Bank Marketing Term Deposit Prediction (MLOps)

An end-to-end MLOps pipeline for predicting customer subscriptions to bank term deposits. This project covers data preprocessing, model training, containerization with Docker, and cloud deployment.

Project Overview

The objective of this project is to build a predictive model to identify customers most likely to subscribe to a term deposit. By targeting the right audience, a bank can optimize its marketing campaigns and increase conversion rates.

Model and Data Pipeline

The following steps were implemented in the model_training.ipynb to develop the inference pipeline:

1. Data Ingestion & Cleaning

• Source: The bank-additional.csv dataset featuring 41,188 records and 21 variables.
• Target Variable: The y column (yes/no) was mapped to binary values (1/0).
• Handling Imbalance: Since only ~11% of customers subscribe, the model uses class-weight balancing to ensure fair prediction.

2. Preprocessing & Feature Engineering

A scikit-learn ColumnTransformer was used to create a reproducible pipeline:

• Numerical Features: Features like age, duration, and economic indicators (euribor3m, nr.employed) were processed using StandardScaler to normalize their scales.
• Categorical Features: Columns such as job, marital, and education were transformed using OneHotEncoder to create dummy variables.
• Robustness: The encoder is configured with handle_unknown='ignore' to prevent API crashes when encountering new categories during live inference.

3. Model Summary:

Algorithm: Logistic Regression with class_weight='balanced'.

• Preprocessing: Data was processed using a ColumnTransformer that applied StandardScaler to numerical indicators (like euribor3m and nr.employed) and OneHotEncoder to categorical variables (like job and education).
• Justification: This model was chosen for its ability to handle the significant class imbalance in the bank marketing data and its low computational overhead, which is critical for real-time inference via a Dockerized Flask API.

4. Model Training

• Algorithm: Logistic Regression (Solver: liblinear).
• Optimization: Trained with class_weight='balanced' to improve sensitivity to the minority class (subscribers).
• Performance: Achieved an AUC-ROC of 0.94, indicating high discriminatory power between subscribers and non-subscribers.

Tech Stack

• Languages: Python 3.9
• Libraries: Pandas, Scikit-Learn, Flask, Joblib, Gunicorn
• DevOps: Docker
• Cloud: Azure / AWS (Linux EC2)

Project Structure

• model_training.ipynb: Exploratory Data Analysis and Model Training code.
• app.py: Flask API script for serving the model.
• final_model.pkl: The serialized ML pipeline (Preprocessor + Model).
• Dockerfile: Instructions to build the Docker container.
• requirements.txt: Python dependencies.

Deployment Instructions

1. Local Setup

# Clone and Install
git clone <your-repo-link>
cd <repo-folder>
pip install -r requirements.txt

# Build and Run Docker
docker build -t bank-marketing-model .
docker run -d -p 5000:5000 --name bank-predictor bank-marketing-model

2. Cloud Deployment

Provision a Linux VM on Azure/AWS.

Configure Security Groups to allow inbound traffic on Port 5000.

Install Docker on the VM, transfer the project files, and run the Docker commands above.

Testing the API

Replace <PUBLIC_IP> with your cloud instance's IP address.

Endpoint: http://<PUBLIC_IP>:5000/predict

Method: POST

Example Request:

curl -X POST -H "Content-Type: application/json" \
-d '{"age": 30, "job": "blue-collar", "marital": "married", "education": "basic.9y", "default": "no", "housing": "yes", "loan": "no", "contact": "cellular", "month": "may", "day_of_week": "fri", "duration": 487, "campaign": 2, "pdays": 999, "previous": 0, "poutcome": "nonexistent", "emp.var.rate": -1.8, "cons.price.idx": 92.893, "cons.conf.idx": -46.2, "euribor3m": 1.313, "nr.employed": 5099.1}' \
http://<PUBLIC_IP>:5000/predict

Expected Response: JSON

{ "prediction": 0, "probability_yes": 0.031885994107234805 }

---

Tips for your Video Demo:

When you record your video, be sure to highlight:

1. The Pipeline: Mention that you didn't just save a model, but a full pipeline that includes preprocessing.
2. Imbalance: Briefly mention using balanced class weights—this shows you understand the business problem of bank marketing.
3. Port 5000: Show the API responding live from your cloud IP address.