Research On The Learning Method Of Marker Distribution Based On Correlation

In recent years,learning with ambiguity is a popular topic in machine learning and data mining areas.There are mainly two sophisticated paradigms for addressing label ambiguity problems at present,namely single-label learning and multi-label learning,respectively.In single-label learning framework,an instance is associated with a single class label,whereas in multi-label learning,an instance may have multiple class labels simultaneously.Obviously,multi-label learning can address more label ambiguity problems compared to single-label learning.However,both single-label learning and multi-label learning actually aim to answer the question “which label can describe the instance”,but they cannot handle the further question with more ambiguity,i.e.,“how much does each label describe the instance”.Thus,to solve such cases,label distribution learning is proposed.Label distribution learning is a further extension of multi-label learning and it is suitable for addressing more complex label ambiguity problems.Besides,it is well known that the key to label distribution learning is to mine and exploit the correlations in data.Based on this,this thesis studies the label distribution learning based on the local sample correlations and local label correlations.Firstly,this thesis proposes a label distribution learning algorithm which exploits sample correlations locally.Current algorithms generally consider the correlations among labels,but have not exploited the correlations between local samples and instances.To exploit such correlations,this thesis clusters the instances into different topics in label space,i.e.,different local samples.The instances in each local sample share similar label correlations.Besides,to encode the influence of local samples,a local correlation vector is designed for each instance based on the clustered local samples.Each item in the local correlation vector represents the impact of each cluster on the instance.Then,the label distribution for an unseen instance is predicted based on the original features and the local correlation vector simultaneously.Experimental results demonstrate that our proposed methods can effectively address the label distribution problems by exploiting local sample correlations,and perform remarkable better than the state-of-the-art methods.Secondly,this thesis proposes a local low-rank label correlation based facial expression emotion distribution learning method.Substantial previous research assumes that each facial expression is associated with one or more predefined affective labels,while ignoring the fact that multiple emotions always have different intensities in a single picture.To depict facial expressions more accurately,this thesis adopts a label distribution learning approach for emotion recognition.Besides,existing works exploit label correlations in a global manner under the assumption that the correlations are shared by all instances.However,the label correlations in facial expression emotion recognition are usually local,where a label correlation may be shared by only a subset of instance.Based on this assumption,a label distribution learning method by exploiting local label correlations is proposed.Moreover,considering the complexity of emotion correlations,a local low-rank structure is adopted to capture the local emotion correlations implicitly,rather than considering the label correlations explicitly.Experiments on benchmark facial expression datasets demonstrate that our method can better exploit emotion correlations and is superior to some state-of-the-art label distribution learning methods.