Most recommendation systems are built on an implicit assumption of infinite supply. In the digital world of Netflix or Spotify, recommending the same blockbuster movie to millions of users simultaneously costs nothing; the "item" is a bitstream that can be replicated endlessly.
However, as Michael Jordan argues in Dr. AI or: How I Learned to Stop Worrying and Love Economics, this "prediction-only" mindset fails when applied to the physical world. In markets for jobs, dating, ride-sharing, or travel, items are rivalrous: a single job can only be filled by one person, a restaurant has finite seats, and a dating prospect has limited time to reply to messages. When recommendation engines ignore these constraints, they create systemic congestion that degrades the experience for everyone.
The Celebrity Bottleneck and Systemic Congestion
When a standard recommendation algorithm identifies a "top-tier" item, it naturally steers all users toward it. In a matching market—like online dating—this creates what Bruno Ribeiro and colleagues describe as a "celebrity bottleneck" in Online dating recommendations: matching markets and learning preferences.
A small subset of popular users becomes overwhelmed by more messages than they can possibly process, leading to communication fatigue. Meanwhile, the vast majority of users receive no engagement at all. This imbalance leads to market inefficiency and high churn. As Yi Su and Thorsten Joachims point out in Optimizing Rankings for Recommendation in Matching Markets, the goal should shift from maximizing individual click probability to maximizing social welfare—the total number of successful matches across the entire platform.
From Prediction to Allocation
To solve this, researchers are moving away from simple preference matching toward resource-aware allocation. This involves treating the recommendation problem as a constrained optimization task.
Joint Optimization vs. Post-Processing
A naive way to handle constraints is post-processing: generate a standard list of recommendations and then truncate it if an item is "sold out." However, Arindam Banerjee and Konstantina Christakopoulou demonstrate in Recommendation under Capacity Constraints that this is highly inefficient.
Instead, they propose integrating capacity limits directly into the objective function. By adding a penalty for exceeding availability into the training process, the model learns to steer users toward under-utilized items that still align with their tastes. This "joint optimization" ensures that the system doesn't just predict what a user likes, but what they can actually attain.
Managing Rivalrous Goods
In the context of employment, Guillaume Bied and the team behind Toward Job Recommendation for All highlight that jobs are strictly rivalrous. Recommending the same position to thousands of qualified candidates is a recipe for mass rejection. Their system, MUSE, accounts for resource exhaustion to ensure that recommendations are distributed across the available job pool, preventing artificial competition for a handful of roles.
Mechanisms for Stability and Fairness
If multiple users want the same limited resource, how does the system decide who gets it? Recent work draws on economic theory to find "stable" solutions.
- Pricing as a Signal: Soham Rajesh Phade and Kannan Ramchandran introduce ILAP in Interactive Learning with Pricing for Optimal and Stable Allocations in Markets. They use Walrasian pricing theory to manage demand. By "pricing" items based on their popularity and remaining capacity, the system can naturally distribute users across the market, minimizing dissatisfaction and ensuring no single resource is over-promised.
- Optimal Transport: In Learning to Recommend via Inverse Optimal Matching, Ruilin Li and Hongyuan Zha use the Wasserstein distance to model matching markets. This allows the system to learn the underlying "rules" of a market—such as why certain matches happen and others don't—while respecting the hard physical reality that every "match" consumes a unit of supply.
- Reducing Envy: Nan Li and Tijl De Bie argue in FEIR that fairness in these systems requires minimizing "inferiority." A recommendation is only useful if the user has a realistic chance of obtaining the item. Recommending a high-demand resource to a user who is unlikely to "win" it is not just a waste of a recommendation slot; it creates systemic envy and frustration.
Applications in Infrastructure and Logistics
The impact of these resource-aware systems extends beyond digital marketplaces into physical infrastructure:
- Traffic Management: Yan Leng and Larry Rudolph show in Managing travel demand that by recommending "secondary" leisure locations to drivers during peak times, a system can reduce traffic delays by over 50% while only slightly nudging users away from their first choice.
- Sustainable Tourism: Damianos Gavalas notes in Tourist trip planning that integrating site capacity into itinerary recommendations prevents overcrowding at popular landmarks, preserving both the local infrastructure and the visitor experience.
By shifting the focus from "what does this user want?" to "how can we best allocate what we have?", these researchers are building a new generation of recommendation systems that are grounded in the economic and physical realities of the real world.