Research On Methods Of EEG-based Emotion Recognition Using Multi-instance Learning

Emotion recognition,as the most basic and important link in affective computing,mainly uses various types of physiological sensors to capture the non-pseudo-intrinsic physiological information and build an emotion model,so that the machine has the ability to perceive and understand human's emotions.Among various physiological modals,electroencephalogram(EEG)signals have been increasingly concerned by related fields because they can directly reflect the activity of the central nervous system in the brain.Traditional EEG-based emotion recognition method is usually based on the assumption of stationarity of signal analysis.The whole EEG belonging to the same affective stimulus is divided into short-time windows,and then each window is treated as an independent sample and assigned the same emotion labels.Since the extracted signal segments may be affected by noise,unrelated brain activity,etc.,and it is not completely guaranteed that the subject will continue to produce EEG segments that are consistent with the emotional stimuli for the entire period of time.Therefore,it is likely to interfere with the classification and discrimination of positive and negative segment samples.On the other hand,since emotion is a complex cognitive process,it cannot be accurately defined by a single dimension,only examining its characteristics in multiple dimensions more effective.In view of the above two issues,this paper deeply discusses the shortcomings of the traditional supervised learning model in emotion modeling,and proposes an EEG emotion recognition framework based on multi-instance learning and multi-instance multi-label learning to make the emotion recognition model more reasonable and effective.The main work and contributions of this paper include the following two aspects:1.An EEG emotion recognition framework based on multi-instance learning was proposed.Using the representation method of multi-instance-bags,the feature vector corresponding to each time window of EEG signal is represented as an instance,while multiple instances that belong to the same affective stimulus form a bag,and make use of the multi-instance learning algorithms in the modeling process to capture important structural information among multiple instances in the bag.The experiment results based on DEAP(A Database for Emotion Analysis using Physiological Signals)dataset show that compared with traditional single-instance methods,theframework exhibits better classification and generalization performance based on multi-instance EEG emotion recognition,and achieved higher classification accuracies of 77.78% and 78.57% on two emotion dimensions(valence and arousal)respectively.2.An EEG emotion recognition framework based on multi-instance multi-label learning was proposed,the multi-dimensional and complex characteristics of the emotion model were fully considered.EEG multi-instances with sliding-time-window features are used to classify multiple emotion dimensions simultaneously to ensure the integrity of emotional expression.Through multidimensional emotion recognition experiments in the DEAP dataset,the results show that compared with the existing research results of multidimensional emotion recognition based on the DEAP dataset,the recognition accuracy of the framework in multiple dimensions has been significantly improved.The highest accuracies in “valence-arousal”,“dominance-arousal”,“dominance-arousal-valence”,and “dominance-arousal-valence-liking” reached 76.68%,73.90%,70.96% and 69.98% respectively.It once again validated the validity of the EEG emotion recognition framework based on multi-instance in quantitative analysis of EEG and multidimensional emotion recognition.