Uncovering the hidden trafficking trade with geographic data and natural language processing

Abstract

Human trafficking is a topic that is both underreported and under-prosecuted in comparison to other crimes. Often this is a result of being unable to detect the crime in the first place. To ease this detection problem, we present a data-driven map visualization and an evidence-based detection tool to improve human trafficking detection. For our tool, we hypothesized that a machine learning model, using natural language processing (NLP), could analyze text to identify socioeconomic patterns in trafficking to detect if trafficking is present. To identify potential patterns, we mapped gross domestic product (GDP) and trafficking information on separate maps. We then statistically drew connections between GDP per capita and reported trafficking rates around the world and in the US, and found a negative relationship between the two variables in the world. We created the detection tool using a logistic regression model on a manually compiled dataset to identify trafficking instances from qualitative data. In making a detection tool, we aimed to draw clearer distinctions between trafficking and other crimes or events. We anticipate that this distinction may lead to more human trafficking reports and prosecutions. Our final trafficking detection tool predicted labor trafficking cases in reports and interviews with 94% accuracy. Our f1 score, another measure of accuracy, was also 94%. We did not find evidence that the model explicitly used socioeconomic patterns to detect trafficking cases, but our analysis suggested that such patterns may have helped the model make predictions.