Fuzzy Support Vector Machines IEEE Transactions on Neural

Fuzzy Support Vector Machines IEEE Transactions on Neural Networks, 2002 Authors: Chun-Fu Lin and Sheng-De Presentation by Zhuang Wang

Outline n n n Introduction SVMs vs Fuzzy SVMs Experiments My figures Drawbacks of the paper

Introduction n Motivation: In many applications (eg. evaluation of credit risk), different data points give different contribution to the decision surface. n How? Treat each point differently. (Give each point a weight or fuzzy membership. )

SVMs vs. FSVMs n Traditional SVMs: To solve the optimal hyperplane problem: (treat each point equally)

SVMs vs. FSVMs (cont. ) n Fuzzy SVMs: (treat each point differently) Difference: each data point is presented like this: (Xi, Yi, si ), where si is a fuzzy membership between [0, 1], New Problem is:

SVMs vs. FSVMs (cont. ) n The optimal problem is different, but the solution is very similar. (only one difference) After reformulation, the problem can be transformed into:

Experiments n Data with time property Assign fuzzy membership according to the time data arrive in the system. n Two class with different weighting Select fuzzy membership as a function of respective class. n Use Class Center to Reduce the Effects of Outliers Assign fuzzy membership according to the distance to class center.

My figures

Figures (cont. )

After decreasing weight 10 times

Drawbacks of the paper n Only toy datasets, no reallife datasets are used in experimental part. n The way to assign fuzzy membership to data points need to be improved.

Thank you! Questions?