You can deploy this model and start earning money today!

ash001/
bank-churn-ann

tf-keras

churn

tabular

binary-classification

tabular-classification

apache-2.0

Views

No views yet

🏦 Bank Customer Churn Prediction (ANN | Keras)

A lightweight Keras ANN for bank customer churn prediction (Exited: 0 = stay, 1 = churn).
This Hugging Face repo hosts the trained model plus preprocessing artifacts used by a Streamlit inference app.

Training → Model → Inference

Training notebook (Colab): https://colab.research.google.com/drive/1ubzL_5BlJwnAVtqoko7IlZjYa0V8913r
Inference app (Streamlit): https://github.com/sparklerz/Deep-Learning-Fundamentals-Suite
(uses pages/01_Bank_Churn_Prediction.py)

What’s in this repo

artifacts/model.h5 (Keras model)
artifacts/scaler.pkl (StandardScaler)
artifacts/label_encoder_gender.pkl (LabelEncoder for Gender)
artifacts/onehot_encoder_geo.pkl (OneHotEncoder for Geography)
artifacts/schema.json (feature order + mappings)
Churn_Modelling.csv (dataset copy used in training)

Inputs (feature order)

The model expects 12 features in this exact order (see artifacts/schema.json):

CreditScore, Gender, Age, Tenure, Balance, NumOfProducts, HasCrCard, IsActiveMember, EstimatedSalary, Geography_France, Geography_Germany, Geography_Spain

Preprocessing (same as training/inference app):

Gender: label-encoded (Female: 0, Male: 1)
Geography: one-hot encoded into the 3 geography columns
All final features are scaled using StandardScaler (from scaler.pkl)

Output

Returns a churn probability in [0, 1].
schema.json stores a default decision threshold of 0.5. (The Streamlit demo app uses 0.35 by default to be more sensitive to churn.)

Quickstart (load artifacts)

python

import tensorflow as tf, pickle, json
from huggingface_hub import hf_hub_download

REPO_ID = "ash001/bank-churn-ann"

model = tf.keras.models.load_model(hf_hub_download(REPO_ID, "artifacts/model.h5"), compile=False)
scaler = pickle.load(open(hf_hub_download(REPO_ID, "artifacts/scaler.pkl"), "rb"))
le_gender = pickle.load(open(hf_hub_download(REPO_ID, "artifacts/label_encoder_gender.pkl"), "rb"))
ohe_geo = pickle.load(open(hf_hub_download(REPO_ID, "artifacts/onehot_encoder_geo.pkl"), "rb"))
schema = json.load(open(hf_hub_download(REPO_ID, "artifacts/schema.json"), "r"))

For an end-to-end preprocessing + prediction example, refer to the Streamlit inference file: pages/01_Bank_Churn_Prediction.py.

Notes / limitations

Intended for learning/demo; may not generalize to other banks/time periods without retraining and monitoring.
Contains .pkl artifacts—only unpickle files you trust.

ash001/bank-churn-ann