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.
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 passenger is picked out of the 891 in the open Titanic example and explained.
One prediction, explained
891 passengers. One of them, explained.
This is the passenger list the Titanic example builds its model from. The passengers regroup by who survived, until one of them steps out and Kaimei shows why the model decided what it did.
-
01 · The list
891 passengers.
Every dot is one row of Kaggle’s Titanic training file, the list the example’s model is trained and validated on. The ringed dot is the first row: Mr Owen Harris Braund.
-
02 · The outcome
342 survived.
That is 38%; the other 549 died. For each passenger the model sees sex, age, ticket class, fare, port, family aboard and the title in the name, and learns which of them separate the two colours.
-
03 · Sex
Most women lived. Most men did not.
233 of 314 women survived: 74%. Of 577 men, 109 did: 19%. A woman on this list was almost four times as likely to survive.
-
04 · Class
Then class.
Split again by ticket class, 91 of 94 first-class women survived. In third class, 47 of 347 men did. Mr Braund is one of those 347.
-
05 · One passenger
Mr Owen Harris Braund.
22, third class, a £7.25 ticket from Southampton, travelling with his brother. He died. The example’s random forest, 30 trees at most 5 deep, gives him a 7.4% chance. The bright dots are the 319 third-class men titled Mr; 36 of them survived.
-
06 · The explanation
Why 7.4%.
Kaimei starts from the forest’s average answer, 41.1%, and measures how far each of his inputs moves it. Being a man (Sex, and Title = Mr, which says the same thing) and third class on a cheap ticket account for nearly all of the drop. They are the splits the dots just made.
KernelSHAP · exact over 58 reference rows · points of P(survived)
-
07 · What if
No ticket would have saved him.
Allowed to change only his class, fare and port, Kaimei’s what-if search finds nothing that flips the answer. Its closest, a £57.81 fare, lifts him to 33.9%, still short of the 50% line. Analysed as a whole, the same forest learned the story the dots tell: sex and title first, then class.
Each dot is a real row of data/train.csv in the Titanic example repository (Kaggle’s training file); the example’s forest is fitted on 80% of these rows and validated on the rest. The explanation, what-if and analysis figures are from its notebook 01. 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.
In the output's own units
Attributions in percentage points of P(survived), dollars of predicted spend, whatever the model outputs.
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.
Upload a model this afternoon.
The free plan covers the whole Titanic walkthrough with room to spare. Your own model is one call away.