|Optimal choice of propensity function for automatic causal estimator selection|
|Organisatsiooni nimi||Wise Plc|
|Kokkuvõte||Automatic selection of causal estimators has a huge potential for quantifying impact of business choices, especially so in areas where the choice cannot be randomized and so can't be handled by the classic A/B testing methods. When the outcome of interest and the choice of 'treatment' that we assess the impact of, are affected by the same external factors (such as properties of an individual customer), an illusion of causal linkage can arise if these factors are not correctly accounted for, known as 'confounding'. Virtually all methods for such a correction involve at some stage the fitting of a 'propensity to treat' function that predicts the probability of an individual being assigned the treatment as a function of their properties; that function also is used in the out-of-sample scoring methods used to select among different estimators on a particular dataset.|
Preliminary experiments (https://openreview.net/forum?id=nRxPliRs4Ar) suggest that using a fully flexible AutoML classifier for the propensity function doesn't lead to optimal estimator fitting and selection. This thesis will experiment with different formulations for the propensity function, such as Naive Bayes, and other approaches, such as label smoothing, to see if these can improve the quality of causal estimator fitting and selection on synthetic data, using https://github.com/transferwise/auto-causality as the starting point.
|Lõputöö kaitsmise aasta||2022-2023|
|Juhendaja||Dr. Egor Kraev (Wise Plc) and Prof. Radwa El Shawi|