Scalable Methods of Lazy Learning with Complex Structured Data Based on Pattern Structures

Nowadays, machine learning is applied in most work areas, including those in which decision-making is a matter of great importance. For example, such areas might include credit scoring, medicine or self-driving vehicles. Due to multiple factors, interpretability, that is opportunity of explaining the reasons for any decision was made is usually more important in these fields than further improvements in the decision making optimality. Most of the commonly used classic machine learning algorithms including artificial neural networks have high efficiency but constitute so-called black boxes. That means that a learned classifier by design cannot be represented in a form of a human readable description. Based on Pattern Structures lazy learning classification algorithm is one of interpretable machine learning methods, which uses similarity between examples of training subset and object to be classified during the work. This paper discusses improving performance of this algorithm by the use of a boosting. The basic idea is to assign weights to training objects and to use these weights to indicate which objects allows to make more accurate predictions and should get more attention at the classification stage. Proposed modification preserves the property of interpretability and able to classify faster by the use of finding a relatively small subset of the most significant example.