Fraud Detection on the Elliptic Bitcoin Dataset: Graphs, Leakage, and Topology
Details
Pyowa returns to Source Allies on Tuesday, October 27th from 5:30–7pm with Dave Keller presenting a hands-on look at fraud detection, machine learning, and the importance of questioning results that seem a little too good.
In this session, Dave will explore his work using the Elliptic Bitcoin transaction dataset to investigate fraud detection. When his initial models produced near-perfect performance, it raised an important question: was the model actually learning meaningful patterns, or was data leakage influencing the results?
Dave will walk through rebuilding the experiment with a temporal train/validation/test split and establishing logistic regression and Random Forest baselines. He’ll also cover how statistical techniques including Spearman correlation, mutual information, and Kolmogorov-Smirnov tests helped investigate relationships between features and time.
Rather than chasing the highest possible model score, this talk is about questioning results, testing alternative explanations, and understanding what the evidence does and does not support.
Hosted at Source Allies. Food and drinks will be provided.
Speaker Bio: Dave is a court researcher and aspiring data scientist with a background in mathematics. He is currently pursuing the MITx MicroMasters in Statistics and Data Science and develops projects focused on machine learning, fraud detection, graph analytics, and Python.