Systemization of Knowledge (SoK): Creating a Research Agenda for Human-Centered Real-Time Risk Detection on Social Media Platforms

Title of the Paper

Systemization of Knowledge (SoK): Creating a Research Agenda for Human-Centered Real-Time Risk Detection on Social Media Platforms

Bibliographic Information

• Subject Areas: Human-Computer Interaction (HCI), Machine Learning (ML), Social Media Risk Detection
• Keywords: Online risks, human-centered machine learning, real-time risk detection, social media, literature review

Research Background and Issues

• Identified Problems or Challenges:

  1. Although real-time risk detection algorithms on social media platforms have made technical progress, they lack evaluation from a human-centered perspective.
  2. There is a trade-off between high accuracy and rapid detection.
  3. Insufficient consideration of the ecological validity of data sources and actual user behavior.

• Significance: Social media has become an integral part of daily life but also poses various risks (e.g., misinformation dissemination, mental health issues, cyberbullying). Timely detection and intervention are crucial for protecting users.

• Research Motivation and Related Work:

  • Current studies focus heavily on technical optimization rather than adaptability to human behavior.
  • Existing reviews have explored certain types of online risks (e.g., cyberbullying and mental health issues), but systematic research on real-time risk detection, particularly from a human-centered perspective, is still in its infancy.

Proposed Solution

• Proposed Methods or Solutions:
  This paper conducts a systematic literature review of academic work on real-time risk detection from 2015 to 2023, proposing a new research framework and addressing key gaps:

  1. Expanding the definition of "real-time" to include preemptive prevention, early post-event detection, and post-event harm mitigation.
  2. Synthesizing current trends in statistical models and deep learning techniques, and analyzing best practices.
  3. Identifying shortcomings in the human-centered perspective within current frameworks.

• Innovations:

  1. Expanded the technical framework to encompass multiple human-centered dimensions, from data collection to model development and evaluation.
  2. Provided recommendations for collecting ecologically valid datasets and optimizing conversational context capabilities.
  3. Proposed a systematic resource allocation strategy for quantifying risk severity.

• Implementation Steps and Key Techniques:

  • Literature Search: Selected 53 relevant papers from five major electronic databases (IEEE Xplore, ACM Digital Library, ScienceDirect, Springer-Link, and ACL Anthology).
  • Coding and Data Analysis: Developed a human-centered coding schema to analyze each paper across multiple dimensions, including data characteristics, model development, evaluation methods, and application studies.
  • Proposed Research Agenda: Offered comprehensive improvement suggestions for data, models, evaluation, and applications.

Research Outcomes

• Specific Findings:

  • Defined and expanded "real-time" to include preemptive prediction, post-event detection, and harm mitigation.
  • Identified technological trends such as adaptive data stream models, deep learning optimization, and domain-driven feature utilization.
  • Highlighted multiple shortcomings in the human-centered perspective of real-time risk detection research and proposed specific improvement directions.

• Comparative Advantages:

  1. A systematic framework covering the entire ecosystem of online risk detection.
  2. Emphasis on dataset ecological validity and dynamic risk detection based on user behavior.
  3. Guiding researchers to focus on human-centered evaluation and practical applications.

• Experimental or Evaluation Results:

  • Detection performance evaluations in all reviewed papers primarily focused on accuracy metrics (e.g., F1 score) and time sensitivity, with limited research addressing user behavior or real-world applicability.
  • A few papers attempted user evaluations but did not deeply explore how models impact actual user experience.

• Limitations and Future Directions:

  1. Lack of ecologically valid datasets targeting victims or specific populations.
  2. Data preprocessing overly relies on static data blocks, failing to dynamically simulate human behavior in real time.
  3. Most evaluation metrics are confined to technical performance, neglecting users' actual reactions and needs regarding risk detection.
  4. Insufficient collaboration between industry and academia, with limited open interfaces and real-world deployment applications.

  Future Recommendations:

  • Develop user-friendly data stream processing tools to enhance ecological validity.
  • Broadly adopt user behavior-based data features to optimize models.
  • Strengthen user involvement in evaluating algorithm performance, incorporating dynamic user feedback from the design phase.
  • Establish multi-stakeholder collaboration mechanisms to promote the development of industrial application interfaces and social value assessment for real-time online risk detection.