| Summary: | 碩士 === 國立成功大學 === 資訊工程學系 === 102 === According to the World Health Organization (WHO), depression is currently one of many serious problems, and awareness of negative emotions is helpful for treating it. Behavioral patterns can either be an antecedent or a consequence of human emotion. For example, usage patterns on smartphones can reflect the user’s emotion. With the popularity of smartphone ownership, researchers are beginning to examine the association of smartphone usage patterns with emotional conditions. This study uses smart phone usage patterns to detect emotional states, aiming to improve self-awareness of negative emotion.
We developed three Visual Analogue Scales to measure and mark the emotional status. The package names of applications shown on the light-on screen are recorded as phone usages. The timeslots were set for each emotion mark in order to determine whether a usage feature is associated with the mark or not. Different users may have different usage patterns that reflect their emotions. We utilize several feature selection methods and classifiers to determine personalized usage features for the machine learning. In summary, we considered four timeslots, five feature selection methods, and four classifiers; each combination can be viewed as a model. Finally, we developed a detection model selection method based on Rank product scoring to narrow down the combinations and to choose the best combination for the detection model.
The user has his/her distinct behavioral pattern on the smartphone. This unique data was used to train our personalized detection model. All personalized detection models achieved an average accuracy of 81.98 %, 84.58 %, and 82.96 % for detecting depression, anxiety, and stress, respectively, and outperformed the two baseline methods: linear regression (as applied by Microsoft’s MoodScope system) and general guessing. The general guessing method considers all detections according to the level of emotional conditions that appears most frequently. The personalized models were sent back to subjects for further evaluation and the results showed that the models predicted their emotional states with an accuracy rate of 85.9%.
We have developed an early negative emotion detection model for smartphones that, after a 14-day personalized training period, is able to detect negative emotional states based on the smartphone usage patterns two hours before detection. This model has a potential for ecological momentary intervention for depressive disorders by envisioning negative emotions and informing the users how they have interacted with the smartphone before they actually reach negative emotional status.
|
|---|
