Explanation of Ensemble Model Predictions

Linear and logistic regressions still remain among the most popular methods in solving problems of regression and classification, despite the fact that their quality (predictive ability) is much worse is more advanced new methods of machine learning. It is justified by the fact that complex models have a poor generalizing ability, which is true in the case of linear and logistic regressions with a large number of parameters, but not fundamentally true for ensemble algorithms (boosting is a prime example). Other factor is a complex interpretation of the model, which is intuitive not as clear as the interpretation of the same linear regression. Interpretending model predictions is as important as building quality predictions. Much attention is paid to machine learning in machine learning signs, as often the number of signs varies from 100 to 1000, depending on the type and complexity of the task, highlighting the most important of them. To solve this problem, there are already metrics (for example, the Gini index, Frequency), using which evaluate the importance of variables, however the problem of all these methods in that they do not show exactly how a particular trait affects model prediction, in connection with which, in practice, these approaches use only in order to throw out from the model all the signs that are lower than divided threshold by one of the metrics. An alternative approach is described to assess the importance of the variables of the model, which does not depend on the complexity models, and to a large extent describes the dependencies in the data, rather than other methods based on the Shapley index from game theory.