Feature selection for pattern mining I must find frequent patterns in temporal data, using a method that was imposed to me. This tool has problems handling these data: processing is long and takes a lot of memory. So, I would like to reduce the number of attributes. Also, I want to keep working with the original attributes (no transformation or combination). In other words, I would like to perform a feature selection. By the way, the data contains both numerical and nominal (or categorical) attributes.
The first thing that comes to mind is to identify groups of correlated (or, more generally: associated) attributes and only keep one to represent each group. But I suppose more advanced, and possibly more efficient, methods exist. However, the only articles I can find on the Web are related to using pattern mining to perform feature selection, which is not what I want to do. Moreover, when I look up the feature selection methods described in the literature, they are designed for classification problems.
So, I was wondering if there are any works focusing on feature selection, specifically for pattern mining.
 A: Not really sure what kind of advanced method you are looking for which typically depends on how big and complicated your data set is(with overlapping class boundaries).If you have overFor selection or extraction one of the commonest method is to find the PCs (Principal Components) using PCA which I am sure you are aware of .Then after that you can remove the features with certain correlation threshold like 0.75 .
I suggest you to look for the book Applied Predictive Modeling by Max Kuhn and Kjell Johnson and you can use CARET package in R if you feel like 
A: The CARET package (and the book) suggest the concept of forward- and backward-feature-selection. This is a generic approach which should always work.
So basically you just start with 1 feature, add another feature and track if the result improves.
And the other way round you would start with all features, remove one and see if results get worse.
However this will not help you to save computation-resources, since this requires running the pattern-matching many times. But this can be done in a palatalized way. And it's a generic approach which respects your pattern-matching method.
