Heuristics of constructing the architecture of an interpreted machine learning model

Abstract

Interpretability is an important vector of development of modern applied artificial intelligence. It is also necessary to understand how and why machine learning models predict the end result. However, the implementation of such models is a complex process due to the need to meet the requirements of interpretability while maintaining high quality approximation. The article presents an overview of heuristics for constructing an interpreted machine learning model, which allows you to determine the most important features when predicting the target class of data. As an example, the subject area of mining was considered, and as a problem - the prediction of seismic hazard in the conditions of mining enterprises. However, the transformed concept of the interpreted machine learning model allows solving problems in many other subject areas, where positive numerical values are defined as input data, and the number of entries in the set does not exceed 50. Such restrictions on the set of input data are dictated by a feature of the real architecture of the interpreted model of applied artificial intelligence. In conclusion, the authors of the article consider methods that will allow to overcome such a “bottleneck” effect.