Qiita G-test guide demystifies CAM, LIME and SHAP

The article is framed as preparation material for the G-test, a Japanese certification on AI literacy, and it groups four techniques by their core idea rather than by math. CAM targets image models, especially CNNs, and highlights the pixels behind a decision as a thermography-like heatmap; the guide notes the more general Grad-CAM as a well-known extension. Permutation Importance works from the other direction, shuffling one feature at a time and reading importance off the accuracy loss.

LIME and SHAP shift from the model to the individual prediction. LIME generates similar dummy data around one case and approximates it with something simple like linear regression or a decision tree, while SHAP borrows the Shapley value from cooperative game theory to divide up each feature's contribution. The guide's cafe example makes SHAP's arithmetic concrete, and its closing summary table lines each method up with its G-test keywords.

The piece closes by arguing XAI is indispensable for using AI safely and trustworthily in business, medical and financial fields. Whether that argument lands for a given team likely depends on whether they can first match the method to the question they need answered — where an image model looked, or how much each factor moved a forecast.