Introduction:

The Bradley-Terry-Luce model (aka Multinomial Logit model) is a classical tool for modelling how humans make choices when presented with a class of alternatives [0]. The model assumes that users act rationally by maximizing their utility. However, it is well known in psychology and behavioural economics that human beings do not always act rationally. Moreover, the Bradley-Terry-Luce model also assumes the 'irrelevance of alternative assumptions' (IIA), which is often violated in practice. Recent work [1,2,3,4] has shown that richer choice models can capture some of these subtler effects and can outperform the Bradley-Terry-Luce model.

Specific Problem:

In the first half of the project, you will spend your time reading and understanding the papers and the code in [1,2,3,4] and draw up the merits and limitations of each work. In the second half of the project, you will apply some of these methods to real-world datasets and try to interpret the models that are learned from them. In particular, we will try to apply these models to datasets involving similarity comparisons, as in [5].

Prerequisites:

1. A grade of 5.5 or above in Modèles stochastiques (or any undergraduate probability course)
2. Good performance in mathematical courses
3. Prior experience in working with real data

If you are interested, please contact me with your CV and transcript.

References:

[0] https://en.wikipedia.org/wiki/Bradley–Terry_model

[1] Seshadri, A., Peysakhovich, A. and Ugander, J., Discovering context effects from raw choice data. In International conference on machine learning (ICML 2019).

[2] Seshadri, A., Ragain, S. and Ugander, J., Learning rich rankings. Advances in Neural Information Processing Systems, (NeurIPS 2020)

[3] Tomlinson, K. and Benson, A.R., Learning interpretable feature context effects in discrete choice. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining (2021).

[4] Tomlinson, K., Ugander, J. and Benson, A.R., 2021, Choice set confounding in discrete choice. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining (2021).

[5] Wilber, M., Kwak, I. and Belongie, S., Cost-effective hits for relative similarity comparisons. In Proceedings of the AAAI Conference on Human Computation and Crowdsourcing (2014).