The Realization Of Label Distribution Learning Algorithm Based On Improved K-means And Adaboost

At present,in the field of machine learning,supervised learning still plays an important role.The main learning methods for supervised learning are single-label learning and multi-label learning.However,since single-label and multi-label learning can only represent the correspondence between the label and the instance,it cannot be well applied in an application where the relevant label is required to indicate the importance to the instance.Thus,label distribution learning has been proposed as a new machine learning paradigm.Under the label distribution learning paradigm,the label corresponding to the instance can represent the degree of description of the instance.The current implementation of label distribution learning algorithms mainly has three strategies:(1)problem transformation,(2)algorithm adaptation,and(3)specialization algorithm.In this paper,based on the second strategy to solve the problem of label distribution learning,the following two label distribution algorithms based algorithm adaptation are proposed.(1)Aiming at the possible relationship between the feature space of the sample instance and the label distribution space in the label distribution data,we propose to implement the label distribution learning by applying the k-means algorithm(KM-LDL).The algorithm takes into account adjacent samples in the feature space,and its corresponding label distribution should also be close in label space.Then,using the k-means algorithm,the samples of the training set are clustered,and then the label distribution set corresponding to each feature vector cluster is found.Finally,the label distribution of the test set is predicted by using the distance between the test set feature vector and each cluster mean vector of the training set as the weight.The algorithm and the existing three kinds of label distribution algorithms are tested on six public data sets.The experimental results show that the algorithm has achieved better results on all five evaluation indicators.(2)It can be observed that the traditional multi-label learning algorithm named BP neural network outputs a probability distribution result through the softmax function beforeoutputting the result,and the form of the label distribution is similar to the probability distribution.Therefore,the results of the probability distribution form obtained by the traditional neural network can be used as the predicted result of label distribution.On this basis,we further propose the BP neural network algorithm combined with AdaBoost to improve the prediction performance,and propose label distribution learning based on BP-Adaboost(LDL-BPAB).Furthermore,we introduce adaptive thresholds to improve LDL-BPAB,and propose label distribution learning based on BP-Adaboost with adaptive threshold improvement(LDL-IBPAB).Experiments on 8 public datasets show that the LDL-BPAB algorithm has achieved satisfactory results,and LDL-IBPAB has achieved better experimental results.