Explainable Machine Learning and Visual Knowledge Discovery

The importance of visual methods in machine learning (ML) as tools to increase the interpretability and validity of models, is growing. The visual exploration of multidimensional data for knowledge discovery of all possible sizes and dimensions is a long-standing challenge. While multiple efficient methods for visual representation of high-dimensional data exist, the loss of information, occlusion, and clutter continue to be a challenge. This chapter starts with the motivation and differences between analytical and visual ML methods, and approaches, showing the benefits of visual methods for ML. Next, the several types of methods to visualize ML models are presented including input-based and structure-based methods accompanied by examples. The major part of the chapter is devoted to the approaches and the theory, to discover interpretable analytical ML models aided by visual methods. It includes theoretical limits to preserve n-D distances in lower dimensions, based on the Johnson-Lindenstrauss lemma, point-to-point, and point-to-graph General Line Coordinate (GLC) approaches. Real-world case studies illustrate algorithms based on GLC models showing their efficiency. It is followed by the methods to scale visual discovery of ML models, to big data, using a combination of GLC, embeddings, lossless visual interpretation of principal component analysis clustering, and other methods. The chapter reviews visual explanation of analytical ML models, including deep neural networks, and outlines of future directions, in visual knowledge discovery and ML.