A Study Of Feature Selection For Multi-Label Classification

With the development of computer technology, multi-label classification has been used in actual application. For example, in the field of computer vision, multi-label learning can be applied to annotating image and video automatically; in the areas of biographical information, multi-label learning can be devoted to forecasting the function of gene; in the realm of text mining, multi-label learning can be employed in document classification. No matter how multi-label classification is used in which field, it will face the problem what is led to by the high dimensional feature spaces, such as, over-fitting. In general, features can be classified into three categories:’relevant’,’irrelevant’ and ’redundant’. Feature selection is aimed at looking for the best feature subset.Multi-label feature selection technology is dedicated to implementing feature selection technology into multi-label data. And like traditional feature selection, multi-label feature selection technology also can be sorted as filter methods, wrapper methods and embedded methods. In general, filter methods have two patterns. One is transforming multi-label data into single-label data and using traditional feature selection to solving problem. Another is improving old evaluation metric or creating new one, however, this ways are not many now. Although, there are some good feature selection algorithm can be used in multi-label classification currently, it is our duty to make the feature subset more and more perfect. Therefore, we do the following work according to the specific problem:(1) At present, common multi-label feature selection means lose sight of the fact, which there is a connection between samples. In fact, similar sample may have similar label set and any sample can be expressed by analogical example. In order to obtain better feature subset, we find the correlation by least square regression at fast, then calculating feature’s representative score by the correlation, and obtaining the feature ranking sequence by the value of representative score.(2) Wrapper methods rely on classifiers and search algorithms. Algorithms, being combined by some three independent algorithms, need enough time to keep good running. In addition, algorithms, being based on genetic algorithm, easily generate local optimal result due to premature phenomena. In order to further optimize the results, we introduce metropolis criterion and big variation to conventional genetic algorithm.(3) A kind of multi-label feature selection method, transforming multi-label data into single-label data before using evaluation criteria (such as mutual information, the F statistics, chi-square, and Relief) to measure relevance between single feature and individual label, ignore usually the correlation among labels. Notwithstanding wrapper methods and embedded methods can acquire better feature subset that can make classifier has a good showing, they have very high computation complexity. In order to consider the correlation among labels and computation complexity, we should distinguish the correlation between feature and label set and the sum of correlation between feature and individual label at prime tense, then using mutual information to measure the correlation between feature and label set and taking advantage of the correlation to gain the better result.