| （英） |
The authors have been conducting a research of video viewers’ emotion estimation by machine learning using bio-signals to improve the performance of the video recommendation system. In our previous studies, the number of videos was limited to 10-30 to avoid subjects’ excessive burden because they had to watch 4-5 minutes videos in the experiments. In addition, on the questionnaire asking which emotions were raised after watching each video, there is a problem that subjects had to recall what emotion is raised at what point in the video. Therefore, in this paper, we report the results of the experiment where the subjects watched 70 short videos of 30-45 seconds. In the analysis, we create the models using Random Forest from bio-signal data of each subject, and conduct the multi-class and binary classifications of emotions. As a result, the accuracies of all the models in the both classifications exceed the chance level when the bio-signal data during watching videos are included in the training. However, as same in our previous studies, the accuracies for unlearned videos are low, and the improvement of this point remains as a future issue. |