Optimization Of FCBF Feature Selection Algorithm And Research On Psychological Stress Assessment Via Speech

The rapid development of information technology has led to large numbers of high-dimensional data that affect the discovery and understanding of knowledge seriously.Therefore,dimensionality reduction is crucial.Feature selection is one of the most widely used and effective method.However,a problem exists in most feature selection algorithms that the runtime and classification accuracy don't match.In order to find a balance between them,this thesis proposes two feature selection algorithms: AWFCBF and SHFS based on FCBF and they achieve prosmising improvement on feature selection and classification accuracy.Feature selection algorithms have been widely used in many fields.This paper focuses on psychological stress assessment via speech.Comparing with rating scales,voice has obvious advantages for psychological stress assessment.Detecting stress through voice is easily accepted by people with non-invasive and portable.However,how to select effective acoustic features is not clear.Featue selection algorithm is one of the most common ways to solve this problem.Therefore,this paper designs experiments that psychological stress is induced by workload.In experiments,we select acoustic feature subsets though SHFS.Then,we classify and analyze the features of SHFS selected.Experimental results show that SHFS can select effective and stable feature subset to assess psychological stress.The main contribution of this paper and innovative points are as follows:1.This thesis proposes AWFCBF algorithm based on FCBF feature selection algorithm.It assigns corresponding weight according to C-correlation.In this way,AWFCBF algorithm enhances the dependence between features with larger C-correlation and decreases the redundancy between features with smaller C-correlation.Experimental results show that AWFCBF feature selection algorithm achieves improvement on classification accuracy and stability.2.This thesis designs SHFS feature selection algorithm through the analyses and studies of AWFCBF in depth.For the features with larger C-correlation,SHFS selects features using sequential backward search strategy,and the smaller using sequential forward search strategy.The experimental results indicate that SHFS not only decreases runtime significantly but also improves the classification performance and stability.3.Using voice to assess psychological stress,this paper employs SHFS feature selection algorithm to select acoustic features.Results suggest that SHFS algorithms can select effective and stable features to assess psychological stress.Furthermore,this thesis analyses acoustic features obtained from SHFS features selection algorithm and analysis results indicate that that spectral features are proper for vowel;the combination of prosodic and cepstral is more suitable for figure;these three types features are the primary choice of sentence.Furthermore,short-term psychological stress is influenced by the level of long-term psychological stress.Therefore,we should consider both short-term and long-term psychological stress to explore the assessment of psychological stress.