EBSVM: A Sparsification Method from LP-Based SVM via Stochastic Functional ξ-Analysis

While SVM being a solid mathematical and state-of-the-art sparse but powerful stochastic machine with top accuracy, Vapnik’s LP based SVM is sparser than the QP generated one though offers similar accuracy. Yet, further sparsity is essential to efficiently work with very complicated and large datasets. In our work, we propose a novel efficient method, EBSVM by imposing additional sparsification on this sparser LPSVM maintaining optimal complexity of the machine to further reduce classification cost while preserving similar accuracy by analyzing the stochastic auxiliary variable, ξ after a compact deterministic mapping of data. Experiments with Gunnar Rätsch benchmark data as well as synthetic data reveals that our model demands classification cost much smaller than the cutting-edge prominent classifiers while posing similar accuracy despite of being simpler to be implemented. On average, it costs 12.4% of the standard QPSVM, 50.3% of the VLPSVM, 68.1% of the EESVM [1], 98.1% of the greedy SpSVM-2 [2]. It insists the least MAC (Machine Accuracy Cost) value as well as GFR(Generalization Failure Rate) considering these machines. Numerical calculus based noise-sensitivity analysis proves our model’s potentiality in working with noisy data. This machine is also expected to work with high efficiency in case of very large and complex data satisfying constraints in real-time.