Using data science to find human trafficking fronts
- Article
- Machine Learning
Human trafficking is hard to see. Traffickers work behind ordinary storefronts and switch identities to avoid attention. Roughly 21 million people are trafficked worldwide each year, seven in ten of them women and girls, and fewer than 50,000 victims are identified and rescued. The trade brings in an estimated 50 billion US dollars (about Php 2.9 trillion) a year. Because the operations are hidden behind legal business names, the main tool against them has been more investigation and enforcement, which is slow.
This post, the first Data Bytes entry of 2025 from UP Data Science Society, was published on February 25, 2025 with content and layout by Victor Luis Ibanez. It summarizes an article by Booz Allen Hamilton on its work with Polaris, a nonprofit that runs the US National Human Trafficking Hotline.
Method
Polaris had noticed that illicit massage businesses in the United States were not isolated shops but parts of organized networks, and had mapped 35 connected businesses in one small city by hand. Booz Allen Hamilton’s data scientists set out to automate that.
Their main tool was a Naive Bayes classifier that predicts whether a massage business is illicit or legitimate from its registered business name. Naive Bayes is one of the simplest machine learning models: it applies Bayes’ theorem, which updates an initial belief when new evidence arrives. The infographic uses a pet example. If a box is equally likely to hold a cat or a dog, and you learn that the animal is quiet, and quiet is true of 80 percent of cats but 30 percent of dogs, the probability that it is a cat rises to 72.7 percent. The classifier does the same with the words in a business name. Alongside it, the team automated collection from public data sources and tracked usernames that reappeared on other social networks and forums to link businesses to each other.
Findings
The name classifier flagged 95 businesses as illicit in the first pass. Once the mapping was automated, the 35 businesses Polaris had connected by hand grew to 900, a roughly 25-fold increase. Several US states adopted the mapping tool for law enforcement, and Polaris used it to move from single cases to enterprise-level investigations.
Takeaway
The post’s closing point is that a model taught in an introductory class solved a problem that was costing lives, and that the same illegal activity can hide in plain sight behind a normal-looking business.
Source: Combating Human Trafficking Using Data Science, Booz Allen Hamilton