Improved Distance Functions for Instance-Based Text Classification

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Instance-based learning (IBL) is a high-performing method for text classification.
• The effectiveness of IBL heavily relies on the chosen distance function for document similarity.
• Existing distance measures may not fully exploit the nuances of text data, such as word frequencies and their ordinal relationships.

Purpose of the Study:

• To evaluate the performance of popular distance measures in text classification.
• To propose novel distance measures that leverage word frequencies and their ordinal properties.
• To compare the efficacy of proposed measures against existing methods and Naïve Bayesian classifiers.

Main Methods:

• Evaluation of popular distance measures for text classification.
• Development of new distance measures, including Value Distance Metric (VDM) and Inverted Specific-Class Distance Measure (ISCDM) variants.
• Comparison using the k-Nearest Neighbors (kNN) algorithm on 18 benchmark text classification datasets.
• Benchmarking against Naïve Bayesian classifiers (MNB, CNB, OVA).

Main Results:

• Nominal distance metrics yield better text classification results than Euclidean distance.
• The proposed ISCDM-based measures significantly outperform VDM-based measures.
• ISCDM effectively utilizes the ordinal nature of term frequencies, leading to more proposed ISCDM variants.
• kNN with proposed distance measures surpasses Naïve Bayesian classifiers on most datasets.

Conclusions:

• Novel distance measures enhance the performance of instance-based learning for text classification.
• ISCDM-based metrics offer a superior approach to document similarity compared to VDM and Euclidean distance.
• The proposed methods provide a competitive alternative to traditional Naïve Bayesian text classifiers.