Growth
Churn, conversion and lifetime value
Why this account is at risk, in terms a retention team can act on, and which change the model responds to most.
Solutions · Technology
For product and platform teams running churn, fraud, pricing and growth models. Explain a prediction from a support ticket, compare a retrain with the model it replaces, and get an alert that names the input that moved.
Where it fits
Growth
Why this account is at risk, in terms a retention team can act on, and which change the model responds to most.
Risk
Explain a held payment or a blocked signup from the model's structured inputs, and store the explanation with the case.
Fintech
The same reasons and records a bank needs, through an API your product team can call.
Operations
Numeric predictions explained in their own units, and watched for drift.
Platform
scikit-learn, XGBoost, LightGBM, CatBoost, Keras and PyTorch uploads, or any model behind an HTTP endpoint.
Release
Ship a retrain as a new version, score traffic in shadow, and compare what the two have learned before switching.
A worked example
Public dataTelco Customer Churn, IBM sample data, via OpenML (IBM).churn-risk-gbmscikit-learn · gradient boosting
A gradient-boosted model predicts whether a subscriber will cancel, from ten account inputs. It is trained on 5,282 subscribers and uploaded with one call; every figure below is what the Kaimei API returned.
POST https://api.kaimei.ai/v1/explanations
From a score to an action
Subscriber A is four months in, on fibre, paying month to month: 78.4% against an average of 25.9%. Kaimei searches for the smallest change that flips the decision, with tenure, dependents and the internet product held fixed.
It finds one: a two-year contract takes the model's answer to 18.5%. For Subscriber C the same change gives 20.0%.
That is a property of the model, not a promise about the customer. It tells the retention team which lever the model responds to, and an experiment tells them whether customers do.
from kaimei import Kaimei
client = Kaimei() # KAIMEI_API_KEY from the environment
client.models.upload(
model, # the fitted GradientBoostingClassifier
name="Subscriber churn", slug="churn-risk-gbm",
features=FEATURES, outputs=OUTPUTS,
X=X_train, y=y_train, # reference rows
)
explanation = client.explanations.create(
model="churn-risk-gbm", input=subscriber, method="shap",
options={"seed": 0, "counterfactual": True,
"frozen_features": ["tenure_months", "dependents", "internet"]},
)
print(explanation.output["value"]) # 0.7841
client.analyses.create(model="churn-risk-gbm", type="full")What it learned
tenure_months is the most influential input, followed by contract and internet. Together they carry 69% of the measured importance (permutation importance over 2000 rows).
2.3% of decisions flip under 5%-σ input noise. Most fragile inputs: monthly_charges, tenure_months.
tenure_months decreases the probability of churn, most sharply between 1 months and 5 months (range 12% → 46%).contract = month-to-month pushes the probability of churn highest and contract = two year lowest (range 8% → 30%).monthly_charges has a non-monotonic effect on the probability of churn: it peaks around 103.5 USD (range 23% → 35%).tenure_months is the most influential input, followed by contract and internet. Together they carry 69% of the measured importance (permutation importance over 2000 rows).On call
Five hundred held-out subscribers are logged as normal traffic, and every input is stable. Then a fibre promotion brings a wave of new sign-ups: 500 subscribers with six months or less on fibre, drawn from the same data. The batch averages a P(churn) of 75%, against 26% before. The alert does not say that the dashboard looks odd. It says tenure, internet product, monthly charge and contract have moved, in that order.
tenure_months, internet, monthly_charges, contract. 500 logged predictions against 5,000 reference rows.Population stability index per input. Bars are cut off at 0.5; the line marks 0.25, where Kaimei calls drift significant.
Who it is for
A Python client that reads the fitted estimator in your process, plain REST underneath, typed errors and retries. The same four calls for every framework.
Open a prediction in the dashboard, change an input and watch the attributions move. No notebook needed.
Answer why for a specific account from a stored explanation, by the id your systems already use.
Scope
We would rather say so here than have you find out in a trial.
| Kind of model | Supported |
|---|---|
| Yes/no, multi-class and numeric predictions on structured data | Yes, uploaded or through your endpoint |
| Image classification | Yes: Grad-CAM for uploaded CNNs; occlusion and LIME for any |
| Tree models with categorical splits, non-linear SVMs, nearest neighbours, models in other languages | Through a scoring endpoint, or from attributions you compute. Running them directly, and ONNX uploads, are coming soon |
| Language models, free text, generation | Coming soon. Until then, a text model's decision can be explained from structured features you derive |
| Recommendation and search ranking, detection, segmentation, audio, video | Coming soon. Image classification works today |
| Raw sequences and time series | Coming soon. Until then, a forecasting model on lags and rolling features works |
Start on the free plan with a model you already have. No sales call required.