Visual Neural Decomposition to Explain Multivariate Data Sets

Abstract

Fig. 1. Visual Neural Decomposition of a chip testing measurement data set with the goal to identify cases in which the target variable (here: jitter) exhibits high values. Each node visualizes parts of the data set depending on its activation.

Random Forest interpretability - explaining classification models and multivariate data through logic rules visualizations

“…Apart from predictive capability, classification models can also serve as descriptive tools, distinguishing data instances among different classes (TAN; STEINBACH; KUMAR, 2005). Once interpretable, such models are suitable for multivariate data explanation, as recently proposed by VA solutions arranging visual representations of SVM and ANN models (GLEICHER, 2013;KNITTEL et al, 2020). Nevertheless, concepts like Emerging Patterns arrange descriptive logic rules (NOVAK; LAVRAC; , where a specific type, called Jumping Emerging Pattern (JEP), provides high discriminative power between classes (KANE; CUISSART; CRÉMILLEUX, 2015;.…”

“…The potential of descriptive logic rules lies in understanding the phenomenon represented by data (NOVAK; LAVRAC; . Data explanation approaches based on classification models visual representations support knowledge generation from complex data (KNITTEL et al, 2020), which is the VA's primary goal involving insights and hypotheses (KEIM et al, 2010;SACHA et al, 2014). Although inspiring, the proposed VA solutions for descriptive purposes make use of black-box models (SVM and ANN) (GLEICHER, 2013;KNITTEL et al, 2020), requiring constraints for reaching interpretability.…”

“…Data explanation approaches based on classification models visual representations support knowledge generation from complex data (KNITTEL et al, 2020), which is the VA's primary goal involving insights and hypotheses (KEIM et al, 2010;SACHA et al, 2014). Although inspiring, the proposed VA solutions for descriptive purposes make use of black-box models (SVM and ANN) (GLEICHER, 2013;KNITTEL et al, 2020), requiring constraints for reaching interpretability. On the other hand, JEPs are naturally interpretable and descriptive, as well as suitable for visualization methods (NOVAK; LAVRAC; .…”

“…In summary, since classification models can be used as predictive and descriptive tools (TAN; STEINBACH; KUMAR, 2005), model interpretability may involve visual explanations for model (MING;QU;ZHAO et al, 2019) and data (GLE-ICHER, 2013;KNITTEL et al, 2020) understanding. Logic rules are helpful in model interpretability solutions (LAKKARAJU;LESKOVEC, 2016;GUIDOTTI et al, 2018b) and can be extracted from DTs MIRANDA;SARDINHA;CERRI, 2021).…”

“…The idea of joining these two concepts has been suggested (GLEICHER, 2013;KNITTEL et al, 2020), using machine learning classification models as surrogates to explore an underlying phenomenon represented by data. The core concept is based on the fact that classification models can be used not only for predictive analysis but also for descriptive purposes (TAN;STEINBACH;KUMAR, 2005).…”