Pattern Classification Using a Penalized Likelihood Method

摘要

Penalized likelihood is a well-known theoretically justified approach that has recently attracted attention by the machine learning society. The objective function of the Penalized likelihood consists of the log likelihood of the data minus some term penalizing non-smooth solutions. Subsequently, maximizing this objective function would lead to some sort of trade-off between the faithfulness and the smoothness of the fit. There has been a lot of research to utilize penalized likelihood in regression, however, it is still to be thoroughly investigated in the pattern classification domain. We propose to use a penalty term based on the K- nearest neighbors and an iterative approach to estimate the posterior probabilities. In addition, instead of fixing the value of K for all pattern, we developed a variable K approach, where the number of neighbors can vary from one sample to another. The chosen value of K for a given testing sample is influenced by the K values of its surrounding training samples as well as the most successful K value of all training samples. Comparison with a number of well-known classification methods proved the potential of the proposed method.