Mixed Causal Structure Discovery with Application to Prescriptive Pricing

Wenjuan Wei, Lu Feng, Chunchen LiuIJCAI 2018 · DOI

TL;DR

This paper introduces a novel method for "mixed causal structure discovery" to build more accurate demand models for prescriptive pricing. By developing a new information criterion (MIC) and an efficient inference algorithm (A* FoBa), they improve the recovery of price-demand relationships, leading to better demand forecasting and ultimately more profitable pricing strategies.

What Problem Does This Paper Address?

Prescriptive pricing aims to find the optimal price strategy to maximize profit by first modeling demand and then optimizing prices based on that model. A crucial step is demand modeling, which involves uncovering the causal relationships between prices, sales, and other influencing factors. Existing methods often struggle with:

  • Mixed Data Types: Real-world pricing scenarios involve a mix of continuous variables (like price, sales, temperature) and discrete variables (like season, day of the week, promotions). Many causal discovery methods are designed for only one type.
  • Accuracy and Efficiency: Accurately recovering these causal relationships is vital, as biased estimates lead to suboptimal pricing. Furthermore, the process needs to be efficient to adapt to fast-changing markets, especially for online pricing.
  • Correlation vs. Causation: Traditional regression-based approaches often confuse correlation with causation, leading to "pseudo causes" that inflate the complexity and reduce the accuracy of the demand model.

This paper tackles these challenges by proposing a method that can handle mixed data types while accurately and efficiently discovering causal structures relevant to pricing.

Methods

The authors propose a novel approach centered around a "mixed causal model" and a new scoring criterion:

  • Mixed Causal Model: They define a causal model that can represent relationships between a mixture of binary, categorical, and continuous variables. For continuous variables, they use linear relationships with additive noise (Laplace distribution for robustness), and for binary variables, a logistic function. They prove this model is "bivariate identifiable," meaning that in simple two-variable cases, the causal direction can be uniquely determined.
  • Mixed Information Criterion (MIC): To score the fit of a causal graph (DAG) to mixed data, they introduce MIC. This criterion is designed to be "locally consistent," meaning it correctly identifies whether adding or removing an edge improves the model fit. MIC uses a relative loss measure, normalized by an optimistic estimate, to be robust to scale variations and estimation errors inherent in mixed data.
  • A* FoBa Inference Algorithm: They develop an efficient inference framework, A* FoBa, for discovering the causal structure. This algorithm combines an A* search strategy with the Forward-Backward greedy (FoBa) algorithm for sparse regression. To further accelerate the search, they incorporate "ancestor-based search space cutting," which prunes redundant search paths based on known causal relationships.

Datasets

The paper uses both simulated and real-world datasets for evaluation:

  • Synthetic Data: They generated data simulating prices and sales of 40 beer products, along with 9 objective factors, across various noise types, sparsity levels, and training/testing data scales. This allowed for controlled experiments to assess performance against ground truth.
  • Real-World Retail Data: They utilized daily prices and sales data for 10 products (milk and bread) from a Tokyo supermarket over two years, along with 69 objective factors (seasonality, weather, guest flow, etc.). This dataset was used to demonstrate the practical applicability of their method.

Key Insights and Ideas

  • MIC's Robustness: The proposed MIC criterion is a significant advancement because it effectively handles the disparate scales and distributions of mixed data types, a common bottleneck in causal discovery for real-world applications. Its relative loss formulation makes it more resilient to noise and scale differences than absolute loss measures.
  • A* FoBa Efficiency: The A* FoBa algorithm, with its search space pruning techniques, demonstrates that efficient causal discovery is achievable even for complex, mixed-variable systems, making it practical for time-sensitive applications like online pricing.
  • Causality for Profitability: The study strongly suggests that accurately identifying causal relationships (rather than just correlations) is not just an academic exercise but directly translates to tangible business benefits, such as higher profits through better pricing strategies.
  • The Value of "Tepid" Days: The real-world results highlight that the proposed method's pricing strategies are particularly effective during periods of lower demand, suggesting it can help businesses navigate challenging market conditions more effectively.

Key Results

  • Structure Discovery: The MIC-A*FoBa method significantly outperformed other benchmark methods in terms of Precision for causal structure discovery across various simulated settings. While Recall was comparable to some methods, the higher Precision indicated more reliable causal relation detection.
  • Demand Forecasting: The method achieved the smallest prediction error (averaged RMSE) for demand forecasting compared to all other tested causal methods.
  • Pricing Accuracy: The pricing strategies derived from the MIC-A*FoBa demand models resulted in the highest gross profits on both simulated and real-world data. On the real retail data, their method achieved the highest validation revenue, outperforming GES, causalMGM, and FoBa.
  • Efficiency: The acceleration policies in A* FoBa were shown to significantly reduce elapsed time for inference, making the method more practical.

Limitations and Future Work

  • Model Assumptions: The current mixed causal model assumes linear relationships for continuous variables and additive noise. While they argue this simplifies optimization, real-world relationships can be more complex. Future work could explore non-linear extensions.
  • Identifiability in Complex Cases: While bivariate identifiability is proven, identifiability in more complex, multivariate mixed-variable settings remains an open area for deeper theoretical investigation.
  • Computational Complexity: Although the A* FoBa algorithm is efficient, causal discovery can still be computationally intensive for very large numbers of variables. Further optimizations or approximations might be needed for extremely large-scale problems.
  • Real-World Data Granularity: The paper mentions using a few products for clarity in the causal graph visualization. Future work could explore the full complexity of larger product catalogs and their interdependencies.

Summary updated

BibTeX
@inproceedings{wei2018mixed,
  author = {Wenjuan Wei and Lu Feng and Chunchen Liu},
  title = {{Mixed Causal Structure Discovery with Application to Prescriptive Pricing}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/711},
}
Figure