Causal inference, that is estimation of impacts of actions, is a powerful analytical approach that has been getting a lot of press recently. The difference to classic supervised ML is that the impact of an action is never possible to observe directly at single data point level, as we never observe the counterfactual outcome – so different techniques are needed.  A large variety of advanced causal inference models have been developed and open sourced recently; yet their very variety makes it hard for a novice to the field to select the best one for their case. That is why we have leveraged three open source libraries from Microsoft, DoWhy, EconML, and FLAML, to create the auto-causality package, which provides automated selection of the model that performs best on your dataset, and visualize its results.

This talk will walk you through a practical application of auto-causality to analysis of real A/B test results from Wise’s feature testing and CRM campaigns, and show how it can help you to not just measure average impact, but to see which customers respond better to what option.

Supported by