Research On Gravitation-based Multi-label Classification

Multi-label classification is one of important models in machine learning,which has been used in various real-world applications,especially for image annotation.This is because different people will tag different labels to the same image,which reflects their various culture background and experience.The research on multi-label classification is hence valuable and significant in both theory and practice.As a classic classification algorithm,k-nearest neighbor method is widely used in the field of the multi-label classification and image annotation.However,existing algorithms are supposed to handle the nearest-neighbors through the local features or labels.The lack of the direct links among data samples makes the loss of interpretability and intuitiveness for k-nearest neighbor method.Therefore,this dissertation focuses on the study of multi-label classification approach based on gravitation for multi-label learning problem.Main contributions of the dissertation are as follows:(1)The relative researches show that there exists the connection between the similarity of multi-label data and the gravitation of objects.Hence,the proposed algorithm introduces Newton's law of gravitation to multi-label nearest-neighbor classification as the gravitation computation.The gravitation computation is able to explore the correlation of data samples that do not only belong to the geometrical feature neighborhood.The special correlation can avoid the loss of information concerning the imbalance and the sparsity of the multi-label model.Besides,the gravitation computation establishes the direct links between the data samples.Due to the direct links,the proposed method selects k largest gravitation data samples as a temporary cluster.Based on the temporary cluster,the proposed method gives the final classification results by gravitation weight.Experiments conducted on the various scales or fields of the datasets validate the effectiveness and adaptability of the proposed method in multi-label model.(2)The nearest neighbor type model is a basic model for image annotation.However,the nearest neighbor type model misses the direct and visual correlations between the images.The proposed method which has advantage on classification task is able to resolve this trouble.The images in the temporary cluster which selected by the proposed method exist the semantic similarity and the feature resemblance between each other.Based on the temporary cluster,the proposed method can improve the accuracy and the credibility for image annotation.Furthermore,overcoming the semantic gap between the available image features and the keywords that people might to use annotate.