Multi-instance Classification Based on Random Point Pattern

Abstract

Aiming at the problem that the performance of bag data classification based on Naive Bayesian model is limited by the similar feature distribution and lack of prior information. A method for multiple-instance classification based on random point pattern is proposed. It is based on random point pattern framework. Firstly, several random point pattern models with different complexity are constructed according to the training set. Then the maximum likelihood estimate (MLE) of model parameters is equivalent to the parameter estimate of the cardinality distribution and the feature distribution. The expectation-maximization (EM) algorithm is used to estimate the feature distribution parameters, and the EM algorithm parameters are initialized by the fuzzy c-means clustering (FCM) algorithm. Finally, the best model is given by calculating the bayesian information criterion (BIC) value of each model. Through two data set classification experiments, the results verify the improvement of the classification performance of the random point pattern model compared with the traditional model, as well as the robustness of the method.