【作者】 吴翔；

【导师】 陈恩红；

【作者基本信息】 中国科学技术大学 ， 计算机应用技术， 2014， 硕士

【摘要】 随着因特网的迅猛发展,在线多媒体娱乐被越来越多的人所接受。同时,在线KTV作为一种新兴的在线多媒体娱乐形式,受到了越来越多的关注。但是受信息过载的影响,人们在成千上万的歌曲中寻找自己的所需,变得越来越困难。有的用户由于平时熟练的歌曲不多,即使在系统中有自己会唱的歌曲,往往也很难快速发现它们；而另一些用户由于会唱的歌曲过多,进入系统后又会显得不知所措。因此使用推荐系统,帮助用户从众多的候选项目中快速地选出符合其喜好的音乐,就显得十分重要。然而,因为在线KTV系统中的分数,代表的是用户演唱水平,而非其喜好程度,所以普通的协同过滤推荐系统在这里并不适用。同时,由于情感等因素的影响,用户的翻唱具有一定的序列特性,而一般的推荐系统并不擅长捕捉这一特点。因此本文对歌曲与用户进行建模,对在线KTV音乐推荐进行了探索,主要研究内容与贡献如下：1.通过对用户历史翻唱记录的学习,捕捉了用户与歌曲以及歌曲之间的联系,提出了个性化马尔可夫映射模型,同时利用用户的长期偏好和短期偏好为用户进行在线KTV音乐推荐。2.在个性化马尔可夫映射模型的基础上提出了一个新的推荐目标函数,在推荐过程中考虑歌曲在模型空间中分布的多样性。同时由于多样性因素在推荐结果中的重要性是可调的,推荐的准确性不至于被过多影响。3.在真实数据集上的实验结果表明,本文提出的模型和算法,可以更好地为用户进行歌曲推荐,具有多样性的推荐方法也可以有效地达到预期的结果。更多还原

【Abstract】 While the Internet is ever growing rapidly, online multimedia entertainment is becoming more and more popular. As a new form of online multimedia entertainment, online KTV also attracts more attention. But due to the effects of the so-called information overload, people find it hard to pick proper songs from thousands of candidates. Some who know little, may take a long time to find a familiar song. While others who know much, may be overwhelmed by all the available content. Therefore, it is important to have a recommender system, helping people pick up their preferred songs.However, scores in a online KTV system represent users’ prformances rather than their preferences and singing records seems to have sequential property. Thus, recommender systems such as collaborative filtering cannot be directly applied here. This thesis built models for both users and songs, and made an attempt for online KTV music recommendation. The major work and contributions are as follows:1. Learnt useful information from user singing history, built relationships among songs and users. Proposed the Personalized Markov Embedding model and made recommendations based on users’ long-term and short preferences.2. Proposed a new goal function for recommendation, considering the diversity of songs’distribution in model space. Made this consideration adjustable in order not to affect the accuracy of the recommendation too much.3. Experimental results on a real world dataset showed that models and algorithms proposed in our thesis can make better recommendations for individuals, and the diversified model also worked as expected.更多还原