基于多模态的音乐推荐系统  

A music recommendation system based on multi-modal fusion

作　　者：龚志 邵曦[1] GONG Zhi;SHAO Xi(College of Telecommunications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003)

机构地区：[1]南京邮电大学通信与信息工程学院,南京210003

出　　处：《南京信息工程大学学报（自然科学版）》2019年第1期68-76,共9页Journal of Nanjing University of Information Science & Technology（Natural Science Edition）

基　　金：国家自然科学基金(70573025)

摘　　要：使用传统协同过滤的方式进行推荐往往会忽视音乐底层特征.通过将音乐的音频特征与歌词信息进行多模态融合,并将融合后的特征信息作为协同过滤推荐的补充,提出了一种基于多模态的音乐推荐系统.主要探讨了音频特征与歌词信息的提取,并在提取歌词信息时利用LDA主题模型进行特征降维.针对多模态融合问题,使用一种特征级联早融合法(EFFC)融合方式,并将多模态融合后的结果与单模态结果进行了比较.对于结果的推荐,以多模态特征信息为依据建立用户兴趣模型,并将该模型通过LSTM神经网络,以过滤与优化协同推荐的用户组.结果表明,基于多模态的音乐推荐系统将推荐结果的误差项平方和(SSE)由传统的2. 009降至0. 388 6,验证了该方法的有效性.Despite the continuous enrichment of music,the underlying music features are often overlooked when using traditional collaborative filtering.By multi-modal fusion of audio features and lyric information and supplementing the fusion information feature as a collaborative filtering recommendation,a multi-modal music recommendation system is proposed.This studyprimarily discusses the extraction of audio features and lyrics information and uses the LDA topic model to reduce the character dimension of the lyrics information.For the multi-model fusion problem,this study proposes an EFFC fusion method,and compares the results of multi-modal fusion with the results using single-mode.For result recommendations,the user interest model is established based on the multi-modal information feature with the input of LSTM networks to filter and optimize the user group.The results show that the multi-modal music recommendation system reduces the SSE of the result from 2.009 to 0.388 6,verifying the effectiveness of the method.

关 键 词：音乐推荐 协同过滤 LDA主题模型 多模态融合 LSTM神经网络 

分 类 号：TN912[电子电信—通信与信息系统]