Research On Ensemble Learning For Depression Recognition Based On Speech

Depression is a serious mental disease which affects more people with the increasing pressure in the society.At present,the diagnosis of depression mostly depends on various scales and communication between patients and doctors,which brings a risk of subjective bias.Speech is one of objective human behavior measurement,which is fast,convenient and non-intrusive.At present,there are some achievements in depression detection based on speech.In the study,ensemble learning was applied to improve it.Ensemble learning is a kind of method which construct various learners to solve one problem,which win the favor of industry and academia.At present,the ordinary ensemble methods train different classifiers on the same dataset and combine them.Ensemble pruning selects a subset from all learners in ensembles to reduce cost and get a better performance.Our work have two contributions.The first one was constructing an ensemble model to detect depression based on speech.We used different speech to generate various classifiers and combine them based on the fact that speech is easily to get.Experiment showed that this ensemble model got a better result than general ensemble model combining various classification algorithm.The second contribution was proposing a new ensemble pruning method based on sample probability to make up some deficiencies in ensemble pruning methods at present.Then we compared them with three other classical ordering-based ensemble pruning methods on 12 UCI(University of California Irvine)datasets.The result showed that the improved algorithm based on probability of samples performed better than three classical ensemble pruning methods.In conclusion,we applied ensemble pruning in depression detection based on speech and improve the accuracy.And we proposed a new ensemble pruning method to improve accuracy of depression detection model.