| In recent years,with the increasing amount of data,it has become very difficult to obtain datasets with all labels in practical applications.Therefore,feature selection using semi-supervised data has become a hot topic in the field of machine learning and pattern recognition research.As an important branch of feature selection,online stream feature selection has been well used in fields such as data mining,which processes the features coming with time in real time to filter out irrelevant and redundant features.Firstly,this thesis proposes a semi-supervised multi-label online streaming feature selection algorithm OFS-Mean for the current state of online streaming feature selection algorithm which is difficult to apply to semi-supervised data for feature selection.The algorithm accomplishes the prediction of missing labels by analyzing the labels in similar instances,using neighborhood rough set theory to calculate the dependency between the features and the predicted labels.The distance between instances and neighbors is used to design an adaptive neighborhood relationship,which enables the algorithm to select a suitable number of neighbors for each instance.A rough set-based feature evaluation method is used to develop a search strategy to complete online feature selection,while the importance of each feature in the feature set is evaluated to achieve the purpose of online redundancy update.To prove the effectiveness of the algorithm in real applications,the OFS-Mean algorithm is applied to a real application scenario of news topic text classification,compared with the comparison algorithm in each evaluation metric,which verifies the generality of the algorithm through experimental data.Secondly,the OFS-Mean algorithm is experimented on 12 publicly available datasets and compared with five traditional multi-label feature selection algorithms on Hamming Loss,Ranking Loss,One-error,Coverage,and Average Precision evaluation metrics to verify the validity.The experimental results are plotted as radar plots and CD plots to check the stability and significant differences of the algorithms,respectively.Similarity tests are designed to verify the feasibility of the algorithm by calculating the similarity between instances and their neighbors in each dataset.The parameters applied in the neighborhood relationship designed by the algorithm are discussed,which describe the reasons for choosing the values to reduce the impact of parameter changes on the algorithm.Finally,to prove the effectiveness of the algorithm in real applications,OFS-Mean algorithm is applied to real application scenarios of news topic text classification and coronary heart disease symptom classification,which are compared with the comparison algorithm in each evaluation metric,validating the generality of the algorithm through experimental data. |
