ShopSphere-AICV-TabTransformer

ShopSphere AICV — TabTransformer

Project

AICV (Algorithm-Impact Churn Variation) tracks sharp spikes in customer cancellations caused by an external platform algorithm shift that changes product visibility or pricing access.

Dataset

Synthetic academic dataset:

• 12,000 observations
• 1,000 customers
• 12 monthly observations per customer
• Historical, Update and Post-Update periods

Model

Custom TabTransformer-style neural network with:

• Categorical embeddings
• Transformer encoder blocks
• Numerical feature projection
• Binary classification head

Final T1 Test Metrics

• Accuracy: 0.7144
• Precision: 0.1940
• Recall: 0.6412
• F1: 0.2978
• ROC-AUC: 0.7414
• Decision threshold: 0.29

Model Configuration

• Embedding dimension: 32
• Attention heads: 4
• Transformer layers: 2
• Feed-forward dimension: 96
• Dropout: 0.15
• Learning rate: 0.0008
• Weight decay: 0.0002

Limitations

The dataset is synthetic and does not represent real ShopSphere customers. Results should not be interpreted as causal evidence that an algorithm update caused churn.

Intended Use

Academic demonstration of predictive marketing analytics, customer churn prediction, and Hugging Face model hosting.