Evaluating Machine Learning models on Discrete Choice data
Responsable(s) :
Gael Lederrey, Anna Fernandez Antolin, Michel Bierlaire
Description :
The objective of this project is to find alternative solutions to the existing Discrete Choice Models (DCMs). The field of Machine Learning (ML) offers many algorithms for multi-class classification. The first step for the student will be to review the literature and identify ML models that can be used for multi-class classification and can lead to similar results as the DCMs. The second step of this project will be to test and validate them on given data. (More information about the data will come later.) The final step will be to use state-of-the-art DCMs and compare them to the ML models. One additional step is the mathematical comparison of these types of models (DCM vs. ML model) to extract the fundamental differences between them. Let me give you an example. Imagine that you have an object O. It has a probability of 60% to belong to class A and a probability of 40% to belong to class B. In ML, most of the algorithms will simply say that O belongs to class A because it's the class with the highest probability. If you have a second object P with the same probabilities, then it will also belong to class A, etc. In the end, for all the objects having the same probabilities, most of ML algorithms will select class A. And this is wrong. DCMs are probabilistic, so they don't have this drawback.
Collaboration with:
Type :
semester project
Pré-requis :
The student needs good knowledge in mathematics (especially probability and statistics) and experience with either ML or DCM. Sufficient programming skills in either Python, Matlab, or Scala (ev. Java) are required to implement and test the models.
Submitted on :
September 06, 2017