Jiricka, L. (2022). From local to global explanations and communication of decision differences between two black-box classifiers [Diploma Thesis, Technische Universität Wien]. reposiTUm. https://doi.org/10.34726/hss.2023.98693
E194 - Institut für Information Systems Engineering
-
Date (published):
2022
-
Number of Pages:
62
-
Keywords:
Explainable Artificial Intelligence; Local explanation; Global explanation; model-agnostic method; black box
en
Abstract:
The use of neural networks, support vector machines and gradient boosting models, among others, has increased significantly in the past 10 years. Especially in crucial areas, the understanding and traceability of the decisions of such black box models is of fundamental interest.Most models are trained on the basis of historical data at a specific point in time. However, the distribution of measured data may change over time and hence the model has to be updated. Comparing different versions of models to detect decision differences supports detecting such concept drifts.DiRo2C (Difference Recognition of 2 Classifiers) aims at recognizing decision differences locally, that is for a specific instance, between two black box classifiers using a modified genetic algorithm to create a synthetic dataset, consisting of data points similar to the to be explained instance and the corresponding decisions of the two black box classifiers. On this synthetic dataset a decision tree is trained to learn and explain decision differences. The main focus of this thesis is the problem of global explainable artificial intelligence through local explainers in the setting of difference recognition of two black box classifiers. We propose approaches to derive global explanations using concepts of local explanation generation of DiRo2C in addition to accompanying strategies to communicate those explanations. Experiments show that a ’Local to Global’ approach using clustering concepts to derive a description of the characteristics of the decision differences of the black box models, yields similar performance to a decision tree trained on the original dataset while reducing the complexity of the explainer.