Automatically correlating public campaign contribution data with criminal data to reveal high-risk links.

At Quantifind we believe in the power of combining two datasets to reveal surprising truths – truths which would not be obvious from either dataset alone. Similar to our nearby physicist friends at SLAC (the Stanford Linear Accelerator Center), we are driven by an innate curiosity to smash two things together just to see what comes out.

In our brand work, this involves correlating social chatter to private financial time series to discover consumer signals that are driving sales. In our financial crimes work, we link entities between external open source data (e.g., news feeds) and our client’s internal data to discover which transacting individuals have suspicious red flags.

We have many stories from these activities that we simply cannot share because one of these data sets is usually private (either confidential revenue or personal data). However, we can demonstrate our capabilities in certain scenarios where the two datasets both come from public sources. In this post, we will demonstrate our linking intelligence by colliding political campaign contribution data from the FEC (Federal Election Commission) with criminal indictment data from the DOJ (Department of Justice). We can then ask the question: Are there suspected criminals making contributions to a given political campaign?

There is an increasing number of financial data sets (e.g., blockchain, Venmo) that are at least partially open to the public. We chose the FEC contribution data set as analogous example of our usual work for the following reasons. Other than being timely and relevant for certain kinds of criminal activity (e.g., campaign fraud, “pay-to-play” schemes), it offers clean data on the identity of the contributors. Whereas the contributors are analagous to customers of a bank, the campaigns on the receiving end of the contributions (including PACs and individual candidates) are analagous the bank itself. Thus we can help the institution (campaign) determine if they are exposed in a way that may require risk-mitigation. For example, a discovered criminal link may cause them to return a contribution, or “exit” a relationship to be on the conservative side.

On the other side, we chose the public DOJ data set to play the “external” role, because it is a high quality and rich source of criminal entity information. This represents only a small part of our full external data options (which also includes news, forums, public social media, and many structured watch lists), but it serves as a useful a focal point due to its precision and potential relevance.

METHODOLOGY AND RESULTS

As opposed to the FEC data, the DOJ data does not come in a clean structured format. To enable our linking, we first parse these unstructured documents into entities (people and organizations) using Named Entity Recognition (NER), as well as extracting metadata that can assist in the eventual entity linking (e.g., age and location). This process essentially turns the DOJ data into a structured database that sets the table for the linking process. Our linking algorithms then attempt to discover specific individuals who exist in both data sets and score those links with a quantitative confidence level. In other words, a specific donor in the FEC data is automatically identified as the same person as a specific criminal in the DOJ data. Without getting into the details, the confidence level depends upon a number of factors, including name rarity, location proximity, and matching other pieces of metadata and context. By using modern machine learning methods, both the entity recognition and linking can be made significantly higher quality, more robust, and more up-to-date than traditional heuristic methods.

When combining the FEC data from the current election cycle (2017-18) with DOJ data spanning nearly a decade, we get several thousand candidate links. Limiting these to “high-confidence” links reduces this number to several hundred. Several of these high-confidence links are shown in the following table.