Classify Retail Returns with PyTorch

Business Context

ShopNow, a mid-sized e-commerce marketplace processing 1.2M orders per month, wants to predict whether a newly placed order will be returned within 30 days. The operations team will use the score to prioritize manual review and optimize reverse-logistics planning.

Dataset

You are given an order-level tabular dataset built from the last 18 months of transactions.

Feature Group	Count	Examples
Customer history	8	prior_orders_90d, prior_return_rate, avg_order_value
Order attributes	10	item_count, total_amount, discount_pct, payment_method
Product signals	9	category, brand_tier, fragile_flag, size_variance_score
Fulfillment	6	shipping_speed, warehouse_region, delivery_delay_days
Behavioral	5	session_length, device_type, referral_channel

Size: 240K orders, 38 input features
Target: returned_30d — whether the order was returned within 30 days
Class balance: 14% positive, 86% negative
Missing data: ~12% missing in behavioral features and ~4% in fulfillment delay fields

Success Criteria

A good solution should achieve ROC-AUC >= 0.84, PR-AUC >= 0.50, and recall >= 0.70 at precision >= 0.45 on a held-out test set.

Constraints

Batch inference must score 1.2M orders nightly in under 10 minutes.
The model should be simple enough to retrain weekly.
The risk team needs feature attribution or at least a clear explanation of the modeling pipeline.

Deliverables

Build a PyTorch-based binary classification pipeline for tabular data.
Preprocess mixed numerical and categorical features, including missing values.
Train and validate the model with an appropriate split strategy.
Report business-relevant evaluation metrics and threshold selection.
Explain why PyTorch is appropriate here and what tradeoffs it introduces versus tree-based models.

Business Context

Dataset

You are given an order-level tabular dataset built from the last 18 months of transactions.

Feature Group	Count	Examples
Customer history	8	prior_orders_90d, prior_return_rate, avg_order_value
Order attributes	10	item_count, total_amount, discount_pct, payment_method
Product signals	9	category, brand_tier, fragile_flag, size_variance_score
Fulfillment	6	shipping_speed, warehouse_region, delivery_delay_days
Behavioral	5	session_length, device_type, referral_channel

Size: 240K orders, 38 input features
Target: returned_30d — whether the order was returned within 30 days
Class balance: 14% positive, 86% negative
Missing data: ~12% missing in behavioral features and ~4% in fulfillment delay fields

Success Criteria

A good solution should achieve ROC-AUC >= 0.84, PR-AUC >= 0.50, and recall >= 0.70 at precision >= 0.45 on a held-out test set.

Constraints

Batch inference must score 1.2M orders nightly in under 10 minutes.
The model should be simple enough to retrain weekly.
The risk team needs feature attribution or at least a clear explanation of the modeling pipeline.

Deliverables

Build a PyTorch-based binary classification pipeline for tabular data.
Preprocess mixed numerical and categorical features, including missing values.
Train and validate the model with an appropriate split strategy.
Report business-relevant evaluation metrics and threshold selection.
Explain why PyTorch is appropriate here and what tradeoffs it introduces versus tree-based models.

Business Context

Dataset

You are given an order-level tabular dataset built from the last 18 months of transactions.

Feature Group	Count	Examples
Customer history	8	prior_orders_90d, prior_return_rate, avg_order_value
Order attributes	10	item_count, total_amount, discount_pct, payment_method
Product signals	9	category, brand_tier, fragile_flag, size_variance_score
Fulfillment	6	shipping_speed, warehouse_region, delivery_delay_days
Behavioral	5	session_length, device_type, referral_channel

Size: 240K orders, 38 input features
Target: returned_30d — whether the order was returned within 30 days
Class balance: 14% positive, 86% negative
Missing data: ~12% missing in behavioral features and ~4% in fulfillment delay fields

Success Criteria

A good solution should achieve ROC-AUC >= 0.84, PR-AUC >= 0.50, and recall >= 0.70 at precision >= 0.45 on a held-out test set.

Constraints

Batch inference must score 1.2M orders nightly in under 10 minutes.
The model should be simple enough to retrain weekly.
The risk team needs feature attribution or at least a clear explanation of the modeling pipeline.

Deliverables

Build a PyTorch-based binary classification pipeline for tabular data.
Preprocess mixed numerical and categorical features, including missing values.
Train and validate the model with an appropriate split strategy.
Report business-relevant evaluation metrics and threshold selection.
Explain why PyTorch is appropriate here and what tradeoffs it introduces versus tree-based models.

Business Context

Dataset

You are given an order-level tabular dataset built from the last 18 months of transactions.

Feature Group	Count	Examples
Customer history	8	prior_orders_90d, prior_return_rate, avg_order_value
Order attributes	10	item_count, total_amount, discount_pct, payment_method
Product signals	9	category, brand_tier, fragile_flag, size_variance_score
Fulfillment	6	shipping_speed, warehouse_region, delivery_delay_days
Behavioral	5	session_length, device_type, referral_channel

Size: 240K orders, 38 input features
Target: returned_30d — whether the order was returned within 30 days
Class balance: 14% positive, 86% negative
Missing data: ~12% missing in behavioral features and ~4% in fulfillment delay fields

Success Criteria

A good solution should achieve ROC-AUC >= 0.84, PR-AUC >= 0.50, and recall >= 0.70 at precision >= 0.45 on a held-out test set.

Constraints

Batch inference must score 1.2M orders nightly in under 10 minutes.
The model should be simple enough to retrain weekly.
The risk team needs feature attribution or at least a clear explanation of the modeling pipeline.

Deliverables

Build a PyTorch-based binary classification pipeline for tabular data.
Preprocess mixed numerical and categorical features, including missing values.
Train and validate the model with an appropriate split strategy.
Report business-relevant evaluation metrics and threshold selection.
Explain why PyTorch is appropriate here and what tradeoffs it introduces versus tree-based models.

Interview Guides

Business Context

Dataset

Success Criteria

Constraints

Deliverables

Classify Retail Returns with PyTorch

Business Context

Dataset

Success Criteria

Constraints

Deliverables

Your Answer

Classify Retail Returns with PyTorch

Business Context

Dataset

Success Criteria

Constraints

Deliverables

Classify Retail Returns with PyTorch

Business Context

Dataset

Success Criteria

Constraints

Deliverables

Your Answer