A MUSIC VIDEO RECOMMENDER SYSTEM BASED ON EMOTION CLASSIFICATION ON USER COMMENTS
Phakhawat Sarakit · Thammasat University Digital Collections · 2015
Along with the concept of collaborative intelligence, user comments are a useful information source for adding more value on online resources, such as music, video, books and other multimedia resources. While several works have been conducted to utilize user comments for sentimental analysis, i.e., LIKE or UNLIKE, there are still few exploitations of such comments for detecting emotion (user mood) on online resources. With emotion recognition, it is possible for us to understand the content of online resources and use the recognition result for further value added service, such as product recommendation. This thesis proposes a two-step method to perform emotion classification using user comments and utilize the result for music video recommendation.In the first step, the emotion filtering tags user comments with three label types of emotional comments, non-emotional comments, and unrelated junk comments. As the second step, the emotion classification aims to classify the emotional comments into six emotion types, including anger, disgust, fear, happiness, sadness, and surprise. With the YouTube API, the total of 85 video clips with 12,000 comments are collected and used for emotion filtering and classification. The emotion filtering detects that 5,345 comments are emotional comments and the emotion classification categorizes them into six emotional classes using 7,722 features (word types) extracted from the dataset. For the classification method, three alternative machine learning algorithms are considered; (1) multinomial naive Bayes (MNB), (2) decision tree induction (DT), and (3) support vector machine (SVM), where the best SVM method obtained at most 76.41% accuracy for the filtering task and 75.68% for the classification task. However, with error analysis, it was found that such low accuracies may be caused by class imbalance where happiness and sadness classes suppress other remaining classes.To improve performance of the emotion classification using unbalanced data, a sampling based algorithm called SMOTE (Synthetic Minority Over-Sampling Technique) is applied. With this preprocess, the SVM classifier yielded the best classification accuracy of 93.30% on the emotion filtering task and 89.43% on emotion classification task. The result improves by 16.9% and 13.76% for filtering and classification, respectively. Finally, a music video recommendation system is developed by suggesting a set of video clips which have a similar emotional profile with the currently focused clip.