摘要
针对音乐平台的迅猛发展,为实现音乐片段的自动推荐,设计基于改进无监督学习的音乐片段自动推荐技术。通过改进的频繁项集挖掘算法对平台用户历史播放数据实施频繁集挖掘。基于残差学习的噪声学习网络改进无监督学习框架(double image prior denoised network,DIPDN),设计掩码图像块去噪网络(masked image patch denoising network,MIPDN)无监督学习框架,实现挖掘数据的去噪处理。依据歌曲节奏对去噪后的数据实施分段处理,设计音频特征提取器提取各段的音频特征,并将其作为基于自编码器的音乐片段推荐模型的输入,实现音乐片段的自动推荐。测试结果表明,设计方法的Precision最高达到0.941,Recall最低仅为0.056,在18:00—22:00这一时段,用户取向捕捉度高达100%。
Aiming at the rapid development of music platform,in order to realize the automatic recommendation of music clips,an automatic recommendation technology of music clips based on improved unsupervised learning is designed.It implements frequent itemset mining on platform user historical playback data through an improved frequent itemset mining algorithm.The noise learning network based on residual learning improves the double image prior denoised network(DIPDN)unsupervised learning framework,it designs the masked image patch denoising network(MIPDN)unsupervised learning framework,and implements the denoising of mining data.It also implements segmented processing of denoised data based on song rhythm,designs an audio feature extractor to extract audio features from each segment,and uses this as input for a music segment recommendation model based on an autoencoder to achieve automatic recommendation of music segments.The test results show that the highest Precision of the design method is 0.941,and the lowest Recall is only 0.056.During the period from 18:00 to 22:00,the user orientation capture is as high as 100%.
作者
孟丽强
MENG Liqiang(Henan Open University,Zhengzhou 450008,China)
出处
《自动化技术与应用》
2025年第9期69-73,共5页
Techniques of Automation and Applications
基金
河南省科技攻关项目(212102210160)
2021年度河南省高等教育教学改革研究与实践项目(2021SJGLX321)。
关键词
改进无监督学习
MIPDN无监督学习框架
音频特征提取器
音乐片段推荐
improving unsupervised learning
masked image patch denoising network unsupervised learning framework
audio feature extractor
music segment recommendation
