 My name is Alicia Gosewska and I am from MI Square Data Lab and today I will talk about lifting interpretability performance through ADO via automated feature engineering. So nowadays we start to pay attention on how to explain the complex models. We want to understand how they work and what to do to improve them. There is a whole idea called an explainable artificial intelligence in short XAI that focuses on methods to help to explain machine learning models. One of the applications of XAI method is to use them to extract knowledge from complex models to build simple interpretable ones. And as a result, we gain better efficiency and do not have to worry about lack of transparency of our models. Now I would like to present a framework for supervised assisted feature extraction for machine learning models. It's called SAFE ML and it uses elastic complex algorithm as a supervisor model to create an interpretable yet still accurate Glassbox model. And the main idea is to train a new interpretable model, a newly engineered feature extracted from the supervisor model. So the method can be described in six steps. First one is to provide a raw tabular data set. Second step is training a supervisor complex machine learning model on the provided data and this model does not need to be interpretable. It just need to be accurate. The first step is to use SAFE to find variable transformations. And for continuous variables, we use the partial dependence profiles to find change points that allow the best binding of variables. For categorical variables, we use clustering to merge some of the levels. The fourth step is optional. Here we perform a feature selection on the new set of features that includes original variables from the raw data and variables transport with the SAFE method. The fifth step is fitting a fully interpretable model on the selected features. Two models used here are, for example, logistic regression for classical problems or for regression problems and simple linear models just to ensure that they are interpretable. And the last step, you can enjoy your fully interpretable and accurate model. We performed a benchmark of SAFE method on binary classification dataset from our OpenML 100 benchmark. Here we can see results for three different supervisor models. On the x-axis is model's performance. On the y-axis is interpretability of a model. The higher, the better, of course. And arrows show the change of performance and interpretability when using SAFE. So each arrow stands for a different dataset. The beginning of the arrow shows the performance and interpretability of a complex supervisor model. And arrowhead shows the performance and interpretability of SAFE-based interpretable model built on features extracted from the complex model, this one from the start of the arrow. And blue arrows are averages built over different datasets. So blue arrows are averages over these gray arrows. So here we can see that using SAFE increases the interpretability, but we do not lose much on model performance. And SAFE method is implemented as an R package, RSAFE. You can find it on GitHub and on Chrome. SAFE is also a collaborative effort. Therefore, I would like to thank my collaborators from Warsaw University of Technology, Anna Kozak and Przemysław Wiecek for working together on the article about SAFE. And I would like to thank Anna Gelak for the effort to put in the development of the R package. And if you are interested in SAFE, you can find more information in our paper in the decisions of our systems journal or in a short introduction to SAFE in our post on the responsible ML blog on Medium. Thank you.