Beyond Match Maximization and Fairness: Retention-Optimized Two-Sided Matching
Pith reviewed 2026-05-21 12:20 UTC · model grok-4.3
The pith
A dynamic learning-to-rank method for two-sided platforms maximizes overall user retention by learning personalized curves and allocating matches where they most reduce churn.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
By modeling each user's retention probability as a function of the recommendations they receive and jointly optimizing the sum of retention effects across both sides of every potential match, MRet reallocates limited matching opportunities to produce measurably higher long-term user retention on two-sided platforms.
What carries the argument
Matching for Retention (MRet), a dynamic learning-to-rank algorithm that learns personalized retention curves from user profiles and interaction history and scores candidate matches by their combined retention impact on both the recommending user and the recommended user.
If this is right
- Matching opportunities are assigned to pairings that produce the largest net increase in total retention rather than the largest number of matches.
- Users whose retention is most sensitive to recommendations receive priority in the allocation of scarce slots.
- The platform no longer relies on fairness axioms to indirectly improve retention; retention is optimized directly.
- Imbalance where a few users receive many matches while most receive none is reduced because retention curves penalize over-exposure.
Where Pith is reading between the lines
- The same retention-curve approach could be tested in other two-sided markets such as job recruiting to reduce applicant and employer churn.
- If retention curves prove stable, platforms might drop separate fairness constraints and treat retention as the sole primary objective.
- Extending the model to include long-term network effects, such as how one user's retention influences their friends' participation, would be a natural next measurement.
Load-bearing premise
Personalized retention curves estimated from each user's profile and past interactions give reliable forecasts of how a specific recommendation will change that user's probability of staying on the platform.
What would settle it
A live A/B test on a two-sided platform that measures the change in user retention rate (e.g., fraction of users still active after 30 days) when MRet replaces a match-maximization or fairness baseline.
read the original abstract
On two-sided matching platforms such as online dating and recruiting, recommendation algorithms often aim to maximize the total number of matches. However, this objective creates an imbalance, where some users receive far too many matches while many others receive very few and eventually abandon the platform. Retaining users is crucial for many platforms, such as those that depend heavily on subscriptions. Some may use fairness objectives to solve the problem of match maximization. However, fairness in itself is not the ultimate objective for many platforms, as users do not suddenly reward the platform simply because exposure is equalized. In practice, where user retention is often the ultimate goal, casually relying on fairness will leave the optimization of retention up to luck. In this work, instead of maximizing matches or axiomatically defining fairness, we formally define the new problem setting of maximizing user retention in two-sided matching platforms. To this end, we introduce a dynamic learning-to-rank (LTR) algorithm called Matching for Retention (MRet). Unlike conventional algorithms for two-sided matching, our approach models user retention by learning personalized retention curves from each user's profile and interaction history. Based on these curves, MRet dynamically adapts recommendations by jointly considering the retention gains of both the user receiving recommendations and those who are being recommended, so that limited matching opportunities can be allocated where they most improve overall retention. Naturally but importantly, empirical evaluations on synthetic and real-world datasets from a major online dating platform show that MRet achieves higher user retention, since conventional methods optimize matches or fairness rather than retention.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript introduces the problem of maximizing user retention on two-sided matching platforms (e.g., online dating) rather than maximizing match count or enforcing fairness. It proposes Matching for Retention (MRet), a dynamic learning-to-rank algorithm that learns personalized retention curves from each user's profile and interaction history. These curves are then used to jointly optimize recommendations by allocating limited matching opportunities to improve overall retention for both the recipient and the recommended users. The authors report that empirical evaluations on synthetic data and real-world datasets from a major online dating platform show MRet achieving higher retention than conventional match-maximization or fairness-based methods.
Significance. If the central empirical claim holds after addressing modeling concerns, the work would be significant for recommendation systems on two-sided platforms where long-term retention drives platform viability (especially subscription models). Shifting from short-term match counts or axiomatic fairness to an explicit retention objective is a practical and timely contribution. The dynamic joint consideration of retention effects on both sides of each match extends standard LTR approaches in a coherent way, and the inclusion of real-world platform data strengthens external relevance. The approach could influence deployed systems if the retention curves prove to be reliable predictors rather than correlational artifacts.
major comments (2)
- [§3] §3 (Retention curve learning): The description of how personalized retention curves are estimated from observational profiles and interaction history does not specify any correction for selection bias or confounding (e.g., highly engaged users both receiving more matches and exhibiting longer retention). This is load-bearing for the central claim, because without isolating the marginal causal effect of a given recommendation on future retention probability, the reported gains on real data may simply reproduce the same engagement correlations already exploited by conventional methods rather than demonstrating true retention optimization.
- [§5] §5 (Empirical evaluation): No details are supplied on the retention curve model architecture, training procedure, baseline implementations, exact definition of the retention metric, or statistical tests used to compare MRet against match-maximization and fairness baselines. This omission is load-bearing because the abstract's assertion of superior retention on both synthetic and real datasets cannot be assessed for robustness, potential post-hoc fitting of curves, or whether the improvements are statistically significant and generalizable.
minor comments (2)
- The abstract would be clearer if it briefly stated the precise operational definition of retention (e.g., days until churn, subscription renewal probability) used in the experiments.
- Notation for the joint retention-gain objective could be introduced earlier to help readers follow the dynamic adaptation step without back-referencing.
Simulated Author's Rebuttal
We thank the referee for the constructive and detailed feedback. The comments highlight important aspects of clarity and rigor that we will address in the revision. We respond to each major comment below.
read point-by-point responses
-
Referee: [§3] §3 (Retention curve learning): The description of how personalized retention curves are estimated from observational profiles and interaction history does not specify any correction for selection bias or confounding (e.g., highly engaged users both receiving more matches and exhibiting longer retention). This is load-bearing for the central claim, because without isolating the marginal causal effect of a given recommendation on future retention probability, the reported gains on real data may simply reproduce the same engagement correlations already exploited by conventional methods rather than demonstrating true retention optimization.
Authors: We acknowledge the concern about selection bias and confounding in observational retention curve estimation. The current §3 describes learning from profile features and interaction history but does not explicitly detail causal adjustments such as propensity weighting. In the revised manuscript we will expand this section to discuss potential confounders, describe mitigation via feature selection and regularization, and clarify that the approach is predictive rather than strictly causal. We will also highlight that synthetic experiments, where the data-generating process is fully controlled, continue to show MRet outperforming baselines, indicating gains beyond simple engagement correlations. revision: partial
-
Referee: [§5] §5 (Empirical evaluation): No details are supplied on the retention curve model architecture, training procedure, baseline implementations, exact definition of the retention metric, or statistical tests used to compare MRet against match-maximization and fairness baselines. This omission is load-bearing because the abstract's assertion of superior retention on both synthetic and real datasets cannot be assessed for robustness, potential post-hoc fitting of curves, or whether the improvements are statistically significant and generalizable.
Authors: We agree that these implementation and evaluation details are essential for assessing robustness and reproducibility. In the revised manuscript we will expand §5 (and integrate relevant appendix material) to specify the retention curve model architecture, training procedure and hyperparameters, exact baseline implementations, the precise retention metric definition, and the statistical tests (including significance levels) used for comparisons. These additions will allow readers to evaluate the reported improvements on both synthetic and real-world data. revision: yes
Circularity Check
No significant circularity; modeling and optimization are independent of inputs
full rationale
The paper defines retention maximization as a new objective, learns personalized retention curves from observed user profiles and histories as a standard supervised modeling step, then uses those curves inside a dynamic LTR optimizer. Empirical gains are reported on held-out synthetic and real-world datasets. No equation reduces the claimed retention gain to a fitted parameter by construction, no self-citation chain carries the central claim, and the derivation does not rename a known result or smuggle an ansatz. The approach is a conventional ML pipeline with external validation benchmarks.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.