TL;DR
This paper introduces a novel framework, BLUE, to infer private valuations in sequential bargaining without relying on strong equilibrium assumptions. By leveraging the fact that bargainers avoid strictly dominated strategies, BLUE derives feasible valuation intervals to guide a Bayesian learning approach, outperforming existing methods on synthetic and real-world data.
What Problem Does This Paper Address?
In online marketplaces like eBay, where buyers and sellers negotiate prices sequentially, understanding a seller's true valuation of an item is crucial for analyzing their behavior and optimizing the platform. However, these private valuations are not directly observable. Traditional methods often assume perfect rationality and convergence to equilibrium, which doesn't hold in real-world scenarios with millions of agents possessing varying cognitive abilities and information. This paper aims to bridge this gap by inferring private valuations from observed bargaining data, even when agents are not perfectly rational.
Methods
The core of the proposed approach, BLUE (Bayesian Learning-based Valuation Inference), involves three key steps:
- Behavior Modeling: Each bargainer's decision-making process (mapping their valuation and bargaining history to an action) is modeled using a recurrent neural network (RNN), specifically a Gated Recurrent Unit (GRU).
- Feasible Interval-Guided Learning: Instead of relying on equilibrium assumptions, the paper leverages a mild rationality assumption: bargainers do not choose strictly dominated strategies. This allows the derivation of "feasible intervals" for each seller's valuation. A novel loss function is then designed, incorporating these feasible intervals, to train the RNN behavior models. This ensures the learned behaviors are consistent with rationality and fit the observed data.
- Bayesian Inference: Using the learned behavior models and the observed bargaining data, Bayes' rule is applied to derive posterior distributions for the sellers' private valuations.
To handle the heterogeneity in bargaining behaviors across different sellers, the authors also propose a clustering algorithm called K-Loss. This algorithm groups sellers with similar behaviors, allowing for more specialized behavior models within each cluster, thereby improving inference efficiency and accuracy.
Datasets
The paper uses the following datasets:
- Synthetic Data I & II: Generated by simulating bilateral bargaining with three types of agent behaviors (simple, random, payoff-maximization) and different valuation distributions (uniform and categorical). These datasets allow for controlled experiments and ground-truth comparisons.
- Real-world eBay Data: Collected from eBay's "Best Offer" platform, containing bargaining information like seller ID, item ID, and thread histories. Sellers with insufficient bargaining threads were filtered out, resulting in a dataset of 32,099 sellers and 358,641 threads.
Key Insights and Ideas
- Rationality without Equilibrium: The paper demonstrates that inferring private valuations is possible even without assuming full rationality or equilibrium play. The weaker assumption of avoiding strictly dominated strategies is sufficient and more realistic.
- Feasible Intervals as a Guiding Principle: Deriving feasible valuation intervals from observed behavior provides a powerful constraint for learning accurate behavior models, which in turn leads to better valuation inference.
- Neural Networks for Behavior Modeling: Using RNNs (like GRUs) to model complex, sequential bargaining behavior is effective and can automatically capture nuanced decision-making patterns.
- Clustering for Heterogeneity: Explicitly addressing seller heterogeneity through clustering (K-Loss) significantly improves inference performance, highlighting that a one-size-fits-all model is suboptimal.
Key Results
- Superior Valuation Inference: The proposed BLUE framework, especially when combined with clustering (BLUE-C), consistently outperforms baseline methods (including other learning-based and sampling-based approaches) in terms of Mean Squared Error (MSE) on synthetic data and Rationality-Compatible Inference Rate (RCIR) on both synthetic and real-world data.
- High RCIR: BLUE-C achieves RCIRs above 90% on synthetic data and the highest RCIR (89.0%) on real-world data, indicating that its inferred valuations are highly consistent with rationality constraints.
- Effective Clustering: The K-Loss clustering algorithm demonstrates superior performance in grouping sellers based on their behavior patterns compared to K-Means and an auto-encoder-based approach.
- GRU Efficiency: For behavior modeling, GRUs are found to be efficient and effective, converging to similar performance as Transformers but with faster training times.
Limitations and Future Work
- Limited Scope of Rationality: While the assumption of not choosing strictly dominated strategies is weaker than full rationality, it still imposes some constraints. The paper acknowledges that more complex irrational behaviors might exist.
- Time-Discounted Valuations: The current framework does not explicitly account for time-discounted valuations, which could be a factor in prolonged negotiations.
- Multilateral Bargaining: The paper suggests extending the framework to multilateral bargaining scenarios, where an agent negotiates with multiple opponents simultaneously.
- Theoretical Foundations: Further exploration of the theoretical underpinnings of the proposed methods is suggested.
Summary updated