How to take advantage of behavioral features for the early detection of grooming in online conversations

• Ano de defesa: 2022
• Autor(a) principal: Flores, Daniela Fernanda Milon
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Behavioral features; Características comportamentais; Classificação com informação parcial; Classificação precoce de texto; Classification with partial information; Detecção de assédio sexual on-line; Early text classification; Online detection of grooming
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-01082022-150335/

Detecting grooming behavior in online conversations has become a growing problem due to the large number of messaging platforms that children and young people use nowadays. The biggest drawback is the lack of tools focused on the automatic prevention of this risk. This paper proposes seven Behavioral Features (BFs) to be used for early grooming detection. A detailed study is conducted to understand the background that allows these features to contribute to tasks of early classification. Besides, we introduce the Behavioral Feature - Profile Specific Representation (BF-PSR) framework as an extension of the well-known Profile Specific Representation (PSR) framework to properly employ the proposed behavioral features. Experimental results reveal that our proposal outperforms all the concurrent methods and obtains state-of-the-art performance in the area of early grooming detection. Specifically, the new BF-PSR framework achieves a gain of more than 40% in effectiveness over five competitors when only 10% of the SGD conversations content is available, thus it shows a substantial advantage to allow the early detection of grooming; besides, it maintains a similar gain in effectiveness as more data arrives. Furthermore, to the best of our knowledge, this is the first work to employ behavioral features for the early detection of grooming. On the other hand, we have assembled two new datasets called PJZ and PJZC to mitigate the lack of data in the grooming detection area. Both sets are publicly available for download aimed at fostering further researches. Additional experiments reveal that our BF-PSR framework outperforms all state-of-the-art methods when processing these new datasets.