Abstract: Traditional music rating recommendation model has lower accuracy, and recommended accuracy has received a great impact, because of user insufficient score and large subjective difference. Mining user interest from huge history behavior data is an excellent method to address those problems of rating model. Features are extracted from user behavior data to establish feature model, which help get corresponding user preferences. FM（Factorization Machine） predicts different types of user behavior, which is the basis of recommendation. FM has made full use of useful music features and user features information. The most important is that FM can simulate hidden factors of user behavior data to fill sparse matrix, and sparsity has little impact on prediction. Compared with traditional recommendation models, ming user interest from behavior data is feasible, experimental results also show that this method has good effect on music recommendation.

Key words: music recommendation, factorization machine, behavior prediction, data ming

摘要： 针对传统音乐评分推荐模式用户评分缺失和主观差异性较大等问题，通过提取用户行为数据构建行为特征模型，用以分析用户行为与兴趣的关联性，并采用因子分解机（Factorization Machine，FM）预测用户行为类型，作为音乐推荐的依据。将FM应用到该方法中，充分利用音乐和用户属性特征，并且通过模拟用户行为特征数据中的隐因子来填充推荐的稀疏矩阵，降低数据稀疏对预测的影响。与传统音乐推荐方法相比，从用户历史行为中挖掘用户兴趣倾向以解决评分模型带来的问题更具可行性，实验结果表明该方法用于音乐推荐也具有良好的效果。

关键词: 音乐推荐, 因子分解机, 行为预测, 数据挖掘