[Paper Review] Community-based Outlier Detection for Edge-attributed Graphs

Supriya Pandhre, Manish Gupta+1 more|arXiv (Cornell University)|Dec 30, 2016

Anomaly Detection Techniques and Applications15 references3 citations

TL;DR

This paper proposes HCODA, a community-based outlier detection method for edge-attributed graphs that jointly models node and edge attributes using a Hidden Markov Random Field (HMRF). By treating outliers as a separate community and applying EM for parameter learning and community inference, HCODA effectively identifies holistic community outliers—nodes or edges that deviate in both structural and attribute contexts—outperforming baseline methods on synthetic and DBLP co-authorship data.

ABSTRACT

The study of networks has emerged in diverse disciplines as a means of analyzing complex relationship data. Beyond graph analysis tasks like graph query processing, link analysis, influence propagation, there has recently been some work in the area of outlier detection for information network data. Although various kinds of outliers have been studied for graph data, there is not much work on anomaly detection from edge-attributed graphs. In this paper, we introduce a method that detects novel outlier graph nodes by taking into account the node data and edge data simultaneously to detect anomalies. We model the problem as a community detection task, where outliers form a separate community. We propose a method that uses a probabilistic graph model (Hidden Markov Random Field) for joint modeling of nodes and edges in the network to compute Holistic Community Outliers (HCOutliers). Thus, our model presents a natural setting for heterogeneous graphs that have multiple edges/relationships between two nodes. EM (Expectation Maximization) is used to learn model parameters, and infer hidden community labels. Experimental results on synthetic datasets and the DBLP dataset show the effectiveness of our approach for finding novel outliers from networks.

Motivation & Objective

• Address the gap in outlier detection for edge-attributed graphs, where traditional methods ignore edge-level attributes despite their importance in real-world networks.
• Detect novel outliers that deviate not only in node attributes but also in the community structure of incident edges, which are missed by node- or link-only approaches.
• Model heterogeneous networks with multiple relationships between nodes by jointly inferring latent communities for both nodes and edges.
• Formulate outlier detection as a community detection problem where outliers form a distinct community, enabling holistic anomaly identification.
• Develop a probabilistic framework that integrates node data, edge attributes, and network structure to improve detection accuracy in complex, real-world graphs.

Proposed method

• Model the graph using a Hidden Markov Random Field (HMRF) that jointly captures node attributes, edge attributes, and network topology.
• Define a joint probability distribution over node and edge community labels, incorporating node data, edge attributes, and structural linkages.
• Use the Expectation-Maximization (EM) algorithm to iteratively estimate hidden community labels (via ICM) and learn model parameters, including weights for node, edge, and clique terms.
• Incorporate three key components in the HMRF potential function: node likelihood, edge likelihood, and triangular clique potential (to model local consistency in communities).
• Assign edge weights based on co-authorship frequency and normalize edge-to-node weights by inverse co-author count to reflect local connectivity.
• Set the number of communities K=4 (for four research areas) plus one outlier community, and use a 1% outlier ratio for parameter tuning.

Experimental results

Research questions

• RQ1 Can a joint probabilistic model that incorporates node attributes, edge attributes, and network structure detect outliers more effectively than methods that consider only nodes or links?
• RQ2 How does the inclusion of edge attributes improve the detection of holistic community outliers in heterogeneous networks?
• RQ3 To what extent does the proposed HMRF-based method outperform baseline community-based outlier detection (CODA) on synthetic and real-world edge-attributed graphs?
• RQ4 How sensitive is the method to hyperparameters like λ₁ (linkage weight), λ₂ (node weight), and λ₃ (triangular clique weight), and what are optimal settings?
• RQ5 Can the method identify meaningful, interpretable outliers in real-world co-authorship networks, such as interdisciplinary researchers or anomalous collaborations?

Key findings

• HCODA outperforms the baseline CODA method on synthetic datasets, demonstrating consistent improvement in outlier detection accuracy across varying outlier injection rates.
• On the DBLP Four Area dataset, HCODA successfully identified 1% of authors as outliers, with 100% of the top-ranked candidates validated as meaningful interdisciplinary researchers via manual inspection.
• The algorithm detected authors like Sameer K. Antani and Ivo Krka as outliers due to co-authoring across research areas inconsistent with their primary publication profiles.
• Anindya Datta and Debra E. VanderMeer were identified as outliers due to co-authoring on topics outside their primary research areas, confirming the method's ability to detect anomalous collaborations.
• Sabyasachi Saha was flagged due to active collaboration across artificial intelligence and information/knowledge management, reflecting the method’s sensitivity to multi-area engagement.
• The method showed robustness to parameter λ₂ and was most sensitive to λ₁ (linkage weight), with lower values yielding better performance, while λ₃ (clique weight) showed optimal performance for values >0.5.