Using LIBSVM to predict authenticity of the user

Question

I am planning on using LibSVM to predict user authenticity in web applications. (1) Collect Data on particular user behavior(eg. LogIn time, IP Address, Country etc.) (2) Use Collected Data to train an SVM (3) Use real time data to compare and generate an output on level of authenticity

Can some one tell me how can I do such a thing with LibSVM? Can Weka be helpful in these types of problems?

Answer 1

The three steps you mention are an outline of the solution. In some more detail:

1. Make sure you get plenty of labeled data, i.e. behavior logs annotated with authentic/non-authentic. (Without labeled data, you get into the pretty advanced field of semisupervised learning, or must consider other solutions.)
2. Design a number of features based on the data that you think predict authenticity well. Try the method and refine it until it works well enough by some statistical standard. Use ten-fold cross validation to assure you're not overfitting.
3. LibSVM can output a probability estimate along with its answer; see section 8 of its manual.