Evidence for the decision.

Temporal landmarking; calibrated boosting; discrete-time hazards

Build member snapshots before expected subscription expiry and reproduce the official churn definition. Compare an elastic-net logistic baseline with gradient boosting and, where complete renewal episodes permit, a discrete-time hazard model. Exclude transactions or usage recorded after each prediction cutoff.

Evaluation: Use month-based holdouts with label-maturation gaps. Report log loss, calibration, precision/recall at fixed contact capacity and performance by tenure and plan. Compare several warning horizons and remove payment/usage feature families in ablations.

Boundary: KKBox is a consumer music service, not a B2B account-revenue dataset. Saved subscribers and retention ROI require an intervention experiment; observed churn prediction cannot establish them.

KKBox Churn Prediction Challenge ↗

Relational inductive bias; temporal heterogeneous graphs; boosting

Select a documented rel-stack engagement task and preserve its official prediction horizon and temporal splits. Compare a transparent SQL aggregation plus boosted-tree baseline with a heterogeneous graph model. Restrict every neighbor, edge and aggregate to information available at the task reference time.

Evaluation: Report the official benchmark metric plus probability calibration, resource usage and performance on users with sparse history. Ablate relation types and compare against identical-budget SQL baselines. Audit neighborhood timestamps to expose graph leakage.

Boundary: Q&A engagement is a product-usage proxy. It does not establish paid SaaS revenue, conversion or the causal effect of a customer-success action.

RelBench rel-stack ↗

Survival analysis; cohort effects; hierarchical shrinkage

Construct repository and contributor cohorts from documented public event types. Define a qualifying contribution and a future return window in advance; exclude obvious automation in a sensitivity analysis. Model time to the next qualifying contribution with survival models and repository-level partial pooling.

Evaluation: Train on earlier cohorts and evaluate later cohorts, with a repository-held-out stress test. Compare against stars, recent event count and simple recency. Report return-risk calibration, time-dependent prediction error and sensitivity to bot rules and archive coverage.

Boundary: Public GitHub activity excludes private work and paid usage. It cannot establish software sales, organization-wide productivity or the causal effect of community initiatives.

GH Archive ↗