When Netflix recommends a movie or Spotify suggests a song, they operate under an assumption of infinite supply. If a million users all decide to stream the same track at the same time, the platform handles it effortlessly. The consumption of digital media is non-rivalrous—one person’s consumption does not diminish another's.
However, when recommendation systems are deployed in the physical world or in matching markets, this assumption of infinite availability breaks down. Consider recommending jobs (Bied et al.), dating partners (Ribeiro et al.), physical services like restaurants (Cao et al.), or one-on-one online tutors (Chen et al.). These resources are strictly finite and often rivalrous. A job can only be filled by one applicant; a restaurant has a fixed number of seats; a human provider has limited time and attention.
When traditional recommendation algorithms—which optimize purely for individual user-item affinity—are applied to finite resources, they suffer from an infinite supply bias. This bias creates severe systemic failures, transforming recommendation engines from helpful guides into sources of market congestion and user frustration.
The Mechanics of Recommendation-Induced Congestion
Why do standard recommendation systems fail when resources are limited? The core issue lies in independent optimization. Standard models calculate a preference score for every user-item pair and recommend the top-scoring items to each user, completely blind to what is being recommended to everyone else.
In any market, user preferences tend to concentrate around a small subset of highly desirable options. When an unconstrained algorithm processes these concentrated preferences, it directs a massive wave of users toward the exact same few items.
Unconstrained Recommendation:
User A ──┐
User B ──┼─> [ Highly Popular Item ] ──> Overwhelmed / Congestion
User C ──┘
Capacity-Constrained Recommendation:
User A ──────> [ Highly Popular Item (At Capacity) ]
User B ──────> [ Compatible Alternative Item A ]
User C ──────> [ Compatible Alternative Item B ]
This creates several distinct failure modes across different domains:
- The Celebrity Bottleneck: In online dating, a handful of highly attractive profiles receive an overwhelming volume of messages, while the vast majority of users receive no engagement at all. As Ribeiro et al. observe, this causes high user churn and prevents the market from clearing efficiently.
- Booking Collisions: In one-on-one service marketplaces, such as assigning teachers to students, independent recommendations cause multiple users to attempt to book the same provider simultaneously, leading to scheduling conflicts and system instability (Chen et al.).
- Over-tourism and Infrastructure Strain: In itinerary planning, recommending the same scenic point of interest to every tourist causes physical overcrowding, destroying the user experience and straining local infrastructure (Gavalas et al.).
As Jordan argues, systems focused solely on isolated predictions inevitably fail in scenarios involving scarcity. Without a mechanism to coordinate recommendations, individual optimizations aggregate into collective inefficiencies. Phade et al. echo this sentiment, noting that directing users toward exhausted inventory results in systemic failure.
Shifting from Prediction to Coordination
To resolve congestion, the recommendation problem must be reframed. Instead of performing independent row-by-row predictions on a user-item matrix, the system must perform joint resource coordination. The goal shifts from finding the absolute highest-affinity item for an individual to optimizing a global allocation strategy that respects capacity constraints.
Researchers have introduced several algorithmic frameworks to achieve this coordination:
1. Two-Sided Matching and Stable Matching
In reciprocal markets like employment or dating, a successful outcome requires mutual interest and respects the finite capacity of both sides. Su et al. introduce frameworks that jointly optimize rankings for all participants simultaneously, maximizing "social welfare"—the total number of successful matches across the entire platform—rather than isolated click-through rates.
To make this computationally viable at scale, Tomita et al. incorporate matching theory directly into the recommendation engine, treating user attention as a finite resource and redistributing suggestions toward compatible but less overwhelmed participants. When preference concentration is high, Nakada et al. demonstrate that stable matching policies maintain a significantly higher total match count than standard methods, preventing platform utility from collapsing when popular nodes become bottlenecks.
2. Flow Control and Optimal Transport
Another approach is to treat recommendation visibility as a fluid dynamic problem where demand must be routed smoothly across available supply. Ribeiro et al. implement explicit flow control by capping the number of times a specific high-demand entity can be suggested across the platform.
Taking this further, Mashayekhi et al. integrate Optimal Transport theory directly into the model's training process. By treating the distribution of job seekers and the distribution of available vacancies as two probability measures, their framework minimizes the cost of shifting recommendations away from congested items. This ensures an even distribution of attention across vacancies without sacrificing the underlying quality of the matches.
3. Diversity Guardrails and Structural Constraints
In highly constrained one-on-one environments, systems can use structural constraints to prevent collisions before they happen. Chen et al. implement "diversity guardrails" that explicitly minimize the overlap between the recommendation lists provided to different users. Here, diversity is not just a tool for user exploration, but a critical operational constraint to ensure system stability.
On a macro level, Larson et al. propose "hypotargeting," a framework that intentionally caps the total number of unique recommendation lists a system can generate for an entire population, providing a structural bound on recommendation variance to ease algorithmic oversight.
Balancing Utility, Fairness, and Sustainability
When a recommendation system enforces capacity constraints, it introduces a fundamental trade-off: some users cannot be shown their absolute top preference because that item is already at capacity. Managing this trade-off requires new definitions of fairness and utility.
Individual Fairness Over Time
If a user is repeatedly denied access to a popular service due to capacity limits, the system becomes unfair. To combat this, Cao et al. introduce a strategy that tracks which users have been "shortchanged" by capacity limits in previous rounds. The system then dynamically prioritizes these users in subsequent rounds, balancing long-term individual fairness with real-time capacity constraints.
Envy and Inferiority
In high-stakes domains like job recruitment or university admissions, recommending a highly competitive, limited resource to a user who has a low probability of securing it is actively harmful. Li et al. develop a framework that quantifies inferiority (the frustration of being recommended items out of reach) and envy (the unfairness felt when others receive better options). By optimizing for these metrics alongside traditional utility, the system avoids setting users up for inevitable rejection.
Sustainable Resource Management
In physical environments, capacity-aware recommendation shifts the system's objective from pure preference satisfaction to sustainable resource management. As highlighted by Gavalas et al., incorporating physical limitations allows platforms to balance individual desires with the preservation of local infrastructure. For example, Halder et al. demonstrate that factoring the ratio of visiting time to queuing time into an itinerary recommendation engine prevents artificial bottlenecks at popular destinations, improving the actual realized utility for the user.
Ultimately, moving beyond the unconstrained resource assumption requires moving beyond pure similarity matching (Basu et al.). By treating attention and physical items as finite, interconnected resources, capacity-constrained recommendation systems transform chaotic, congested environments into balanced, sustainable ecosystems.
Key Foundational Work
Online dating recommendations: matching markets and learning preferences (2014) — contributed 3 concept elements to this direction.
Technical Paper Recommendation: A Study in Combining Multiple Information Sources (2011) — contributed 1 concept elements to this direction.