14 January 2025
10:00 Doctoral defense IC3 Auditorium
Theme
Determining relevance of social-media posts for forensic event analysis
Student
José Dorivaldo Nascimento Souza Junior
Advisor / Teacher
Anderson de Rezende Rocha
Brief summary
When a large-scale forensic event occurs, posts related to the event are immediately shared on social media. This information can be potentially useful for later forensic inspection, since it can show different perspectives at different moments of the event. However, the analysis of social media data related to an event can be hampered by the large amount of irrelevant items retrieved during the collection process, such as memes and images from previous events. Manual sanitization of the dataset is unfeasible, since it can contain thousands of items. Therefore, we studied the use of machine learning techniques that use few labeled data to speed up this process and reduce the human effort required. Our work followed three paths. The first sought to provide a good representation of the posts, exploring pre-trained neural networks and feature fusion; the second aimed to perform classification with few samples, employing semi-supervised techniques ranging from graph-based methods to graph neural networks and the incorporation of data from previous events; and the last one aimed to add interactivity to the procedure, using instance selection and active learning techniques. We demonstrated through a series of experiments that following these directions improved the overall performance of the methods for this task.
Examination Board
Headlines:
Anderson de Rezende Rocha IC / UNICAMP
Cristina Nader Vasconcelos Google
João Paulo Papa FC / UNESP
Marcelo da Silva Reis IC / UNICAMP
Levy Boccato FEEC / UNICAMP
Substitutes:
Alexandre Mello Ferreira EEP
Marcos Medeiros Raimundo IC / UNICAMP
Jefersson Alex dos Santos Univ. Sheffield/United Kingdom
Moacir Antonelli Ponti ICMC / USP
William Robson Schwartz DCC / UFMG