Explanation methods — Kaimei API Documentation

method on POST /v1/explanations picks the attribution algorithm. GET /v1/models/{model} lists the methods a model supports in explain_methods, the recommended one first.

Method	What it computes	Needs	Notes
`shap`	KernelSHAP. Weighted linear regression over feature coalitions, each coalition averaged over background rows. Attributions sum to `output.value − base_value`.	Tabular, runnable	Up to 12 features every coalition is enumerated, so the result is exact for the sampled background and reproducible. Above that a coalition budget (`n_samples`) is drawn.
`shapley`	Permutation Shapley sampling. Marginal contributions over `n_permutations` random feature orders.	Tabular, runnable	Each `top_features` row carries a `stderr`. The default `auto` for tables.
`lime`	A local linear surrogate fitted to perturbed neighbours of the input (feature noise for tables, on/off regions for images).	Any runnable model	`options.local_r2` reports the surrogate's fit; `kernel_width` the locality.
`occlusion`	Hides each image region (blur for photos, gray for thumbnails) and measures the change in the target class's log-odds.	Image, runnable	The default `auto` for images. Works for remote endpoints too.
`gradcam`	Class-activation map from the last convolutional feature map of an uploaded network, read by central differences.	Image, uploaded structure	`options.layer` picks another spatial layer. Heatmap in `[0, 1]`.
`auto`	`shapley` for tables, `occlusion` for images.	—	Stored under the method actually used.
attributions	No computation. You send the attributions, Kaimei stores and ranks them.	Nothing	`method` is recorded as `provided`. See bring your own attributions.

Wide inputs are handled by screening: above 96 features, Shapley sampling first finds the features that move the output (in groups when there are tens of thousands), samples over the strongest 48, and reports the rest as one residual so the attributions still sum to f(x) − E[f(z)]. LIME and KernelSHAP refuse inputs wider than 4,096 features.

Image explanations time one copy of the picture first and coarsen their region grid (and LIME's samples) so the whole explanation finishes within the API's 40-second budget; the response then also carries grid_requested or n_samples_requested and budget_seconds.