Comparative analysis of interpretable artificial intelligence methods
Article Sidebar
Issue Vol. 38 (2026)
-
Comparative analysis of Next.js and Astro frameworks
1-5
-
Behavioral analysis of ransomware threats to ESXi Hypervisors: a machine learning-based predictive model
6-10
-
Performance and usability evaluation of a VR virtual museum application
11-18
-
Comparative analysis of selected data visualization methods
19-25
-
A performance comparison of web programming interfaces: GraphQL, gRPC and Thrift
26-31
-
A review of security mechanisms in electronic payment systems
32-42
-
Clustering methods in machine learning
43-50
-
Comparative analysis of interpretable artificial intelligence methods
51-58
-
Comparative analysis of machine learning classifiers
59-65
-
Analysis of the impact of machine learning algorithms on the quality of generated sounds
66-72
-
Comparative analysis of the functionalities of applications supporting the self-control process of anticoagulation therapy
73-80
-
Application of machine learning for predicting Formula 1 race results
81-86
-
Analysis of latency reduction and performance improvement methods in selected VR applications
87-94
-
Integrating deep learning image analysis into Web GIS applications: A Hybrid Flask - Spring Boot architecture for automated place detection
95-101
Main Article Content
DOI
Sustainable Development Goals (SDG)
- Quality education
- Industry, Innovation, Technology and Infrastructure
Authors
Abstract
The aim of this article is to analyze and compare methods for explaining the results of artificial intelligence methods. Three methods were analyzed: Grad-CAM, SHAP, and LIME, evaluated in terms of their effectiveness on different data types. The analysis used five datasets: Iris, Wine Quality, Brain Tumor Dataset, PHCD, and WheatGrain. Two datasets are tabular, two are image, and one is mixed. SHAP and LIME were applied to tabular datasets, while all three methods were used for image data. Grad-CAM proved the fastest and most effective in locating key regions, while SHAP was slower but more accurate in pixel attribution, and LIME achieved the lowest precision. For tabular data, SHAP provided more accurate and consistent explanations than LIME, especially for high-dimensional datasets.
