ShopSphere AICV TabTransformer (v2)
Custom TabTransformer-style PyTorch model for churn prediction, trained on a synthetic ShopSphere e-commerce dataset for an academic Data-Driven Marketing Analytics project.
Dataset
Synthetic academic dataset: 12,000 observations, 1,000 customers, 12 months each, covering Historical / Update / Post-Update periods.
Baseline comparison
| Model | Accuracy | Precision | Recall | F1 | ROC-AUC | Threshold |
|---|---|---|---|---|---|---|
| Logistic Regression | 0.5094 | 0.1403 | 0.8176 | 0.2394 | 0.7347 | 0.35 |
| Gradient Boosting | 0.1450 | 0.0985 | 0.9882 | 0.1792 | 0.6824 | 0.02 |
| TabTransformer | 0.4456 | 0.1290 | 0.8471 | 0.2240 | 0.7325 | 0.17 |
Decision threshold
Selected on the validation set by maximizing expected net value under the assumption that a missed churner costs 25x more than an unnecessary retention outreach (₹5,000 lost value vs. ₹200 intervention cost) — not by F1.
Limitations
The dataset is synthetic and does not represent real ShopSphere customers. Results, including the baseline comparison, should not be interpreted as causal evidence that the platform algorithm update caused churn, and the hyperparameter search covered a small, compute-budget-limited grid rather than an exhaustive search.
Intended use
Academic demonstration of predictive marketing analytics and model hosting.