Multi-objective evolutionary optimization for dimensionality reduction of texts represented by synsets.

Area of Science:

• Natural Language Processing
• Machine Learning
• Information Retrieval

Background:

• Spam filtering faces challenges from linguistic complexities like polysemy and synonymy.
• Existing methods struggle with irrelevant or confusing words, impacting classification accuracy.
• Pre-processing is crucial to handle these linguistic issues before model building.

Purpose of the Study:

• To address limitations in spam filtering accuracy caused by linguistic phenomena.
• To introduce feature reduction as a multi-objective optimization problem.
• To compare different dimensionality reduction strategies for text representation in spam filtering.

Main Methods:

• Utilizing a Multi-Objective Evolutionary Algorithm (MOEA) to solve feature reduction.
• Implementing lossless (synset grouping), low-loss (synset grouping + irrelevant info removal), and lossy (irrelevant info removal) strategies.
• Experimentally comparing lossless and low-loss text representation schemes.

Main Results:

• The proposed MOEA framework effectively handles feature reduction as a multi-objective problem.
• The low-loss dimensionality reduction strategy demonstrated significant improvements in classifier efficiency.
• Experimental results validate the effectiveness of the low-loss approach over lossless methods.

Conclusions:

• Feature reduction framed as a multi-objective optimization problem offers a flexible approach to text classification.
• The low-loss strategy, balancing generalization and information relevance, is particularly effective for spam filtering.
• This work provides a novel method for enhancing spam filter performance through intelligent feature reduction.