How meghnakorim/americanexpress-user-churn-prediction-ann works
Project type and purpose
• This is a data-science/ML pipeline project that builds and evaluates an Artificial Neural Network (ANN) for user churn prediction using an American Express customer dataset.
• It covers the full workflow: data ingestion → preprocessing → model training → evaluation → single-user inference → conceptual 3D visualization.
File structure
Top-level files:
• AmericanExpress-Data-Analysis-for-User-Exit-Prediction.csv – raw data source
• churn_prediction.py – main Python script (pipeline entry point)
• requirements.txt – dependency list
• README.md – documentation and usage
There are no subdirectories; all logic lives in a single script. The project has a linear pipeline structure rather than a microservice or multi-tiered web app.
README for additional insights
• Data preprocessing uses pandas, NumPy, scikit-learn (label encoding, one-hot encoding, StandardScaler).
• Model built with TensorFlow/Keras: input layer, 2 hidden layers (ReLU), output layer (sigmoid).
• Evaluation via accuracy score and confusion matrix (Matplotlib).
• A conceptual 3D pipeline visualization is mentioned (likely Matplotlib/Plotly).
a. Main components:
1) Data Source (CSV file)
2) Data Ingestion & Preprocessing Module
3) Model Training Module
4) Model Artifact (trained weights)
5) Evaluation & Metrics Module
6) Single-User Prediction Component
7) Visualization Component (3D pipeline, confusion matrix plot)
b. Relationships and interactions:
• Data flows from CSV → preprocessing → training.
• Preprocessing outputs training and test sets → training module.
• Training module writes model artifact → evaluation module reads model + test set.
• Evaluation outputs metrics → directed to console and visualization component.
• Prediction component reads model artifact + new user data → outputs churn probability.
c. Architectural patterns and principles:
• Linear data pipeline (sequential stages).
• Modular separation: each stage is a logical block (even if implemented in one script).
• Reproducibility principle: requirements.txt pins dependencies.
d. Key technologies and libraries:
• Data handling: pandas, NumPy
• Preprocessing: scikit-learn (LabelEncoder, OneHotEncoder, StandardScaler)
• Model building: TensorFlow/Keras
• Evaluation/Visualization: Matplotlib (and potentially 3D lib)