Multilayer machine learning with visual knowledge discovery

Abstract

Increasing domain experts’ trust in Machine Learning (ML) models requires explainable models. However, employing explainable techniques alone can be insufficient for getting such trust in the model. The reasons include insufficient model accuracy, overfitting, overgeneralization, too complex explanations, and others. Integration of multilayer/multilevel ML and lossless visualization approaches for high-dimensional data is proposed to help address these challenges. Multilayer/multilevel ML approaches have been extensively developed to capture complex relations especially within modern deep learning (DL) approaches. Graph-based lossless visualization approaches for high-dimensional data emerged recently in ML as an opportunity to represent high-dimensional ML information without loss of n-D information as an alternative to lossy dimension reductions methods. These approaches within a Visual Knowledge Discovery (VKD) paradigm use General Line Coordinates (GLC) to create explainable and reversible n-D representations in a visual form with intent to increase user’s trust in ML models. However, when viewing vast volumes of data, occlusion affects visual approaches. Both multilayer/multilevel and visualization approaches have challenges of becoming too complex for the user. This paper adapts to these challenges by developing multilayer/multilevel ML approaches based on hyper-rectangles/hyperblocks/hyperboxes and a classifier decomposition at several levels as simple base elements for both multilayer/multilevel ML and visualization. The efficiency of the proposed integrated approach is demonstrated in several cases studies on numeric data and images. Major benefits from this approach are increased generalization and less complex interpretable rules to increase user confidence in each classification.