In recent years, support vector machine has become one of the most important classification techniques in pattern recognition, machine learning, and data mining due to its superior classification effect and solid theoretical base.
However, its training time will increase dramatically as the number of samples increases, and training will become more sophisticated when dealing with problems involving multiple classifications. A quick training data reduction approach MOIS appropriate for multi-classification tasks is presented as a solution for the aforementioned issues. While eliminating redundant training samples, the boundary samples that play a vital role are chosen in order to considerably reduce training data and the problem of unequal distribution between categories.
The experimental results demonstrate that MOIS may maintain or even improve the classification performance of support vector machines while substantially enhancing training efficiency. On the Opt digit dataset, the suggested method improves classification accuracy from 98.94% to 99.05%, while training time is reduced to 15% of the original; in HCL2000, the proposed method improves classification accuracy from 98.94% to 99.05%. When the accuracy rate is marginally increased (from 99.29% to 99.30%) on the first 100 categories dataset, the training time is dramatically reduced to less than 6% of the original. Additionally, MOIS has a high operational efficiency.
0 comments:
Post a Comment