This repo contains the source code for the research paper http://arxiv.org/abs/2107.11357. The paper proposes a Joint Shapley value that extends the widely-used Shapley value, a key measure of feature importance in game theory. The Joint Shapley value can be used to establish the importance of coalitions of features, as well as individual features, whilst maintaining the axiomatic foundations of the original Shapley value.
- Presentation given at the University of Toronto's Vector Institute Seminar, 26 July 2021:
https://www.youtube.com/watch?v=Wrd7JzYA2sE
The Rotten Tomatoes Walkthrough Jupyter notebook is a walk-through of the Joint Shapley methodology described in the associated paper, using the Rotten Tomatoes movie review database.
The Boston Housing Dataset Walkthrough offers a similar approach, but for the Boston housing dataset.
The datasets folder contains datasets used in the notebooks:
rt-polaritydatacontains the Rotten Tomatoes data. This is divided into positive reviews (*.pos) and negative reviews (*.neg). The 'mangled' suffix indicates that the order of the reviews has been scrambled.
The presentations folder holds presentation materials regarding the Joint Shapley approach.
Chris Harris - [email protected]
Richard Pymar - [email protected]
Colin Rowat - [email protected]