as an optimal allocation under capacity constraints and mutual preferences.

Stable matching (NTU): Classical Gale–Shapley deferred acceptance, used in settings where ordinal preferences are available, yields stable matchings immune to blocking pairs (Tomita et al., 2022).
Transferable utility (TU): When mutual "utilities" can be aggregated, the equilibrium matching problem admits a closed-form, entropy-regularized solution (e.g., Choo–Siow, Shapley–Shubik), often implemented by scalable Iterative Proportional Fitting or Sinkhorn-style updates (Tomita et al., 2023, Nakada et al., 2024, Tomita et al., 2022). These models achieve fairness via congestion penalties: as one user is matched frequently, their marginal match rate is automatically reduced.

2.4 Neural and Deep Sequential Architectures

RRSs increasingly leverage neural approaches:

Random CNN and RL-based feature selection: The RRCN approach combines attribute-based user embeddings with random convolutional filters over non-adjacent features, further enhanced by RL-guided attribute selection for salience (Luo et al., 2020).
Sequential and co-attention models: The ReSeq model encodes both users' dynamic interaction sequences, performing co-attention over time to capture evolving reciprocal preferences and employing self-distillation for efficient inference (Zheng et al., 2023).
Image-based models: Architectures based exclusively on user profile images demonstrate that in visually-dominated domains, end-to-end models can produce high-precision reciprocal matches with minimal reliance on auxiliary side information (Neve et al., 2021).

2.5 Causal and Counterfactual Models

Causal methods explicitly recognize the impact of historical exposure bias and unequal treatment, introducing inverse propensity scoring and counterfactual risk minimization via self-normalized objectives (e.g., SNIPS) to debias observed data and deliver more equitable recommendations (Kawamura et al., 3 Aug 2025, Yang et al., 2024). Multi-model, treatment-specific architectures further estimate distinct potential outcomes for every possible intervention (recommendation) pattern, enabling direct optimization of match coverage and bilateral stability metrics (Yang et al., 2024).

Fairness in RRSs is structurally more complex than in one-sided recommenders. Two-sided platforms need to ensure that reciprocal matches do not concentrate excessively on a small subset of users (typically those with high "attractiveness" or popularity scores), a problem often called opportunity inequality or envy.

Social welfare (SW): Maximization of expected matches (total utility) via linear or entropic assignments; can lead to highly skewed allocation (Tomita et al., 20 Jan 2026, Tomita et al., 2024).
Nash social welfare (NSW): Alternating maximization of the geometric mean (product) of users' utilities guarantees (approximate) envy-freeness under mild conditions, striking a fairness–efficiency trade-off (Tomita et al., 20 Jan 2026, Tomita et al., 2024).
Opportunity as a divisible good: Inspired by fair division, each user's "chance to be recommended" is formalized as a divisible resource. Algorithmic policies are then selected to minimize pairwise envy and maintain parity in opportunities for exposure (Tomita et al., 2024).
Sinkhorn-based algorithms and allocation scaling: Scaling to massive platforms is feasible via entropic-regularized optimal transport solvers (mini-batch, factor-model enhanced Sinkhorn iterations), allowing reciprocal matching among millions of users (Nakada et al., 2024).

Empirical results demonstrate that fairness-aware allocation (using NSW or $\alpha$ -SW models) often achieves near-optimal total matches with drastically reduced opportunity envy, as measured by both theoretical envy counts and Gini indices (Tomita et al., 20 Jan 2026, Tomita et al., 2024).

4. Evaluation Metrics and Methodological Innovations

Evaluating RRSs requires metrics sensitive to bilateral success and coverage, as well as traditional information retrieval criteria. Recent literature introduces new evaluation protocols:

Coverage-adjusted and stability-adjusted precision/recall: These metrics subtract double-counted mutual recommendations and focus on the fraction of true matching pairs recovered with bilateral endorsements (Yang et al., 2024).
Balanced NDCG (RNDCG): Weighted averages of NDCG by population size prevent one side from dominating aggregate performance scores (Yang et al., 2024).
Envy measurement and Gini indices: Quantify unequal match or recommendation allocation across user populations (Tomita et al., 20 Jan 2026, Tomita et al., 2023).
Reciprocity-specific discounted gain (DCG/NDCG): Variants that quantify reciprocal rank alignment and mutual relevance (Prabhakar et al., 2017).
Long-tail coverage: Fraction of low-exposure users appearing in top- $k$ lists, crucial for equity assessment (Kawamura et al., 3 Aug 2025).

Independently assessing both sides and reporting bilateral or aggregate metrics has become standard in rigorous RRS evaluations.

5. Applications and Practical Systems

RRSs are central to a range of platforms:

Online dating: The canonical RRS application, with extensive deployment of CF and matching-theoretic RRSs (Zhao et al., 2013, Xia et al., 2015, Neve et al., 2021).
Job/candidate matching: Both recruiter and candidate preferences are accommodated in hybrid, TU-based, and fairness-aware RRSs (Ruijt et al., 2021, Tomita et al., 2023, Tomita et al., 2022, Nakada et al., 2024).
MOOC peer and study group formation: Attribute-matching models with symmetric re-ranking and reciprocity enforcement optimize for mutually compatible learner groups (Prabhakar et al., 2017).
Expert/mentee matching, roommate assignment, collaborative filtering for mutual compatibility: Emerging applications reflecting the growing relevance of two-sided reciprocation requirements.

Privacy-preserving RRS variants using Bloom filters, local differential privacy, and decentralized architectures have also been implemented for sensitive domains (e.g., HR recruitment) (Voigt et al., 2021).

6. Theoretical Foundations and Limitations

Theoretical analysis has established foundational results on the efficiency, optimality, and regret of online reciprocal matching algorithms:

Sequential RRSs: Under minimal feedback and clusterable preference assumptions, the SMILE algorithm achieves near-optimal match discovery rates, closely approaching clairvoyant upper bounds (Vitale et al., 2018).
Worst-case analysis: In the absence of structure, random selection matches are minimax-optimal, demonstrating fundamental limits on purely exploratory matching (Vitale et al., 2018).

Open issues include scalability of full joint matchers, adaptation to non-stationary markets, direct modeling of time-dynamics and textual content, and robust online learning of preference distributions and exposure propensities (Zheng et al., 2023, Yang et al., 2024).

7. Explanation, User Trust, and Behavioral Effects

Explanations in reciprocal environments serve dual purposes: increasing user acceptance, and managing the risk/cost of initiating contact.

Reciprocal explanations: Providing both the recipient's and the candidate's reasons (top- $k$ attribute correlations or content signals) can increase acceptance and trust, especially in high-cost or emotionally-charged applications (Kleinerman et al., 2018).
Personalization of explanation strategy: Effectiveness of reciprocal vs. one-sided explanations depends on user heterogeneity and the perceived cost of engagement, requiring system designers to calibrate explanation delivery accordingly (Kleinerman et al., 2018).

This highlights the behavioral nuance of RRS design, with implications for both engagement metrics and user satisfaction.

The study and deployment of reciprocal recommender systems is characterized by deep theoretical, algorithmic, and practical innovation, driven by the inherent complexity of mutual acceptance and fairness constraints. Across domains and architectures, the field continues to advance rigorous, scalable methods for equitable and efficient two-sided matching.