基于面部表情的学习困惑自动识别法

Automatic Learning Confusion Recognition Using Facial Expressions

The role of emotion in learning remains largely unexamined and certainly undervalued in the education literature， despite recent molecular research found the evidence of the impact of emotion to learning. Learning emotion recognition is regarded as one of crucial technology in next-generation intelligent tutoring system. Among learning emotion， confusion is one of the most common one， which has a great influence on the learning outcomes. However， automatic recognition of learning confusion is a challenge because of its strong implicitness，compared to basic emotion such as happy and angry. We designed a confusion-inducing experiments within an online test environment and applied machine learning algorithms to detect the confusion emotion reflect from the face of subjects. The subjects were asked to solve a series of test questions with different difficulties within limited time. A high-solution professional camera was set in front of subjects to capture face image in high frequency. Then， the 78 key feature points of human face were extracted as feature vector， and fed in to a machine learning model to predict confusion. To get a robust prediction model， six popular classification algorithms， i. e. ， logistic regression， support vector machine， K-nea-rest neighbor， decision tree，random forest，deep learning，were tested. A 10-fold cross validation was used to compare the six algorithms on five performance metrics that calculated using subjects， self-reported confusion labels. The experimental results show that most of the classification algorithms can detect studentsy learning confusion effectively and the average accuracy is over 58 % ， and the random forest is the best prediction model with an average accuracy of 71. 18%. The proposed method in this study can provide technical support for the learner affect modeling in the next generation of intelligent tutoring system. The current results are promising，but the evaluation revealed some limitations that could further improve the research. Firstly， the self-reported emotion label is subjective and could be combined with field observation to improve the precision of labeling. In addition，the representation ability of feature point is still very limited. Therefore，in the future，we will attempt to use directly the image to detect confusion with deep neural network technology. Lastly，we also plan to investigate the possibility to simultaneously detect multiple learning emotions， such as confusion， boring and engagement.