Updated: Apr 20, 2022
Written by Manya Malhotra
Edited by Missy Bridgwater and Shathviki Krishnaraj
On March 14th, 2022, Our Future Of Change Event Planning Directors Shathviki Krishnaraj and Shivani Dave had the opportunity to speak with Pia Ramchandani. Pia is a doctoral candidate in the Operations, Information, and Decisions Department at the Wharton School at UPenn. Recently, she partnered with the TellFinder Alliance, a network of counter-trafficking agencies, to work on algorithms that look at the risk of trafficking in commercial sex supply chains. Prior to doing her PhD, she served as a director in PricewaterhouseCoopers’ AI Accelerator lab and continues to work for their responsible AI team. She has an interest in data science for social good and ethical AI.
In this exclusive interview, Ramchandani said that she originally was not interested in Artificial Intelligence, but when working at PricewaterhouseCoopers, many problems needed data to be addressed and analyzed. She decided to pursue a PhD to apply the skills she learned in data science and solve problems in the social-impact space. She noted that she looked for socially important problems that could benefit from data analysis. By reading about the academic research on human trafficking, she was able to find work on analyzing deep web data to provide insights into patterns related to trafficking.
Ramchandani discusses that when analyzing data and web pages, her team looks for deceptive recruitment patterns used to find and exploit victims. For example, they may flag a webpage that is supposedly recruiting for modeling or secretarial services, yet is selling sex elsewhere or the linked opportunities involve sex work. With this example, the person applying is unknowingly applying to a job that involves sex, and this may sadly lead to sex trafficking. But, of course, this is just one example of deceptive recruitment. Ramchandani mentions being surprised by the number of types of deceptive trafficking recruitment there were. Working with TellFinder, a counter-trafficking alliance, they were able to uncover over 25 types of deceptive recruitment from a range of job advertisements. She agreed with saying that women and the LGBTQ+ community are often targeted when it comes to human trafficking and emphasized that it is usually the economically constrained communities that fall victim to deceptive recruitment.
From her research, she mentions that she studies data from adult services websites, yet the actual recruitment posts involve less than a percent of the activity on these websites. Hence, the data can be traumatizing for the researchers. Therefore, she made an effort through data optimization to avoid having someone review millions of posts that are really mentally and emotionally challenging. In an effort to improve the learning of the algorithm and optimize the program, she and her team designed an active learning process that adds a label to posts being reviewed. She mentions her research restructuring the active learning process. The network discovery objective made sure that the model was improving while trying to discover the criminal organizations' popular recruitment-to-sale trafficking routes.
Awareness of deceptive recruitment tactics is crucial to prevent the risk of trafficking. It is also important to point out the coordinates of trafficking routes on networks to law enforcement agencies. To implement a method called predictive model selection mentioned in her research, Ramchandani discusses testing out different model structures to figure out which worked best for her project. Feeding into the active learning process to find the network seemed to be the most important step.
The idea of prediction scores that she mentions in her research is to predict the likelihood that the post involves sex recruitment. This helps evaluate whether a post has deceptive recruitment, even if the post does not contain words like ‘sex work’. Having phone numbers or similar styles and links that lead to ads with sex work is a flag. The prediction scores are meant to inform the risk flags for some of these entities. She suggests that figuring out the best way to incorporate it into decision-making is still in process as her programs contain some biases: the data used from these deep web services are only in the English language, hence leaving bias towards English speakers. Her research is also only being applied to some parts of the world, so she mentions her intent and hope to expand to another extension. This will ensure that the process covers different locations and eventually focuses on Europe and India, in addition to the US.
Her advice toward those who want to pursue a career in technology with the intent of social awareness is to work with domain experts to become sensitive to some of the particularities of the domain and to also hone in on the objectives of actual on-the-ground field workers who are dealing with that work. Not being shy and reaching out to people to find a space for your skills is the way to go. Ultimately, the OFOC interview with Pia Ramchandani proves that there is much anti-trafficking work to be done in many fields, especially in STEM, so never give up on utilizing your strengths to help others and prevent this global crisis.
Learn more about Pia Ramchandani and her work by reading her article "Unmasking Human Trafficking Risk in Commercial Sex Supply Chains with Machine Learning," linked here: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3866259.