Subscriber renewal classifier — three-class churn tree
A trained classification tree predicting whether a subscriber renews, downgrades or cancels — split rules, sample counts, per-class counts and Gini impurity on every node, seven leaves.
For the data scientist explaining a churn model
Scenario
A classification tree that predicts what a subscriber does at renewal: renews, downgrades to a cheaper plan, or cancels. The features are plain business facts — contract type, tenure, support calls, monthly charge and data use — so the model can be explained to people who do not build models. The tree is three levels deep, with six splits and seven leaves. The counts are illustrative but add up consistently.
Annotation key
decisiontree:ml— machine-learning tree modeclasses: renews, downgrades, cancels— the class order used by everyvalue=[...]listimpurity: gini— the impurity measure shown on each nodesplit "Tenure <= 12 months"— a split node; the quoted text is its rulesamples=,value=[...],gini=— how many subscribers reach the node, how many fall in each class, and how mixed the node istrue/false— which branch of the parent split the node sits onleaf ... class=renews— a prediction; the class is the majority in that leaf
How to read
Start at the top and follow the branch where the rule holds (true) or fails (false). The first question is the contract: month-to-month subscribers go down the true branch, where tenure and support calls or monthly charge separate them further. The largest leaf is annual-contract subscribers paying $90 or less — 1,900 of them, 90% renewing. Some leaves are weak predictions: the "downgrades" leaf for longer-tenure month-to-month subscribers paying more than $70 holds only 42% downgraders. Reading the class counts, not just the label, shows how confident each prediction is.