amadeus-research-publications

PCMC-Net: Feature-Based Pairwise Choice Markov Chains

← Back to Publications

Publication Details

• Authors: Lhéritier, A.
• Venue: ICLR 2020
• Year: 2020
• Link: OpenReview

Abstract

PCMC-Net is a novel choice modeling framework that combines deep neural networks with the theory of Pairwise Choice Markov Chains (PCMC). The model uses learned feature representations to predict probabilities of choosing one option over another in a set, capturing complex substitution effects. PCMC-Net can handle high-dimensional feature data (e.g., images or text describing options) and still produce choice probabilities that are consistent with rational utility theory. Experiments on public choice datasets demonstrate that PCMC-Net achieves state-of-the-art prediction accuracy, while offering more interpretable structure than a standard neural network softmax approach.

Key Contributions

• Novel neural architecture for discrete choice modeling
• Integration of Markov chain theory with deep learning
• Efficient handling of variable-sized choice sets
• Strong empirical results on travel choice prediction
• Theoretical analysis of model properties and convergence

Citation

@inproceedings{Lheritier2020PCMC,
  author    = {Alix Lh{\'e}ritier},
  title     = {PCMC-Net: Feature-Based Pairwise Choice Markov Chains},
  booktitle = {Proceedings of the 8th International Conference on Learning Representations (ICLR 2020)},
  year      = {2020}
}

---

← Back to Publications