Ομιλία Daniel Schmidt (Monash University, Australia)_ Σεμινάριο Τομέα Μαθηματικών 29/3/2024

Σας ανακοινώνουμε τη διάλεξη του  κ. Daniel Schmidt  Associate Professor of Computer Science at the Department of Data  Science and AI, Monash University, Australia,         (https://research.monash.edu/en/persons/daniel-schmidt).

 Η ομιλία θα πραγματοποιηθεί την Παρασκευή 29 Μαρτίου 2024 & ώρα 13:00 στην αίθουσα Σεμιναρίων του Τομέα  Μαθηματικών ΣΕΜΦΕ.

 

Τίτλος : Prevalidated ridge regression as a highly-efficient drop-in  replacement for logistic regression for high-dimensional data

 

Abstract: Linear models are widely used in classification and are  particularly effective for high-dimensional data where linear decision boundaries/separating hyperplanes are often effective for separating  classes, even for complex data. A recent example of a technique  effectively utilising linear classifiers is the ROCKET family of  classifiers for time series classification. One reason that the ROCKET  family is so fast is due to its use of a linear classifier based  around standard squared-error ridge regression. Fitting a linear model  based on squared-error is significantly faster and more stable than  fitting a standard regularised multinomial logistic regression based  on logarithmic-loss (i.e., regularised maximum likelihood), as in the  latter case the solutions can only be found via a numerical search.

 While fast, one drawback of using squared-error ridge-regression is  that it is unable to produce probabilistic predictions. I will  demonstrate some very recent work on how to use regular  ridge-regression to train L2-regularized multinomial logistic  regression models for very large numbers of features, including  choosing a suitable degree of regularization, with a time complexity  that is no greater than single ordinary least-squares fit. This in  contrast to logistic regression, which requires a full refit for every value of regularisation parameter considered, and every fold used for  cross-validation. Using our new approach allows for models based on  linear classifier technology to provide well calibrated probabilistic  predictions with minimal additional computational overhead. If time  permits, I will also discuss some thoughts on when such linear  classifiers would be expected to perform well.

Από την Επιτροπή Σεμιναρίων