Beyond Relevance: A Demand Balancer Model for Rental Platforms with Single-Unit Inventory (WSDM 2025)
Source: https://dl.acm.org/doi/abs/10.1145/3701551.3706126
About this paper
This WSDM 2025 paper tackles a distinctive problem in rental marketplaces: unlike e-commerce with abundant stock, rental platforms (e.g., vacation rentals, short-term accommodation) operate with single-unit inventory. Each listing can be booked by only one user for a given time window, so optimizing purely for user relevance causes demand to concentrate on a small set of popular listings, while the long tail of supply stays vacant.
The proposed demand balancer model goes "beyond relevance" by jointly optimizing:
- User-side relevance — matching queries/intent to listings;
- Platform-side demand distribution — steering traffic across the inventory to improve booking coverage and supply utilization.
- The cascade of retrieval → ranking → generation remains the mainstream architecture, with agentic paradigms making retrieval strategy itself learnable.
- High-quality logs and instruction data remain critical; synthetic data risks distribution shift.
- Offline metrics (nDCG, MRR, Recall@k) increasingly diverge from online satisfaction; LLM-as-judge needs human cross-validation.
- Latency, cost, explainability, and safety are hard constraints for industrial deployment.
- An interpretable ensemble of graph and language models
- Applying Deep Learning to Ads Conversion Prediction in Last Mile Delivery
- Automated Query-Product Relevance Labeling using Large Language Models
- Behavior Modeling Space Reconstruction for E-Commerce Search (Jan 2025)
- Behavior-driven query similarity prediction based on pre-trained language models
- Better to Ask in English: Cross-Lingual Evaluation of Large Language Models
Why single-unit inventory matters
In classic recommendation, exposing an item does not consume it. In rentals, a successful recommendation *removes* the item from the pool, creating dynamic, time-sensitive inventory constraints. This makes the ranking problem closer to an allocation/balancing task than a pure relevance-ranking task.
Context within search & recommendation
The post situates this work in the broader IR/RecSys landscape:
Engineering deployment checklist
| Check | Question | Suggestion | |-------|----------|------------| | Data | PII in training/index? Version control? | Sharded indexes, anonymization, rollback-able embedding versions | | Latency | p99 budget? Retrieval steps? | Cascade + early stopping, hot-query caching, async reranking | | Quality | Do offline gains convert to CTR/satisfaction? | Interleaving experiments, human audits, citation checks | | Safety | Poisoning/bias via open retrieval? | Source whitelists, adversarial detection, output filtering | | Cost | Per-query token/GPU cost? | Small-model routing, distillation, hybrid sparse+dense |
Related entries
Takeaways
1. For researchers: focus on how platform-level objectives (demand balancing) can be integrated with user-level relevance without hurting experience; check whether papers report significance and compute cost. 2. For engineers: extract pluggable modules (encoders, rerankers, balancers) and assess integration cost with existing stacks. 3. For product managers: track user-perceivable benefits (latency, trustworthiness, consistency), not just offline nDCG.
> Note: Specific quantitative results should be verified against the original PDF via the ACM link above; this report is based on the title, abstract metadata, and public information.