[Paper Review] Causal Inference for Early Detection of Pathogenic Social Media Accounts

Hamidreza Alvari, Paulo Shakarian|arXiv (Cornell University)|Jun 26, 2018

Complex Network Analysis Techniques6 references3 citations

TL;DR

This paper proposes a causal inference framework for early detection of pathogenic social media (PSM) accounts using time-decay causality scores and community structure from user action logs, achieving an F1-score of 0.63 and 84% precision in detecting PSMs within 10 days of their activity, without relying on network structure, content, or cascade paths.

ABSTRACT

Pathogenic social media accounts such as terrorist supporters exploit communities of supporters for conducting attacks on social media. Early detection of PSM accounts is crucial as they are likely to be key users in making a harmful message "viral". This paper overviews my recent doctoral work on utilizing causal inference to identify PSM accounts within a short time frame around their activity. The proposed scheme (1) assigns time-decay causality scores to users, (2) applies a community detection-based algorithm to group of users sharing similar causality scores and finally (3) deploys a classification algorithm to classify accounts. Unlike existing techniques that require network structure, cascade path, or content, our scheme relies solely on action log of users.

Motivation & Objective

• Address the challenge of early detection of pathogenic social media (PSM) accounts that spread harmful misinformation before they go viral.
• Overcome limitations of existing methods that depend on network structure, content, or cascade paths, which are often unavailable or unreliable.
• Develop a causality-based detection framework that relies solely on user action logs to identify PSM accounts in real-time.
• Investigate whether causality scores are more concentrated within communities than across communities to improve classification accuracy.
• Enable timely intervention by detecting PSM accounts within a short time window around their initial activity.

Proposed method

• Compute time-decay causal metrics using a sliding window and exponential decay function to model dynamic causal influence over time.
• Apply the Kleinberg-Mishra causality metric to quantify user influence on message virality, adjusted for temporal decay via $\xi_{K\&M}^{I}(i) = \frac{1}{|\mathcal{T}|}\sum_{t'\in\mathcal{T}}e^{-\sigma(t-t')}\times\epsilon_{K\&M}(i)$.
• Construct a user-action graph from co-posting behavior, where users are connected if they post the same message chronologically.
• Use the Louvain algorithm to detect communities in the action-based graph to identify clusters of users with similar causality patterns.
• Perform a two-sample t-test to validate that within-community causality scores are significantly higher than between-community scores.
• Integrate causality scores and community structure into a classification pipeline for PSM detection, using supervised learning and proximity-based scoring.

Experimental results

Research questions

• RQ1 Can time-decay causal metrics effectively distinguish pathogenic social media (PSM) accounts from normal users based solely on action logs?
• RQ2 Is there a significant difference in causality score distribution between users within the same community versus across different communities?
• RQ3 Can community structure derived from action logs enhance the accuracy of PSM account detection compared to individual causality scores?
• RQ4 How timely and effective is the proposed framework in identifying PSM accounts within the first 10 days of their activity?
• RQ5 Does the integration of causality and community structure outperform baseline methods that rely on network structure or content analysis?

Key findings

• The time-decay causal metric alone achieved an F1-score of 0.63 in detecting PSM accounts using supervised learning.
• Supervised learning with decay-based causality features identified 71% of PSM accounts within the first 10 days of their activity.
• The community detection-based framework achieved a precision of 0.84 in identifying PSM accounts within 10 days of their initial activity.
• Misclassified accounts in the initial detection window were successfully detected within 10 days after the initial window, indicating strong temporal recovery capability.
• The two-sample t-test rejected the null hypothesis ($p$-value = 4.945e-17), confirming that causality scores are significantly higher within communities than across communities.
• The framework operates effectively without requiring network structure, content, or cascade path data, relying only on user action logs.