Reproducible
The same request, the same numbers
KernelSHAP enumerates every feature coalition for models with up to 12 inputs, so the result is exact for the reference sample. Sampling methods take a seed.
Solutions · Research and teaching
For academic groups, research labs and public bodies. Exact where it can be, seeded where it samples, and open about its methods: results another group can reproduce, and a model anatomy students can walk through weight by weight.
What you get
The properties a reviewer, a co-author or a student will ask about.
Reproducible
KernelSHAP enumerates every feature coalition for models with up to 12 inputs, so the result is exact for the reference sample. Sampling methods take a seed.
On the record
An explanation is stored with its method, options, seed, model version and reference sample. An analysis is stored with its row counts and model calls.
Exportable
Structures, attributions, behaviour curves and drift reports come back as JSON, ready for your own figures and tables.
Teachable
Model anatomy draws every layer, split and weight, and a trace follows one prediction through them.
Honest
Permutation sampling reports a margin of error for each attribution. LIME reports how well its local fit holds. Non-monotonic effects are named as such.
Private
Self-host the API, explain through your own endpoint, or upload shapes only.
Methods
Kaimei implements published methods and says which one produced each number.
| Method | What Kaimei computes | Reference |
|---|---|---|
| KernelSHAP | Shapley values against a reference sample; exact by enumeration up to 12 inputs. | Lundberg & Lee, 2017 |
| Permutation Shapley | Monte Carlo Shapley values over random feature orderings, with a standard error per input. | Štrumbelj & Kononenko, 2014 |
| LIME | A weighted local linear surrogate, for tables and image patches, with its fit reported. | Ribeiro, Singh & Guestrin, 2016 |
| Grad-CAM | Class evidence on the last convolutional feature map of an uploaded network. | Selvaraju et al., 2017 |
| Occlusion | The change in the class score as patches of the image are hidden. | Zeiler & Fergus, 2014 |
| Permutation importance | How far the model's output moves when one input is shuffled across the reference rows. | Breiman, 2001 |
| Partial dependence and ICE | The average and per-row response as one input is varied, with a monotonicity score. | Friedman, 2001; Goldstein et al., 2015 |
| Interaction strength | The share of two inputs' joint effect that is not additive. | |
| Population stability index | Drift of each input, and of the output, between reference rows and logged predictions. |
Each method is described, with its options and limits, in Explanation methods and Analyses.
Worked examples
| Example | Data | Model |
|---|---|---|
| Banking and lendingRead the example | Yeh, I.-C. and Lien, C.-H. (2009). The comparisons of data mining techniques for the predictive accuracy of probability of default of credit card clients. Expert Systems with Applications 36(2). Source | credit-default-forestscikit-learn random forest |
| Capital marketsRead the example | Kenneth R. French Data Library, 49 Industry Portfolios, monthly value-weighted returns. Source | sector-rotationscikit-learn gradient boosting |
| InsuranceRead the example | Dutang, C. and Charpentier, A. CASdatasets: freMTPL2freq, French motor third-party liability claims. Source | motor-frequencyscikit-learn gradient boosting |
| HealthcareRead the example | Strack, B. et al. (2014). Impact of HbA1c measurement on hospital readmission rates: analysis of 70,000 clinical database patient records. BioMed Research International. Source | readmission-30dscikit-learn gradient boosting |
| TechnologyRead the example | IBM sample data: Telco Customer Churn. Source | churn-risk-gbmscikit-learn gradient boosting |
| Energy and industryRead the example | Salam, A. and El Hibaoui, A. (2018). Comparison of machine learning algorithms for the power consumption prediction: case study of Tetouan city. IRSEC. Source | load-forecastscikit-learn neural network |
client.models.upload, with reference rows drawn from the training data.Teaching
The live model anatomy draws a real random forest and lights up the path one row takes through it, split by split.
Model anatomyThe Titanic walkthrough is the teaching example: upload, explain, analyse, monitor and version a model, with every output saved.
The notebooksA short guide to reading attributions without over-reading them: not causation, not a fairness test, not a check of the data.
Reading an explanationWho it is for
Exact and seeded attributions for a paper's figures, with the options that reproduce them stored next to the result.
A free plan that fits a course, an open client, and a model anatomy that makes a trained network something you can point at.
A plain-English account of what a model relies on, and a record of each decision and its explanation, kept under your own retention rules.
Coming soon
What we are building now, and what to use until each one ships.
Token and passage attributions for text classifiers and language models, and for the score of a chosen answer from a generative model.
Inputs that arrive step by step or as a signal, attributed per time step, frequency band or frame.
Group fairness and disparate-impact measures as their own analysis, stored with its seed and options like every other result.
Gradient-based attributions computed by Kaimei for uploaded networks. Until then, store the ones you compute beside the built-in methods.
Gradient attacks such as FGSM and PGD beside the robustness analysis, which measures decision flips under random noise today.
The same calls from R, for statisticians and for teaching. Until then, the REST API works from any language.
Said plainly. Attributions describe the model, not the world: Kaimei does not establish cause and effect. Analyses use sampling, so set a seed when you need to reproduce a result exactly.
What works today, and everything on the way, is in Supported models and what’s coming.
Start on the free plan, or write to us about a course or a research project.