eduzhai > Applied Sciences > Computer Science >

A Neural-embedded Choice Model TasteNet-MNL Modeling Taste Heterogeneity with Flexibility and Interpretability

  • Peter
  • (0) Download
  • 20210216
  • Save

... pages left unread,continue reading

Document pages: 33 pages

Abstract: Discrete choice models (DCMs) and neural networks (NNs) can complement eachother. We propose a neural network embedded choice model - TasteNet-MNL, toimprove the flexibility in modeling taste heterogeneity while keeping modelinterpretability. The hybrid model consists of a TasteNet module: afeed-forward neural network that learns taste parameters as flexible functionsof individual characteristics; and a choice module: a multinomial logit model(MNL) with manually specified utility. TasteNet and MNL are fully integratedand jointly estimated. By embedding a neural network into a DCM, we exploit aneural network s function approximation capacity to reduce specification bias.Through special structure and parameter constraints, we incorporate expertknowledge to regularize the neural network and maintain interpretability. Onsynthetic data, we show that TasteNet-MNL can recover the underlying non-linearutility function, and provide predictions and interpretations as accurate asthe true model; while examples of logit or random coefficient logit models withmisspecified utility functions result in large parameter bias and lowpredictability. In the case study of Swissmetro mode choice, TasteNet-MNLoutperforms benchmarking MNLs predictability; and discovers a wider spectrumof taste variations within the population, and higher values of time onaverage. This study takes an initial step towards developing a framework tocombine theory-based and data-driven approaches for discrete choice modeling.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×