基于用户行为轨迹的在线音乐偏好模型  

Online Music Preference Model Based on User Behavior Trajectory

作　　者：刘义理[1] 朱茂然[1] 胡莼 LIU Yili;ZHU Maoran;HU Chun(School of Economics and Management,Tongji University,Shanghai 200092,China)

机构地区：[1]同济大学经济与管理学院,上海200092

出　　处：《复旦学报（自然科学版）》2022年第3期342-352,共11页Journal of Fudan University：Natural Science

摘　　要：将音乐情感识别集成到音乐推荐系统中时存在用户感知情感与歌曲/歌词预期情感不一致的困难。为了解决这一难题,本文通过分析用户行为轨迹来构建用户音乐偏好。对用户播放行为中记录的歌曲使用LDA方法识别歌曲歌词,生成客观文本向量,基于对应的用户评论生成主观文本向量,再将两个向量融合为表达用户音乐偏好的综合文本向量,然后使用用户播放行为特征来处理时间和播放次数两个要素对用户偏好的衰减影响,并使用用户行为统计特征来平衡用户长期行为特征对用户当前偏好的影响,从而建立基于用户行为轨迹的用户在线音乐偏好模型。通过采集网易云音乐数据进行实证研究发现,本模型的推荐效果要高于单纯使用歌词文本向量的偏好算法。本研究为基于情感识别的在线音乐推荐提供了新的思路和方法。Inconsistency between felt emotion and perceived emotion is one of key difficulties of integration of music emotion recognition and music recommender system. To solve this problem, this paper constructed users’ music preference by analyzing users’ behavior trajectory. Based on user playback behavior records, objective text vectors of the songs’ lyrics and subjective text vectors of user reviews about the songs by LDA method were generated and fused as comprehensive text vectors to express user music preferences. User play behavior characteristics were introduced to handle the attenuation effect of time and play times on user preferences. User behavior statistical characteristics were integrated to balance the impact of user long-term behavior characteristics on user current preferences. Thereby a user online music preference model based on user behavior trajectory was established. Experiments with data from NetEase Cloud Music were carried out and findings showed that the recommendation effect of this model is higher than that of the preference algorithms simply using lyrics text vectors. This research provides a new idea/method for online music recommendation based on emotion recognition.

关 键 词：用户偏好 音乐情感识别 行为轨迹 音乐推荐系统 

分 类 号：TP182[自动化与计算机技术—控制理论与控制工程]