Explainable AI for static & Time Series Models, Deltika

Kevin Daniels

Data Science Fellows June 2020 Cohort

Abstract

The project focuses on building and improving and the explainability method called through “Counterfactual samples” for static and time series data. The objective is to find the smallest change in the feature values required to achieve a desired target from a trained model. 

Challenges (at least two)

1. Creating Counterfactual Samples through Gaussian Mixture Models adding high dimensionality.
2. Creating a toy dataset from scratch adding time series correlated features through defined rules and patterns.

Achievements (according to KPIs)

1. Efficient data generation.
2. Training Time Series Model with ARIMA and LSTM
3. New explainability method by creating Counterfactual Samples to understand feature importance.

Future project development 

The next steps of the project will be to apply Time Series into the Gaussian Mixture Model data generation and implement it to the toy dataset that has been created. The objective will be to create Counterfactual samples that will revolutionize the way of explaining feature differences and having a Baseline Model that can be used for more complex datasets in the future.