Devdit/Datathon_TabTransformer

🧠 TabTransformer Multitask Model for Churn, Tenure, and LTV Prediction

This model is a multitask TabTransformer implemented in PyTorch, designed to perform:

• Binary classification for customer churn
• Regression for customer tenure
• Regression for customer LTV (Lifetime Value)

It is saved as a pickle file: model.pkl and includes all custom layers (e.g., positional encoding).

---

🧩 Model Architecture

• Tabular input with:
  • x_num: Numerical features (projected into latent space)
  • x_cat: Categorical features (embedded + transformer)
• Transformer-based attention over categorical embeddings
• Multi-head output for multitask predictions:
  • Churn: Sigmoid activation for binary classification
  • Tenure and LTV: Linear regression heads

---

🧪 How to Use

1. Install Dependencies

pip install torch pandas


import torch

# Load the model from pickle
with open("model.pkl", "rb") as f:
    model = torch.load(f)

model.eval()


# Example dummy input
x_num = torch.rand((1, 10))          # Replace 10 with your actual num_features
x_cat = torch.randint(0, 5, (1, 3))  # Replace with your actual number of categories

# Predict
churn_prob, predicted_tenure, predicted_ltv = model(x_num, x_cat)

print("Churn probability:", churn_prob.item())
print("Predicted tenure:", predicted_tenure.item())
print("Predicted LTV:", predicted_ltv.item())


(
    churn_prob: FloatTensor of shape (B, 1),
    predicted_tenure: FloatTensor of shape (B, 1),
    predicted_ltv: FloatTensor of shape (B, 1)
)