Predicting Gender and Political Affiliation Using Mobile Payment Data

TL;DR

This study demonstrates that machine learning models can predict user gender with high accuracy and political affiliation with moderate accuracy based on Venmo transaction data, highlighting the potential of mobile payment data for social attribute inference.

Contribution

It is the first to apply latent attribute detection methods to mobile payment data for predicting social attributes like gender and political views.

Findings

Gender prediction accuracy of 91%.

Political affiliation prediction accuracy of 63%.

Utilized TF-IDF and SVM on Venmo data.

Abstract

We explore the understudied area of social payments to evaluate whether or not we can predict the gender and political affiliation of Venmo users based on the content of their Venmo transactions. Latent attribute detection has been successfully applied in the domain of studying social media. However, there remains a dearth of previous work using data other than Twitter. There is also a continued need for studies which explore mobile payments spaces like Venmo, which remain understudied due to the lack of data access. We hypothesize that using methods similar to latent attribute analysis with Twitter data, machine learning algorithms will be able to predict gender and political affiliation of Venmo users with a moderate degree of accuracy. We collected crowdsourced training data that correlates participants' political views with their public Venmo transaction history through the paid Prolific service. Additionally, we collected 21 million public Venmo transactions from recently active users to use for gender classification. We then ran the collected data through a TF-IDF vectorizer and used that to train a support vector machine (SVM). After hyperparameter training and additional feature engineering, we were able to predict user's gender with a high level of accuracy (.91) and had modest success predicting user's political orientation (.63).