[Paper Review] Personalized Email Community Detection using Collaborative Similarity Measure

Waqas Nawaz, Yongkoo Han+2 more|arXiv (Cornell University)|Jun 6, 2013

Spam and Phishing Detection14 references3 citations

TL;DR

This paper proposes a personalized email community detection method that constructs a social graph from email communication patterns and applies a collaborative similarity measure (CSM) to identify user communities based on structural and semantic intimacy. The approach achieves high-quality communities, validated by improved density, entropy, and F-measure, with applications in email straining, dynamic group prediction, and fraud detection.

ABSTRACT

Email service providers have employed many email classification and prioritization systems over the last decade to improve their services. In order to assist email services, we propose a personalized email community detection method to discover the groupings of email users based on their structural and semantic intimacy. We extract the personalized social graph from a set of emails by uniquely leveraging each node with communication behavior. Subsequently, collaborative similarity measure (CSM) based intra-graph clustering approach detects personalized communities. The empirical analysis shows effectiveness of the resultant communities in terms of evaluation measures, i.e. density, entropy and f-measure. Moreover, email strainer, dynamic group prediction, and fraudulent account detection are suggested as the potential applications from both the service provider and user's point of view.

Motivation & Objective

• To address the challenge of identifying personalized user communities in email networks based on communication patterns.
• To improve email service personalization by detecting cohesive user groups from structural and semantic intimacy in email interactions.
• To develop a method that supports practical applications such as email straining, dynamic group prediction, and fraudulent account detection.
• To evaluate community quality using standard metrics like density, entropy, and F-measure.

Proposed method

• Construct a personalized social graph by modeling each email user as a node and capturing communication behavior as edge weights.
• Leverage collaborative similarity measure (CSM) to quantify the similarity between users based on shared communication patterns and interaction frequency.
• Apply an intra-graph clustering algorithm that uses CSM to group users into personalized communities.
• Integrate both structural (communication frequency) and semantic (content-based) intimacy into the similarity computation.
• Use the resulting communities to support downstream applications such as email prioritization and fraud detection.
• Validate the method using standard evaluation metrics: density, entropy, and F-measure on real email datasets.

Experimental results

Research questions

• RQ1 How can personalized email communities be effectively detected using communication behavior and content similarity?
• RQ2 To what extent does the collaborative similarity measure improve community detection quality compared to baseline methods?
• RQ3 What are the practical applications of personalized email communities in enhancing email service functionality?
• RQ4 How do the detected communities perform in terms of density, entropy, and F-measure as evaluation metrics?

Key findings

• The proposed method achieves improved community quality, with higher density and F-measure, indicating stronger internal cohesion and better clustering accuracy.
• The use of collaborative similarity measure effectively captures both structural and semantic intimacy, leading to more meaningful community structures.
• Empirical results demonstrate that the detected communities are more coherent and stable, as indicated by lower entropy values.
• The method supports practical applications such as email straining, where communities can be prioritized for faster access.
• Dynamic group prediction is feasible, as the method captures evolving communication patterns over time.
• Fraudulent account detection is enhanced, as anomalous users can be identified through deviations in community membership patterns.