Visualizing Classification Results: Confusion Star and Confusion Gear

Recent developments in machine learning applications are deeply concerned with the poor interpretability of most these techniques. To gain some insights process designing data-based models it is common to graphically represent algorithm’s results, either their final or intermediate stage. Specially challenging task plotting multiclass classification results as they involve categorical variables (classes) rather than numeric results. Using well-known MNIST dataset and a simple neural network an example, this paper reviews existing techniques visualize from those centered on particular instance set instances, representing overall performance metric. As commonly summarized form confusion matrix, special attention paid its graphical representation. From analysis, new visualization tool derived, which presented two forms: star gear. The errors, while gear focuses hits. proposed tools also evaluated when facing: (i) balanced imbalanced classifiers issues; (ii) problem errors different orders magnitude. By using shapes instead colors value each matrix cell, significantly improve readability matrices. Furthermore, we show how area enclosed by stars gears directly related standard metrics. graphic can be usefully employed performances sequence classifiers.