Business Context
Datadog wants to improve adoption of a newly launched observability workflow. The growth team has labeled a subset of customer accounts that either adopted the workflow within 30 days or did not, while many newer accounts remain unlabeled. You need to compare a supervised approach for predicting adoption with an unsupervised approach for discovering account segments that may inform targeting.
Dataset
You are given an account-level dataset built from Datadog product usage over the first 14 days after signup.
| Feature Group | Count | Examples |
|---|
| Product usage | 12 | hosts_reporting, apm_services, logs_ingested_gb, dashboards_created |
| Team activity | 8 | active_users, monitors_created, alerts_acknowledged, notebooks_opened |
| Account metadata | 6 | company_size_band, cloud_provider, region, plan_tier |
| Support & sales signals | 4 | csm_touch_count, trial_days_remaining, demo_completed, ticket_count |
| Target label | 1 | adopted_workflow_30d |
- Size: 82K accounts, 30 input features
- Labeled subset: 49K accounts with known 30-day adoption outcome
- Unlabeled subset: 33K accounts with no target yet
- Class balance (labeled only): 28% adopted, 72% did not adopt
- Missing data: ~9% missing in support/sales fields, ~4% missing in some usage fields for low-activity accounts
Success Criteria
A strong solution should:
- Build a supervised model with ROC-AUC >= 0.82 and PR-AUC >= 0.58 on a held-out test set
- Produce an unsupervised segmentation with clusters that are stable and interpretable enough for GTM or product teams to act on
- Clearly explain when supervised learning is preferable, when unsupervised learning is preferable, and how the two can complement each other
Constraints
- Batch scoring must finish daily for ~100K accounts
- The growth team needs interpretable outputs, not just black-box scores
- The solution should work even when a large fraction of accounts are temporarily unlabeled
Deliverables
- Train a supervised model to predict 30-day workflow adoption on labeled accounts
- Build an unsupervised clustering pipeline on all accounts and profile the resulting segments
- Compare the two approaches: objective, data requirements, outputs, and evaluation methods
- Recommend how Datadog should use both models in production
- Provide metrics, feature importance, and cluster summaries
