Support Vector Machines. Similarity functions to work with heterogeneous data and classifying documents

Abstract

The objective of Data Mining (DM) is to classify information from the real world. That kind of information is commonly heterogeneous data: information that needs different kind of data to be represented. How to deal with heterogeneous data has been usually something DM lacks about because DM is not deeply used with real world problems. Different solutions has been shown and our objective is to show a new one using similarities and Support Vector Machines (SVM). How to use similarities instead of kernels in SVM and later how to combine similarities to work with heterogeneous data. The idea is that any type of data will have a similarity related and then all this similarities will be combined to output a result. What makes this idea powerful is the way we can combine similarities, it can be practically anything while other methods to work with heterogeneous data only do linear combinations.First of all understand how SVM works and what does it means to use similarities instead of Kernels. Later implement in a SVM library what explained before and show it working with an example. We will work with documents so it would be also required to do some NLP, learn about a NLP is another of my goals. Another of our goals is to use OO techniques and get a good design. Make our framework easy to be modified by anybody. Make an easy implementation. The objective is to extend the library used not to fork it.