Improvements on ν-Twin Support Vector Machine.

Area of Science:

• Machine Learning
• Computational Intelligence
• Pattern Recognition

Background:

• ν-Twin Support Vector Machine (ν-TWSVM) is a binary classifier that determines two nonparallel hyperplanes.
• ν-TWSVM solves two related Quadratic Programming Problems (QPPs), which can be computationally intensive.

Purpose of the Study:

• To propose novel binary classifiers, Improvements on ν-Twin Support Vector Machine (Iν-TWSVM) and Iν-TWSVM (Fast), that improve upon ν-TWSVM.
• To enhance computational efficiency and maintain comparable generalization ability compared to ν-TWSVM.

Main Methods:

• Iν-TWSVM solves one smaller-sized QPP and one Unconstrained Minimization Problem (UMP).
• Iν-TWSVM (Fast) avoids a QPP by transforming it into a unimodal function solvable via line search, alongside solving a UMP.
• Iν-TWSVM incorporates structural risk minimization (SRM) by adding a regularization term.

Main Results:

• The proposed Iν-TWSVM and Iν-TWSVM (Fast) classifiers demonstrate faster performance than ν-TWSVM.
• The classifiers achieve generalization ability comparable to ν-TWSVM.
• Experiments on UCI and NDC datasets, along with color image pixel classification, validate the efficacy of Iν-TWSVM.

Conclusions:

• Iν-TWSVM and Iν-TWSVM (Fast) offer a more efficient alternative to ν-TWSVM for binary classification tasks.
• The novel formulations lead to reduced computational burden without sacrificing classification accuracy.
• The methods are applicable to diverse datasets and practical problems like image analysis.