Fig. 1

EIG enhancements to identify relevant features per sample or task. a Top: Different linear and nonlinear paths for EIG. (1) Linear path in original feature space (O-L-IG, black line), (2) neighbors path in the original feature space (O-N-IG, blue line), (3) linear path in the hidden feature space (H-L-IG, gray line), and (4) neighbors path in the hidden feature space (H-N-IG, green line). Bottom: visualization of such paths for a specific sample using splicing data from [16] and PC1, PC2 of the feature space (see Additional file 1: Section S1). b Different group-agnostic (zero and encoded-zero) and group-specific (k-means, median, close, random) baselines. Encoded-zero baseline is generated by decoding a zero vector in the latent space. A close baseline is created by taking a baseline point which is close to the sample in euclidean distance (see the “Methods” section). c Framework to identify significant features that distinguish sample 5 from baseline 3. The digit images show mean of 300 examples of sample digit 5 and median of 300 examples of baseline digit 3. The distribution plots are illustrative only. They show difference between distribution of attributions of two sets of samples for a significant and a not significant feature