LVQ algorithm with instance weighting for generation of prototype-based rules

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Mining

Background:

• Prototype-based rules offer comprehensible data representation, similar to crisp and fuzzy logic.
• Similarity-based methods, including Learning Vector Quantization (LVQ), are highly accurate data mining approaches.
• LVQ's performance is sensitive to prototype initialization and optimization.

Purpose of the Study:

• To enhance the accuracy and simplicity of LVQ models.
• To introduce novel methods for prototype initialization and LVQ cost function modification.

Main Methods:

• Developed context-dependent clustering for prototype initialization.
• Modified the LVQ cost function to incorporate class-dependent training vector distribution.
• Applied the enhanced LVQ approach to benchmark datasets.

Main Results:

• The proposed methods yield simple and accurate data models.
• Effectiveness demonstrated across various benchmark datasets.
• Prototype-based rules derived from the enhanced LVQ approach show strong performance.

Conclusions:

• The novel prototype initialization and cost function modifications significantly improve LVQ.
• This approach offers a powerful method for creating accurate, interpretable prototype-based rules.
• The findings contribute to advancing similarity-based data mining techniques.