L1/Lp regularization of Differences
Marcel A. J. van Gerven, Tom Heskes · 2008
In this paper, we introduce L1/Lp regularization of differences as a new regularization approach that can directly regularize models such as the naive Bayes classifier and (autoregressive) hidden Markov models. An algorithm is developed that selects values of the regularization parameter based on a derived stability condition. for the regularized naive Bayes classifier, we show that the method performs comparably to a filtering algorithm based on mutual information for eight datasets that have been selected from the UCI machine learning repository. 1