Anomaly Detection in Networks | SNA Lecture Part 2

Added:

Contextual Subjectivity and Algorithmic Bias in Network Anomaly Detection

A significant critique of network anomaly detection, particularly in social network analysis (SNA), is the highly subjective nature of defining what constitutes 'normal' behavior. Critics argue that establishing statistical baselines often marginalizes non-conformist, minority, or emergent communities by misclassifying their unique interaction patterns as 'anomalies' or threats. This can result in systemic algorithmic bias, censorship, and disproportionate surveillance of vulnerable groups. Furthermore, social networks are inherently fluid and context-dependent; a connection or behavior deemed anomalous in one context may be entirely benign and adaptive in another. Opponents of purely algorithmic detection suggest that reductionist mathematical models (whether feature-based or community-centric) strip away essential qualitative context, leading to high false-positive rates and failing to capture the complex, evolving nature of human social dynamics.

Basic Graph Theory: Understanding of nodes, edges, adjacency matrices, and fundamental graph representations.
Social Network Analysis (SNA) Metrics: Familiarity with centrality measures (degree, betweenness, closeness) and clustering coefficients.
Introduction to Community Detection: Knowledge of how cohesive subgroups or communities are identified in graphs (e.g., modularity, clustering).
Foundational Machine Learning: Understanding basic outlier/anomaly detection concepts and unsupervised learning techniques.
Graph Neural Networks (GNNs) for Anomaly Detection: Leveraging deep learning architectures like GCNs and Graph Autoencoders for complex pattern recognition.
Dynamic and Temporal Graph Mining: Exploring advanced algorithms that handle continuously evolving networks and time-series graph data.
Scalable Graph Analytics: Implementing anomaly detection on massive, real-world graphs using big data frameworks like Apache Spark (GraphX) or Neo4j.
Domain-Specific Applications: Applying network anomaly detection to real-world scenarios such as financial fraud detection, cybersecurity intrusion detection, and social media bot identification.
1.1K views15likes1:06:04@lablcs2Original Release: 2022-04-12

This lecture covers two primary approaches for detecting anomalies in static graphs: (1) Feature-based methods that use handcrafted features such as node-level metrics (eigenvector centrality, betweenness, clustering coefficient), dyadic features (reciprocity, common neighbors, Adamic-Adar index), ego-network properties, and global network properties, which are fed into classifiers to identify anomalous nodes or edges; and (2) Pattern-matching approaches like ODDBALL that define specific types of anomalies (near-cliques, near-stars, heavy vicinities, dominant links) and use feature dependencies (number of nodes vs. edges, total weight vs. edges, principal eigenvalue vs. total weight) to detect deviations from normal patterns in ego-networks.