[Paper Review] Improving the methods of email classification based on words ontology

Foruzan Kiamarzpour, Rouhollah Dianat+2 more|arXiv (Cornell University)|Oct 22, 2013

Spam and Phishing Detection10 references3 citations

TL;DR

This paper proposes a novel email spam classification method that enhances accuracy by integrating word ontology with ensemble decision trees. By leveraging semantic relationships from an ontology to refine feature selection and combining multiple decision trees via voting, the approach achieves improved spam detection performance over traditional keyword-based methods.

ABSTRACT

The Internet has dramatically changed the relationship among people and their relationships with others people and made the valuable information available for the users. Email is the service, which the Internet provides today for its own users; this service has attracted most of the users' attention due to the low cost. Along with the numerous benefits of Email, one of the weaknesses of this service is that the number of received emails is continually being enhanced, thus the ways are needed to automatically filter these disturbing letters. Most of these filters utilize a combination of several techniques such as the Black or white List, using the keywords and so on in order to identify the spam more accurately In this paper, we introduce a new method to classify the spam. We are seeking to increase the accuracy of Email classification by combining the output of several decision trees and the concept of ontology.

Motivation & Objective

• To address the growing challenge of email spam due to increasing volume and complexity.
• To improve classification accuracy beyond traditional keyword-based filtering methods.
• To integrate semantic knowledge from word ontology into machine learning models for better feature representation.
• To develop a hybrid approach combining decision trees and ontology reasoning for robust spam detection.

Proposed method

• Construct a domain-specific word ontology to capture semantic relationships among email terms.
• Use the ontology to enrich feature extraction, transforming raw keywords into semantically meaningful attributes.
• Train multiple decision trees on the ontology-enhanced feature set to improve generalization.
• Combine the outputs of individual decision trees using a voting mechanism to produce a final classification.
• Apply semantic similarity measures from the ontology to resolve ambiguity and reduce false positives.
• Optimize the ensemble model using cross-validation to balance precision and recall.

Experimental results

Research questions

• RQ1 Can integrating word ontology into email classification improve detection accuracy compared to traditional keyword methods?
• RQ2 How does semantic enrichment of features affect the performance of decision tree models in spam filtering?
• RQ3 To what extent does combining multiple decision trees enhance robustness and reduce overfitting in spam detection?
• RQ4 What is the impact of ontology-based feature selection on reducing false positives in email classification?

Key findings

• The proposed method achieved a significant improvement in classification accuracy compared to baseline keyword-based approaches.
• Incorporating word ontology reduced false positive rates by enhancing semantic understanding of email content.
• The ensemble of decision trees outperformed single-tree models, demonstrating better generalization on unseen data.
• Semantic feature enrichment led to more consistent performance across diverse spam categories.