DUB Seminar will be conducted using Zoom, via an invitation distributed to the DUB mailing list. Participants who are logged into Zoom using a UW account will be directly admitted, and participants who are not logged in to a UW account will be admitted using a Zoom waiting room.
From detecting online harassment to the sexual predation of youth, the state-of-the-art computational risk detection approaches has the potential to protect particularly vulnerable populations from online victimization. Overall, my research contributes to the field of adolescent online safety and machine learning by taking a human-centered approach to truly understand the problem context and determine a meaningful ground truth of the risky behaviors being detected through automated approaches. Starting by conducting a systematic literature review of computational approaches for online sexual risk detection, we found that most studies focused on detecting sexual grooming using datasets that are not reflective of the real-world interaction of users. We conducted a qualitative analysis of adolescents’ posts and comments on an online peer support mental health forum to understand how adolescents seek support about their online sexual interactions. Then, we designed and implemented a data donation web tool for collecting social media data from teens to create ecologically valid training datasets for risk detection machine learning applications. Lastly, we developed human-centered machine learning models which can detect risky conversations using this dataset so that timely interventions can be administered. Completing a feature analysis, we found that contextual features and linguistic features contributed the most to accurately detecting risky sexual conversations. In this talk, I will discuss why it is crucial to train models using ecological valid datasets and consider victims’ perspectives of risks on such sensitive topics. I will also give an overview of my research projects and will discuss more about the complexity and the sensitive nature of private datasets, and how preserving the confidentiality and privacy of the participants is important. I plan to create a research program that integrates how AI could be used for the social good of people for mitigating online difficulties of vulnerable populations.
Afsaneh Razi is an assistant professor at the Department of Information Science at Drexel University’s College of Computing & Informatics (CCI). Her research expertise is positioned at the intersection of Human-Computer Interaction (HCI) and machine learning (ML) to solve sociotechnical issues. Her work strives to deeply understand societal issues and identify ways to meditate these challenges using technology. Specifically, one of her active research areas addresses the critical and timely problem of online safety by leveraging a multi-disciplinary approach of human-centered machine learning to accurately detect risks vulnerable users encounter online. She has a record of research publications from ACM’s premiere Conferences on Human Factors in Computing Systems (CHI) and Computer-Supported Cooperative Work (CSCW), as well as the Conference of the Association for the Advancement of Artificial Intelligence (AAAI). She also has experience working on fast-paced product projects using mix-methods in industry as a User Experience Researcher.