What Kaimei can and can't explain
A plain list of the models Kaimei explains, the ones it explains only through your own scoring endpoint, and the ones it does not support.
#Explained today
Kaimei explains models that make one prediction from a fixed set of inputs. Each input is a number, a category, a yes/no flag, or an image.
| Kind of model | Examples | How it connects |
|---|---|---|
| Yes/no and multi-class decisions on structured data | Credit approval, fraud flags, churn, claims triage | Upload it from scikit-learn, Keras or PyTorch, or connect a scoring endpoint. |
| Numeric predictions on structured data | Demand forecasts from features, pricing, loss estimates | The same. |
| Image classification | Defect detection, document type from a scan, species | Upload a convolutional network, or connect a scoring endpoint. |
| Models Kaimei cannot run itself, on the same kinds of input | XGBoost, LightGBM, CatBoost, large forests, models written in other languages | A scoring endpoint you host, or attributions you compute. |
#Not supported
- Free text and language models. Kaimei has no text input, so it cannot attribute a prediction to words or sentences. That rules out sentiment or topic classifiers on raw text, and large language models such as chat assistants, summarisers and code generators. If a text model's decision is made from structured features you derive, such as scores or counts, you can explain those features.
- Generative models that produce text, images or audio. An explanation needs a single score or label to explain.
- Raw sequences and time series. There is no sequence input, so models that read a history step by step are not supported. A forecasting model that takes fixed features, such as lagged values and rolling averages, works.
- Several outputs at once. Kaimei explains one label or one number per prediction. Register one model per output instead.
- Boxes, masks and rankings. Object detection, image segmentation, recommendation and search ranking.
- Audio and video.
- Clustering and reinforcement learning. A model that outputs a single anomaly score can be registered as a numeric prediction.
#Partly supported
- Networks with attention, recurrent or upsampling layers. Kaimei can draw the architecture but cannot run those layers. Connect a scoring endpoint to explain the model, provided its inputs are structured data or images.
- Large forests. Kaimei runs forests of up to 30 trees and 4,000 nodes itself; larger ones need a scoring endpoint.
- Very wide inputs. Above 96 inputs, the Shapley methods first find the inputs that matter and report the rest as one remainder. LIME and KernelSHAP stop at 4,096 inputs.
- Grad-CAM needs an image model whose layers were uploaded. Occlusion and LIME work for any image model.
#What Kaimei does not do
- Measure fairness or disparate impact.
- Establish cause and effect.
- Serve your model to customers. It calls your model, or runs an uploaded copy, to explain and analyse it.
- Check your data beyond comparing live traffic with the reference data.
- Delete records on a schedule. Records are kept until you delete them.
If your model is not covered here, tell us about it.