KDD Cup: Using SVMs

Why SVMs?

1. Several studies (see entries in blog) say its a good technique

Software
So far, I have found only one
Its a two step process - learn on the training set and predict using the testing set.
The window version seems buggy - it ran much faster on my less powerful Linux box.

Data format

1. Treat each query as a document
2. Training set:
• ...
• What if there are several classes?
3. Testing set:

Open Issues:

1. Scalability - what's the max size of dimensions and data that it can process in a reasonable amount of time? (could not process 100,000 records and 500,000 features)
   
2. Data pre-processing
3. There are several kernel functions and associated parameters. By trial and error, need to determine which should be used.
4. Need to build a smaller subset of the data for experimenting