I performed Random Forest–based feature importance analysis on the MNIST dataset, focusing only on digits 0 and 1. When I visualize the importance map (see image below), it doesn’t resemble the expected shapes of 0 and 1. What confuses me is that there is no visible coloring (importance) in the upper and lower parts of the digit 0, where I expected to see some contribution. Why might this be happening? • Is it because the pixels in those regions are not contributing to the classification between 0 and 1? • Or could this be due to normalization, model bias, or the way PFI permutes features? I would appreciate an explanation of what this pattern means in the context of feature importance for digit classification.
