Business Context

Microsoft Marketplace wants to predict whether an active seller will become inactive in the next 60 days so the account team can intervene early. You are given historical seller-level snapshots and need to design a preprocessing and feature engineering pipeline suitable for a production classification model.

Dataset

The training data is exported from Azure Data Lake and contains one row per seller per month.

• Rows: 240K seller-month records across 24 months
• Target: inactive_60d = 1 if the seller has no sales activity in the following 60 days
• Class balance: 11% positive, 89% negative
• Missing data: 18% missing in campaign features, 9% in support metrics, sparse values for new sellers

Success Criteria

A strong solution should produce a reproducible preprocessing pipeline, avoid temporal leakage, and improve over a simple one-hot + median-imputation baseline. Good enough means PR-AUC >= 0.42, ROC-AUC >= 0.80, and recall >= 0.70 at precision >= 0.35 on the holdout period.

Constraints

• Batch scoring must run nightly in Azure Machine Learning for ~35K active sellers
• The account team needs interpretable features and stable behavior over time
• Training data is temporal; random splits that leak future information are not acceptable

Deliverables

1. Define a preprocessing strategy for missing values, outliers, categorical encoding, and scaling
2. Engineer predictive features from raw monthly seller activity without leaking future data
3. Train at least one baseline and one stronger model, and justify the choice
4. Evaluate using time-based validation and business-relevant thresholds
5. Explain how the feature pipeline would be deployed and monitored in production