个性化音乐推荐算法的研究与实现

发布时间：2018-01-20 16:53

本文关键词： 音乐推荐协同过滤 Spark 列表推荐空间模型　出处：《山东大学》2017年硕士论文　论文类型：学位论文

【摘要】：在互联网时代,各类的音乐网站提供了成千上万的歌曲,满足了人们对音乐的需求,为人们带来了极大的方便。但是,数字音乐的迅速发展造成了音乐信息的过载,面对海量的歌曲人们难以快速找到符合自己兴趣的音乐。音乐推荐系统的出现能够为用户推荐可能喜欢的音乐,帮助用户快速的发现或者找到自己想要的歌曲。这种推荐服务能够为用户提供良好的使用体验,带来商业利益,因此音乐推荐领域也成为工业界和学者们重视的研究方向。协同过滤算法广泛应用于各领域的推荐系统中,但是存在冷启动和可扩展性差的问题;传统的推荐算法没有考虑到音乐行为中特有的连续性行为和环境的影响,造成了推荐效果不理想。针对这些问题,本文对音乐推荐算法进行了深入的研究,对协同过滤算法进行了改进及基于大数据分析平台的并行化实现;同时,基于音乐收听的特点,实现一种音乐列表的推荐算法。首先,本文对于协同过滤算法进行了研究,针对基于ALS模型的方法进行了改进,根据用户属性信息对用户进行聚类分群,提出分群协同过滤算法。同时,对分布式计算技术进行研究,基于Spark平台实现算法的并行化设计与实现。基于Spark平台的分群协同过滤算法提高了推荐的准确性和效率,并解决用户的冷启动问题。然后,本文对于音乐收听具有连续性且与上下文环境有关的特点,实现了一种基于空间模型的音乐列表推荐算法。通过将三者之间的概率关系映射为空间距离,构建用户-歌曲-标签空间模型,实现对用户的个性化并且连续的歌曲推荐,满足用户当前音乐需求。最后,基于上述两种推荐技术本文设计了一种多功能音乐推荐系统。基于Spark分布式平台,从实际需求出发,实现对用户长期喜好预测的音乐推荐和当前会话期内需求偏好的连续列表推荐功能。
[Abstract]:In the Internet era, all kinds of music websites provide thousands of songs, to meet the needs of people for music, for people to bring great convenience. The rapid development of digital music has resulted in the overload of music information. In the face of massive songs people can not find music that meets their interests quickly. The emergence of music recommendation system can recommend music that users may like. Help users quickly find or find the song they want. This recommendation service can provide users with a good use of experience, bring business benefits. Therefore, the field of music recommendation has become an important research direction in industry and scholars. Collaborative filtering algorithm is widely used in recommendation systems in various fields, but there are some problems such as cold start and poor scalability. The traditional recommendation algorithm does not take into account the unique continuous behavior and the impact of the environment in the music behavior, resulting in unsatisfactory recommendation effect. In view of these problems, this paper makes a deep research on the music recommendation algorithm. The collaborative filtering algorithm is improved and parallelized based on big data analysis platform. At the same time, based on the characteristics of music listening, a music list recommendation algorithm is implemented. Firstly, the collaborative filtering algorithm is studied, and the method based on ALS model is improved. Cluster collaborative filtering algorithm is proposed according to the user attribute information. At the same time, the distributed computing technology is studied. The parallel design and implementation of the algorithm based on Spark platform. The clustering collaborative filtering algorithm based on Spark platform improves the accuracy and efficiency of recommendation and solves the cold start problem of users. In this paper, a music list recommendation algorithm based on spatial model is implemented for music listening which is continuous and context-dependent. By mapping the probabilistic relationship between the three to spatial distance, this paper proposes a new algorithm for music list recommendation based on spatial model. The user-songs-label space model is constructed to realize the personalized and continuous recommendation of songs to meet the current music needs of users. Finally. This paper designs a multi-function music recommendation system based on the above two recommendation technologies. Based on the Spark distributed platform, it starts from the actual requirements. It implements the function of predicting the long-term preferences of the user and the continuous list recommendation of the demand preference in the current session.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【相似文献】