What to do with many super sparse features, in a classification problem?

I am facing a classification problem. There might be 50K observations and 1M factor (only in 0 or 1)features, while those features are all sparse, with little 1 and many 0 (the proportion of 1 is almost under 5% for any feature). I am wondering what can I do besides SVD and PCA(if possible)? Is feature selection possible? Do I have to consider association rules? Thanks a lot.