Many problems in machine learning have to deal with wide data – manymore features than observations. Most of the features are of no use,and even the useful ones are often too sparse. For these problems L1regularization and its variants have proven to be useful for both featureselection and complexity control. This talk is a review of a number oftopics in this area, with a focus on computational aspects.
This is joint work with Jerome Friedman, Rob Tibshirani, and our pastand present students.
- SMILE in Paris