I have some QSAR models which are based on SVM, DT, MLP and Ensemble. After preprocessing the networks are trained to predict the IC50 of some drugs. I cant figure it out why the performance of SVM is so weak in this regression problem?
If you need to train a good model using SVM, I suggest the parameters of SVM that don't use the default. Like RBF SVM, both gamma and c parameters are the vital rule to influence results. Moreover, there are many kernel functions, including linear, polynomial, RBF, sigmoid, ..... You need to understand the advantages and disadvantages of these kernel functions, and then to determine which kernel function could be used to fit your problem.
In summary, the kernel function selection and their parameters could be the key to improve your performance of SVM.
Thank you very much for your answer. Actually I used 4 algorithms and SVM performance was worst. I need to know the reason to be able to explain it in my manuscript.
If the other algorithms have the good results, I consider the SVM performance should be not worst because these algorithms indicate that the certain functions are really existed to classify the class. This is unlikely to be caused by insufficient SVM capabilities. Therefore, I suggest to try to turn SVM parameters. If you don't understand how to setting the optimal SVM parameters, the optimization algorithm could be help you to find the optimal combination of parameters, including particle swarm optimization, genetic algorithm, and so on.