ComputeFest 2019: Symposium on Data Science, Machine Learning, and Fairness in Computational Science

Computefest 2019

Computefest 2019

Hosted by the Institute for Applied Computational Science (IACS), ComputeFest is an annual winter event of knowledge and skill-building activities in computational science, engineering and data science. The workshop content compliments the curriculum taught in DataFest.

IACS Symposium: "Data Science at the Frontier of Discovery: Machine Learning in the Physical World"

Tuesday, January 22nd, 2019 Harvard University Science Center, Hall B, 1 Oxford Street, Cambridge MA 02138

Notes

https://christophm.github.io/interpretable-ml-book/ https://www.fatconference.org/2019/program.html

Most of these tools won't work with MLaaS but could be wired into a framework.

https://www.oreilly.com/learning/introduction-to-local-interpretable-model-agnostic-explanations-lime

On Fairness and Interpretability

Model Agnostic Methods for Interpretability and Fairness

look at local local perturbations
decision boundaries
shapeley values

Local Perturbations

lime provides local modifications around the input values driving target

https://github.com/marcotcr/lime

input gradients around spend and volume 4 month sliding
use to clarify impact of feature

https://arxiv.org/abs/1611.07634

hold all factors constant
example was prob of default relative to debt to income
plot holds

BILE Decision Boundary

spend and lift -> SpendLiftV6

Shapely Values

requires retraining 2^F model retraining to determining interpretability.

Workshop

load the training data
perform the core splitting based on the features and the labels
consider the heatmap as a 3d map where one could apply the facet plot
LICE plot uses the values as fixed for the training data then moving each of teh values through
consider looking at points near the decision boundary
ensure that fairness testing are part of the visualization not the pipeline