Bank Customer Churn Prediction using ANN, TabNet, FT-Transformer and Autoencoder

[kapoorraaghav]

Pull Request for DL-Simplified 💡

Issue Title : #924

• Info about the related issue (Aim of the project) :
• Name: Raaghav Kapoor
• GitHub ID: kapoorraaghav
• Email ID: raaghavkapoor@gmail.com
• **Idenitfy yourself: GSSoC Participant

Closes: #issue number that will be closed through this PR

Describe the add-ons or changes you've made 📃

Give a clear description of what have you added or modifications made

Type of change ☑️

Added a complete end-to-end deep learning project for predicting bank customer churn. Here's what's included:

EDA: class distribution, feature distributions by churn status, correlation heatmap, churn rate by country and gender. The dataset has about a 20/80 class imbalance, which is handled using SMOTE.
ANN: 3-layer feedforward network (128, 64, 32) with BatchNorm and Dropout, used as a baseline. EarlyStopping and ReduceLROnPlateau callbacks are included.
TabNet: built from scratch in pure TensorFlow, with no external library. It uses sequential attention across 3 steps to select important features at each step.
FT-Transformer: a custom FeatureTokenizer layer embeds each scalar feature into a d-dimensional token. Then, 2 transformer blocks with MultiHeadAttention capture feature interactions.
Autoencoder + Classifier: an unsupervised pre-training stage where the encoder compresses input to 16 dimensions, followed by a supervised classifier trained on the compressed representations.
SHAP: KernelExplainer on the best-performing model for feature contribution analysis.
Keras Tuner: RandomSearch over ANN hyperparameters (hidden units, dropout rate, learning rate) with 10 trials.
Streamlit Web App: a dark-themed interactive demo that takes customer details as input. It runs all 4 models simultaneously and shows per-model churn probability, ensemble average, a risk badge, bar chart comparison, and model performance table.

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Code style update (formatting, local variables)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

How Has This Been Tested? ⚙️

Describe how it has been tested
Notebook was run end to end on the Churn Modelling dataset (10,000 rows) in a
conda environment with Python 3.9, TensorFlow 2.10, Keras 2.10.
All 4 models trained successfully with EarlyStopping — no overfitting observed.
SMOTE applied successfully — class balance confirmed after resampling.
Model weights saved and reloaded correctly — predictions verified on test set.
Streamlit app tested locally — all 4 models load and return predictions without errors.
ROC-AUC verified on held-out test set (80/20 stratified split) for all models.

Checklist: ☑️

[github-actions]

Our team will soon review your PR. Thanks @kapoorraaghav :)

[kapoorraaghav]

Hi @abhisheks008 any updates?