A Research Of Feature Selection And Feature Learning

With the arrival of information age,the learning data tends to be of larger scale,higher dimensions,and with complex noises,which gives rise to challenges maching learning research.As a consequence,it is of great value and significance to fully analyze the data and extract the key features and potential information.In this dissertation,we discuss the reseach status and related works of feature selection and feature learning respectively.Feature selection aims at selecting an optimal subset from the original feature set by a certain strategy.Existing feature selection algorithms mainly evaluate the impor-tance of features by considering their relevance to target tasks.In supervised learning,the statistical correlation between features and target is considered.In unsupervised learning,the discrimination ability of features is evaluated according to the degree they agree with the structure of samples.In addition to considering the relevance of features,we put forward a new algorithm for feature selection called FSIR2,which integrates feature relevance and redundancy.The FSIR2 evaluates feature relevance based on spectral feature selection theory,and considers the redundancy between features.By maximizing the feature relevance and minimizing the redundancy combinatorially,the algorithm can obtain a low redundancy feature subset.FSIR2 can be applied to both supervised learning and unsupervised learning.Unlike feature selection,feature learning is more general which focuses on map-ping the original feature sets into a new feature space to get the optimal expression of data.Existing feature learning algorithms can be divided into two types:the traditional learning algorithms and the neural network based algorithms.Extensive methods have been proposed for supervised feature learning based on convolutional neural networks and recurrent neural networks,in the meantime the research which makes full use of the large amounts of low cost unsupervised data for feature learning is inadequate.In this dissertation,we propose a convolutional AutoEncoder model called SoundAutoEn-coder.The algorithm conducts unsupervised feature learning for audio data.On one hand,it applies convolutional AutoEncoder in order to fully tap the effective informa-tion in audios;on the other hand,on account of the natural consistency of image data and audio data in videos,it extracts semantic information of images by mature visual recognition model,which then give guidance to the feature learning process.In experiments,for FSIR2,supervised learning and unsupervised learning are con-ducted on 10 datasets,the classification and clustering accuracy as well as redundancy are evaluated for the selected feature subset.Consider the results on each data set on average,compared with the relatively best algorithm MCFS,FSIR2 improves the clus-tering accuracy and NMI by 4%,decreases the redundancy rate by 5%,and matches with MCFS on classification accuracy.For SoundAutoEncoder,three datasets are used for training a classifier to evaluate classification accuracy for the learned features.Com-pared to SoundNet,the classification accuracy of SoundAutoEncoder were raised by 0.6%on DCASE-2016,6.9%on ESC-10,and 6.3%on ESC-50.