Causal feature selection for predicting interventions

Abstract

With the advent of high throughput experiments, researchers are more and more often confronted to really big datasets especially in biology. The number of variables can reach several thousand and to deal with this increasing number of features, researchers are using feature selection techniques to reduce the size of the datasets and so reduce the costs linked to the experimentation. In machine learning, the objective is to learn a model on a dataset in order to predict the target variable of a second one drawn from the same distribution. Sometimes, the distribution is not the same between the two databases. For example, in biology, an experiment on the genes can be done on population from various places on earth. Since the genomes between different population present some variations, it is important to take them into account before selecting genes to predict the result of the experiment. These variations can also come from external agents like drugs which manipulate some features. In this thesis, we focused on data generated from causal graph because it is representative of the phenomena met in biology like gene network. We were interested in a particular situation where the target is provided for an unmanipulated dataset but the objective is to predict the target of a manipulated database. We developed a three stages process to find an optimal subset of features involving causal feature selection, filtering and two algorithms developed in the framework of this thesis. The first step was to find a small set of features made of both highly correlated manipulated and unmanipulated variables. The two following steps focused on retrieving the manipulations and removing the bad features. The whole process was able to significantly increase the prediction efficiency compared to classical feature selection techniques.