Sliding into My DMs: Detecting Uncomfortable or Unsafe Sexual Risk Experiences within Instagram Dire

Link: https://scholar.google.com/citations?view_op=view_citation&hl=en&user=sV43f5AAAAAJ&sortby=pubdate&citation_for_view=sV43f5AAAAAJ:hFOr9nPyWt4C

Journal: Proceedings of the ACM on Human-Computer Interaction (PACMHCI'24) Keywords: Online Safety, Youth Online Risks, Machine Learning, Deep Learning

My notes

Summary

Use CNN and ML to detected sexual risks in insta DM conversations and train a classifier to identify the risky level. Measurement, qualitative analysis, and case studies on results. Related Technique: ML, DL, NLP

Pros

1. Focus on a minor group lacking attention

2. Carefully selected data

3. Comprehensive benchmark comparison experiments

4. Nice detailed and comprehensive qualitative analysis and case studies

Cons

1. It might be better if language models can be set in comparison experiments

2. Other cons are mentioned in the "limitation and future work" section

Details

Abstract

• RQ1: Based on the first-person accounts of youth, what attributes can help us best predict whether sexual risk is present within a private social media conversation?

• RQ2: a) Can we accurately predict if a given message is sexually risky? b) If so, can we assess its risk severity level?

• RQ3: a) How are the contextual, linguistic, and semantic features most predictive of sexual risk inform our understanding of the sexual risk behaviors of youth online? b) What are the most common reasons for misclassifications?

Data

15,547 instagram DM conversations from 150 filtered adolescents PAN12 as a baseline

Findings

1. Overall, we observed that all traditional classifiers had the best performance with the combination of features (LIWC, Sentiment, TF-IDF, and Sexual Lexicons) plus contextual factors (age, gender, and relationship), compared to having features separately or the combination of features without the contextual features

2. Both yes. For binary classification, CNN and RF performed better in conversation-level. For risky level, the CNN classifier outperformed SVM, RF, and LR.

3. Refer to case studies for details.

Conclusion

The first work that analyzes machine learning approaches on private social media conversations of youth to detect unsafe sexual conversations. In addition, this work highlights the importance of contextual and implicit features on identifications of unsafe sexual conversations and provides a good indication of how different methods and features perform when addressing this problem.