This repository contains the Master Thesis "Synthetically Extending a Dataset to Improve Machine Learning Prediction - An Application in Tabular Marketing Data" in .pdf format, together with the raw and un-cleaned Jupyter notebook files. Message me for any questions: https://www.linkedin.com/in/gwmdebruin/

ABSTRACT

Good machine learning (ML) algorithms start with good data to train on, but this data is not always available. A potential way to solve this problem is by generating synthetic data to complement a real training dataset. This approach has already been used to boost ML performance in for example the field of computer vision (Frid-Adar et al., 2018; Borkman et al., 2021), but the subject remains largely unexplored when it comes to its application in business and economics. Here, we consider a tabular, binary classification dataset describing customer responses to a marketing campaign. We investigate whether we can use a conditional tabular generative adversarial network (CTGAN) to create synthetic data that improves the performance of a random forest (RF) algorithm trained on a combination of real and synthetic data. Two approaches are considered: (i) an “internal” approach, where the real train set is complemented by data from a CTGAN generator that was trained on the real train set itself, and (ii) an “external” approach, where the largest part of the real train set is considered to be outside data which is only accessible to the CTGAN. In this approach only a small subset of the train set is directly accessible, and this set is combined with synthetic data generated by the CTGAN. We use both methods to train RF algorithms, and employ a bootstrap approach to get an estimate of the RF performance metrics. For approach (i), no significant increase or decrease in RF performance is observed. Approach (ii) does show a considerable increase in RF performance: when adding a synthetic dataset with 99k observations to a small real data set with 165 observations, the absolute increase in accuracy, precision, recall, F1 and AUC-ROC is 1.7%, 4.7%, 18%, 18.7% and 3.5%, respectively. The boosted performance is on par with using the full real training dataset for training the RF, indicating that real data can be replaced by synthetic data without compromising on the RF’s performance. Moreover, it is found that the increase in precision and recall has a strong effect on the profitability of the marketing campaign that we consider, leading to a profit increase of 99% for the particular case when adding synthetic data to the 165 real observations.