Research On Piano Music Recommendation Algorithm Based On Convolutional Neural Network

The amount of online digital music is increasing rapidly,which brings users difficulty in finding their favorite ones.Music recommendation algorithms try to make personalized recommendation based on the analysis of users' preferences,which can reduce the effects of information overload and thus improve users' experience.Our work is towards the designing of a content-based recommendation algorithm for digital piano music.Specifically,our work includes the following three parts.(1)Feature extraction and selection for piano music classification.The first set of features considered includes time-domain,spectrum and amplitude,which characterized the rhythm,melody and pitches of a music.In addition,an abstracted feature: the notes,is considered.The notes are automatically extracted by our improved note recognition algorithm.In the algorithm,we proposed to use moving frames to separate notes and thus the noise of doubled wave is averted.The use of both frequency features and music notes feature makes our work distinctive in the field of piano music classification.(2)Classification of piano music.We considered four categories: Blues,Classical,Jazz,and Pop.A convolutional neural network(CNN)is trained to classify a music into these categories.In the training phase,we studied the effects of two activation functions: ELU and ReLU,and two gradient descent strategies: RMSProp and Adam.Experimental results showed that ELU is more stable than ReLU,and Adam is faster than RMSProp.Different configurations of music features were also investigated.Results showed that with the music notes feature used the accuracy of our algorithm can be improved by 1.5%,beyond the 96% accuracy achieved where notes are not used.(3)Classification based piano music recommendation.We firstly identify the reason why CNN's classification result is not suitable for recommendation,and then proposed a recommendation method based on the classification of music segments.Specifically,a piece of music is divided into several segments,whose probability distribution on candidate categories are calculated.A threshold model is proposed to distinguish pseudo discrete and real discrete,which is showed to improve the accuracy of classification.With an analysis on the relation of a piano music's category and the user's history,a recommendation is made.Experiments were carried out on piano music collected from HQGQ.com and music.163.com.Simulated user behavior data generated by a HMM model is used for evaluation.Results showed that our recommendation algorithm's accuracy could reach 50% or higher,which is competitive to related music recommendation methods.