基于状态转移的奖励值音乐推荐研究

Research on Music Recommendation of Reward Value Based on State Transfer

摘要:

听音乐有助于纾解人们的压力，现已成为大众娱乐的一种重要方式。互联网的发展使人们听音乐变得方便，但同时也使得“信息过载”的问题变得日益严重。尽管各大公司平台纷纷推出了针对音乐的推荐系统来解决这个问题，但现有传统的推荐系统并不能保证用户体验，用户对精准推荐的需求仍是强烈。为解决“信息过载”问题的同时并保证用户体验，本文提出了基于状态转移的奖励值算法。该算法包括对用户自身喜好建模，并利用用户集数据提出的音乐流行度和用户从众度，根据用户喜好、音乐流行度以及状态转移概率定义奖励函数。所提出的算法能个性化的对音乐库数据进行筛选和聚类。在对数据进行处理时，采用Davies-Bouldin指数对声乐特征进行离散化。在模型训练时，采用基于列表距离最小化的计算方法对参数进行选择。通过在Million Song Dataset开源音乐数据集上的实验，表明在算法中加入音乐流行度对推荐效果有一定影响，本文所给出的推荐算法能够提高推荐的效果，说明了本文算法的有效性。

Abstract:

Listening to music is helpful in relieving the pressure of people, and has become a major entertainment for the general public. The development of the Internet makes it convenient for people to listen to music, but it also makes the problem of “information overload”more and more serious. Although many internet companies have launched music recommendation system to solve the problem, the existing recommendation systems cannot guarantee good user experience. As a result, there is still a populardemand for precise recommendation for music. In order to solve the problem of “information overload” and guarantee good user experience atthe same time, this paper presents a reward value algorithm based on state transition.Specifically, the user preference model is first built; then, the music popularity and user conformity is proposed based on user data; finally, the reward function is defined based on user preference, music popularity and state transition probability. The proposed algorithm can individually screen and classify the data from the music library. In the algorithm, the DaviesBouldin exponent is used to discretize vocal characteristicswhen processing data; the algorithm based on list distance minimization is used to select parameters during the model training. The experiments are conducted on the Million Song Dataset , and the results show that the music popularity has certain influence on the recommendation effect of the algorithm. The recommendation algorithm proposed in this paper can improve the performanceof recommendation, which proves the effectiveness of the proposed algorithm.