基于增强学习的个性化音乐情感分类系统研究

发布时间：2019-04-30 08:01

【摘要】：音乐是情感的重要载体,是人们日常生活中不可或缺的重要因素。随着数字音乐数量的快速增长,人们对音乐情感分类和检索的需求也在不断增加。目前的音乐情感分类研究均把情感分类看作一个静态的过程,并且忽视了用户对音乐情感理解的主观偏好。本文主要研究基于增强学习的个性化音乐情感分类系统。本文以心理学经典情感模型VA模型为基础提取了一个包含12种情感类别的情感空间,以歌曲的音频特征为分类基础,应用并比较多种分类算法训练了一个全新的音乐情感分类静态模型。考虑到静态分类模型本身的误差和用户对音乐情感理解的主观偏差,在该静态模型的基础上,本文对用户在收听音乐过程中的行为进行了分析,构建了一个新颖的基于增强学习的动态模型。该模型通过对用户行为的学习动态调整情感分类结果,以实现对用户情感偏好的个性化定制。本文从商业音乐网站收集了均匀涵盖12种情感类别的600首歌曲,对歌曲进行了音频特征提取和特征筛选,建立了一个全新的音乐情感分类训练集,实现了一个新颖的音乐情感分类原型系统。本文通过实验分析了静态情感分类模型的分类正确率,通过对用户行为的模拟验证了动态学习模型的可行性和有效性。小规模的用户体验调查分析结果表明本文所研究的原型系统具有较好的应用效果。
[Abstract]:Music is an important carrier of emotion and an indispensable factor in people's daily life. With the rapid growth of the number of digital music, the demand for music emotion classification and retrieval is increasing. The current research on music emotion classification takes emotion classification as a static process and ignores the subjective preference of users for music emotion understanding. This paper mainly studies the personalized music emotion classification system based on enhanced learning. Based on the classical emotional model of psychology, VA model, this paper extracts an emotional space containing 12 emotional categories, and classifies the audio features of songs as the basis of classification. A new static model of music emotion classification is trained by applying and comparing a variety of classification algorithms. Considering the error of static classification model itself and the subjective deviation of user's understanding of music emotion, this paper analyzes the behavior of users in listening to music on the basis of this static model. A novel dynamic model based on reinforcement learning is constructed. The model adjusts the emotional classification results dynamically by learning the user's behavior to realize the personalized customization of the user's emotional preference. This paper collects 600 songs covering 12 emotional categories evenly from the commercial music website, carries on the audio feature extraction and the feature screening to the songs, and establishes a brand-new music emotion classification training set. A novel prototype system of music emotion classification is implemented. In this paper, the classification accuracy of static emotion classification model is analyzed by experiments, and the feasibility and effectiveness of dynamic learning model is verified by simulation of user behavior. The results of a small-scale user experience survey show that the prototype system studied in this paper has a good application effect.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN912.3;TP18

【相似文献】