Abstract: Predictor screening rules, which discard predictors from the design matrix before fitting the model, have had sizable impacts on the speed at which sparse regression models, such as the lasso, can be solved in the high-dimensional regime. Current state-of-the-art methods, however, face difficulties when dealing with highly-correlated predictors, often becoming too conservative.
In this talk we introduce a new screening rule that deals with this issue: The Hessian Screening Rule, which offers considerable improvements in computational performance when fitting the lasso. These benefits result both from the screening rule itself, but also from much-improved warm starts.
The Hessian Screening Rule also presents a welcome improvement to the construction of the lasso path: the set of lasso models produced by varying the strength of the penalization. The default approach, to a priori construct a log-spaced penalty grid, often fails in approximating the true (exact) lasso path. Leaning on the information already used when computing the Hessian Screening Rule, however, we can improve upon the construction of this grid by adaptively picking penalty parameters along the path.
Organiser: Umberto Picchini (firstname.lastname@example.org). Please contact me if you wish to be informed of future seminars.
The seminar will take place in the room PASCAL and also on zoom (a password is needed: email email@example.com) https://chalmers.zoom.us/j/69071292832
Physical attendance in Pascal is very welcome, but only up to 30 people.
Pascal and Online