Fraud Detection in Crowdfunding Campaigns
On the show today, we are joined by Beatrice Perez. Beatrice is a postdoctoral researcher at Dartmouth College. She discusses her study titled I call BS: Fraud Detection in Crowdfunding Campaigns.
Machine learning has largely been used for bank fraud detection but finds sparse application in detecting fraudulent campaigns on crowdfunding platforms. Beatrice started by explaining the current process of detecting and taking down fraudulent campaigns by most host platforms - a mechanical process. To apply machine learning in this field, she had to first define what fraud in crowdfunding meant. For instance, a person could be truthful about the ailment of a loved one but would not remit the contributions made. This is a fraud but was outside the scope of her study.
She also explained the data collection process of retrieving properties of various crowdfunding campaigns. The data was cleaned to have only medical-related campaigns. Beatrice also discussed the rigorous process of data annotation. She explained the methods used to categorically tag a campaign as fraudulent or non-fraudulent. Some of the analyses she performed on the campaign description and comments include sentiment analysis, complexity and language choice, named entity recognition, etc. She detailed the emotional analysis carried out on images attached to campaigns. After performing these analyses, Beatrice was faced with over 5000 features and had to perform the KS test (Kolmogorov–Smirnov test) to reduce the features.
Going forward, she discussed the results from her model in terms of accuracy and AUC score. In addition, she explained how the model revealed the effect of images on the fraud classification. She also spoke about how to scale this research to other fields, perhaps building a general deception model.
Beatrice also discussed some constraints that prevent the models from being production-worthy. Rounding up, she gave some advice on how users can spot a potential fraudulent campaign on crowdfunding platforms. You can follow Beatrice and her co-authors on Twitter @bmpmila, @sara_rm84, and @kourtellis.
Beatrice Perez
Beatrice Perez is a Research Scientist in Boston, MA working with software defined radios and web information analysis. She’s worked in purely analytical projects (like the one presented in the podcast) and in projects that combine data analysis with hardware sensors and embedded systems. She’s passionate about privacy and the social aspect of digital technologies; things like identity, privacy, fraud, disinformation. She received her PhD in September 2020 from University College London (UCL) and since then she’s worked at Dartmouth College, the US Army Corps of Engineers Research Lab, and most recently Riverside Research.
Thanks to our sponsors for their support
