Feature selection for ordinal text classification.

Area of Science:

• Machine Learning
• Natural Language Processing
• Data Mining

Background:

• Ordinal classification, or ordinal regression, is crucial for rating data on discrete scales.
• It's increasingly important in sentiment analysis and opinion mining for product review data.
• Feature selection is vital for efficiency and preventing overfitting in classification tasks.

Purpose of the Study:

• To address the lack of feature selection methods for ordinal classification.
• To propose and evaluate novel feature selection techniques specifically for this task.
• To compare these new methods against existing approaches using real-world product review data.

Main Methods:

• Developed six novel feature selection methods tailored for ordinal classification.
• Tested these methods on two product review datasets.
• Compared performance against three established literature methods using support vector regression algorithms.

Main Results:

• All six proposed feature selection metrics significantly outperformed baseline techniques.
• The novel methods demonstrated superior stability, being an order of magnitude more stable.
• Consistent outperformance was observed across both datasets and learning algorithms.

Conclusions:

• The proposed feature selection methods are highly effective for ordinal classification tasks.
• These methods offer substantial improvements in efficiency and stability for sentiment analysis.
• This work advances the field of ordinal classification by providing robust feature selection strategies.