Platform · Explanations

Why this prediction, for this input.

Kaimei queries the model repeatedly and measures how far each feature moves the output from the population baseline. Below, one subscriber is picked out of the 1,761 a telecom churn model was tested on, and explained.

One prediction, explained

1,761 subscribers. One of them, explained.

These are the subscribers a telecom churn model was tested on, none of them seen in training. They regroup by who left, until one steps out and Kaimei shows why the model scored that account as it did.

  1. 01 · The held-out set

    1,761 subscribers.

    Every dot is one subscriber in the Telco Customer Churn data, IBM’s public sample of 7,043 telecom accounts. These are the quarter held out of training, so the model has never seen them. The ringed dot is row 2,292 of the file. Call it Subscriber A.

  2. 02 · The outcome

    467 churned.

    That is 26.5%; the other 1,294 stayed. For each subscriber the model sees ten account inputs, among them tenure, monthly charge, contract, internet service and payment method, and learns which of them separate the two colours.

  3. 03 · Contract

    The contract comes first.

    420 of 974 month-to-month subscribers churned: 43%. Of 787 on a one- or two-year contract, 47 did: 6%. A month-to-month subscriber in this set was about seven times as likely to leave.

  4. 04 · Internet service

    Then the service.

    Split again by internet service, 294 of 535 month-to-month fibre subscribers churned: 55%. Of 274 on a longer contract with no internet service, 2 did. Subscriber A is one of the 535.

  5. 05 · One subscriber

    Subscriber A.

    Four months in, on fibre, month to month at $85.65, paying by electronic check. This subscriber churned. The example’s model, 120 boosted trees of depth 3, puts the chance at 78.4%. The bright dots are the 166 month-to-month fibre subscribers with six months of tenure or less; 116 of them churned.

  6. 06 · The explanation

    Why 78.4%.

    Kaimei starts from the model’s average answer, 25.9%, and measures how far each input moves it. Fibre service, four months of tenure and a month-to-month contract account for 41.5 of the 52.5 points between the two. Two of the three are the splits the dots just made.

    KernelSHAP · exact over 58 reference rows · 59,428 model calls · points of P(churn)

  7. 07 · What if

    A two-year contract flips it.

    With tenure, dependents and the internet product held fixed, Kaimei’s what-if search looks for the smallest change that crosses the 50% line. It finds one: a two-year contract takes the same subscriber to 18.5%. That is a property of the model, not a promise about the customer. Analysed as a whole, the same model ranks tenure, contract and internet service first.

Each dot is a real row of the Telco Customer Churn data (IBM sample data, via OpenML): the 1,761 of its 7,043 subscribers held out of training. The model is the gradient-boosted classifier of the technology example, fitted on the other 5,282; on these 1,761 its AUC is 0.850. The average, the prediction, the attributions and the what-if are what the Kaimei API returned for row 2,292. Attributions are shown to one decimal, so the bars can add up a tenth or two away from the total. The order of dots within a group and the shades within each colour are for legibility only.

In practice

The same numbers in the dashboard, the API and your notebook.

Edit any input in the dashboard and re-explain, or call the API from your own code with the same result.

Exact

Every coalition, when it fits

Up to 12 inputs: every feature coalition over a fixed reference sample, so the same request returns the same numbers.

Units

In the output's own units

Attributions in percentage points of P(churn), megawatts of forecast load, whatever the model outputs.

Logged

Linked to its prediction

Every explanation is stored and linked to the prediction it explains, so a decision can be revisited later.

tabularKernelSHAPShapley values by weighted linear regression over feature coalitions. Exact for small executable models.
tabularPermutation ShapleyShapley values by sampling feature orderings, for larger inputs and remote models.
tabular · imagesLIMEA local surrogate fitted around the input. Works for any model, including a scoring endpoint.
imagesGrad-CAMReads the last convolutional feature map of an uploaded network, or the layer you choose.
imagesOcclusionHides each region in turn and measures the drop. Any image model, remote or uploaded.
tabularCounterfactualsThe smallest change that flips the decision, step by step, with chosen features frozen.

Coming soon

Coming soonText and tokensAttributions to words and passages for text classifiers and language models.
Coming soonSequencesModels that read a history step by step, attributed per time step as well as per input.
Coming soonIntegrated Gradients and DeepLIFTGradient attributions computed for uploaded networks. Store your own beside the built-in methods today.
Coming soonBoxes and masksObject detection and segmentation, explained per box or mask, not only per image label.
Coming soonSeveral outputsMulti-output and multi-horizon models, with every output explained in one call.
Coming soonAudio and videoAttributions per time span, frequency band or frame.

Upload a model this afternoon.

The free plan covers 10 models, 500 explanations and 20 analyses a month. Your own model is one call away.