[Paper Review] In-Network Neighborhood-Based Node Similarity Measure: A Unified Parametric Model

Yang Yu, Jian Pei|arXiv (Cornell University)|Oct 13, 2015

Complex Network Analysis Techniques27 references3 citations

TL;DR

This paper proposes a unified parametric model for in-network node similarity based on neighborhood features, enabling flexible, application-specific similarity measures. It introduces four instance measures—using maximal common neighborhoods, neighborhood patterns, random walks, and k-hop neighbors—demonstrating their effectiveness in real-world and synthetic networks through extensive experiments.

ABSTRACT

In many applications, we need to measure similarity between nodes in a large network based on features of their neighborhoods. Although in-network node similarity based on proximity has been well investigated, surprisingly, measuring in-network node similarity based on neighborhoods remains a largely untouched problem in literature. One grand challenge is that in different applications we may need different measurements that manifest different meanings of similarity. In this paper, we investigate the problem in a principled and systematic manner. We develop a unified parametric model and a series of four instance measures. Those instance similarity measures not only address a spectrum of various meanings of similarity, but also present a series of tradeoffs between computational cost and strictness of matching between neighborhoods of nodes being compared. By extensive experiments and case studies, we demonstrate the effectiveness of the proposed model and its instances.

Motivation & Objective

• Address the lack of systematic methods for measuring node similarity based on neighborhood structure in large networks.
• Develop a flexible, parametric framework that supports various similarity meanings by plugging in different neighborhood features.
• Provide a spectrum of similarity measures balancing computational cost and topological matching strictness.
• Ensure theoretical properties such as metric transformation and normalization are preserved across instances.
• Validate the model’s effectiveness through extensive experiments on real-world and synthetic networks.

Proposed method

• Propose a unified parametric model that generalizes neighborhood-based similarity by allowing customizable features and aggregation functions.
• Define four instance similarity measures: SimMCN (maximal common neighborhoods), SimNP (neighborhood patterns), SimLW (labeled random walks), and SimKN (k-hop neighbors).
• Use a feature set S of neighborhood patterns with parameters τ (pattern frequency threshold) and n (pattern length) to define SimNP.
• Apply normalization and metric transformation techniques to ensure similarity measures are scalable and mathematically sound.
• Leverage spectral clustering on the similarity matrix to evaluate structural role detection without full topological access.
• Use synthetic graphs with planted anomalies to evaluate rank promotion and robustness of similarity measures.

Experimental results

Research questions

• RQ1 How can we systematically model node similarity based on neighborhood structure rather than proximity or relative proximity?
• RQ2 What are the key features of neighborhoods that can effectively capture different application-specific meanings of similarity?
• RQ3 How do different neighborhood-based similarity measures trade off between computational cost and topological matching accuracy?
• RQ4 Can neighborhood-based similarity measures be transformed into normalized similarities or metrics to support downstream applications?
• RQ5 How effective are these measures in real-world tasks such as anomaly detection and node clustering?

Key findings

• The proposed model successfully captures diverse meanings of similarity through four distinct instance measures, each tailored to different application needs.
• All four similarity measures can be normalized and transformed into valid metrics, ensuring mathematical robustness.
• SimNP, based on neighborhood patterns, effectively partitions nodes into structural roles in the NetSci co-author network, identifying hubs, outliers, and mainstream nodes.
• In anomaly detection tasks, SimLW and SimKN significantly promote the ranks of planted anomaly nodes, with average rank promotion of 12.5 and 14.3 respectively.
• The model demonstrates strong performance on both synthetic and real-world networks, showing resilience and adaptability across diverse network structures.
• The method enables effective node clustering without full topological access, relying only on pairwise similarity matrices derived from neighborhood features.