Predicting BTD Purchases from Marketing Campaign Data

Project Goal

Can we create a data processing pipeline and machine learning classifier that will predict whether a bank client will purchase a subscription to a bank term deposit (BTD)?

Data source

This project makes use of the UCI bank marketing data set to train and evaluate the classifiers defined in this repository.

Running the experiments

Installation

1. Setup Docker for desktop
2. Pull the docker image for this repo: docker pull pauldhein/telemarket-model:latest

Dataset preprocessing

This experiment processes a CSV dataset into train/dev/test folds after first converting the label column to binary and all string-valued columns to categorical data. Categorical data columns are also one-hot encoded to aid in training. The script can be run with the following command:

docker run -it pauldhein/telemarket-model:latest python dataset_processor.py -p ../data/bank-full.csv

The -p flag specifies the full path of the CSV file to be processed.

Model training

This experiment performs a grid search over specified hyperparameter sets for five different classification models. Extra processing is done during the experiment to:

1. Account for class imbalance in the labels
2. Select the top 20 features to improve classification using a mutual information statistic
3. Standardize the feature values using min-max standardization

The experiment can be run using the following command:

docker run -it pauldhein/telemarket-model:latest python model_trainer.py -p ../data/bank-full

The -p flag provides the prefix-path of the dataset to be used for training and evaulation.

Model evaluation

This experiment verifies the performance of the various saved classifier configurations on either development data or test data. The experiment can be run using the following command:

docker run -it pauldhein/telemarket-model:latest python model_evaluator.py -t -d

The -t and -d flags specify to evaulate on the test set and the development set respectively.

Experimental Results

Evaluation metrics for the test set

Model Name	Accuracy	Precision	Recall	F1	AUC-ROC
Logistic Regression	0.8623	0.4324	0.8677	0.5772	0.8647
Random Forest	0.9053	0.5870	0.4238	0.4922	0.6938
Naive Bayes	0.8371	0.3616	0.6592	0.4670	0.7589
Linear SVM	0.2765	0.0501	0.3161	0.0865	0.2939
Neural Network	0.8555	0.4242	0.9350	0.5836	0.8904

ROC curve for the test set