Research On Key Recommendation Methods And Technologies In Digital Music Marketing

The growth and popularity of streaming music has changed the way people consume music.Users can listen to online music content at anytime and anywhere.This change trend has prompted the digital music marketing mode to gradually change to a user-centric personalized marketing strategy.The birth of recommendation system made it possible.It incorporates multiple recommendation algorithms and strategies(such as user portraits,collaborative filtering,content filtering,etc.)to capture the user's interest preferences and recommend the users with contents they are probably interested in.However,compared with other types of recommended content,there are many unique and challenging problems in digital music marketing,which causes that the existing recommendation algorithms cannot bring a good recommendation effect.To this end,how to propose a practical and effective music recommendation algorithm to improve the marketing efficiency of digital music is one of the important problems to be solved urgently.Based on the research results of music preference theory and recommendation technology,this paper uses big data and machine learning technology to deeply excavate user attributes,music metadata and user behavior data,and proposes a series of personalized,precise and intelligent music recommendation algorithms.The research results has played a key role for digital music in content distribution,user experience and commercial monetization.It achieved the marketing purpose of reducing costs and increasing efficiency,and also provides practical guidance and optimization suggestions for marketing methods in other content fields.Meanwhile,the research results of this paper have further expanded the application of marketing theory and service science,and injected new impetus into the development of the digital music industry.Specifically,the main research contents and innovations include:(1)Aiming at the problem of sparseness of behavioral data in digital music marketing,a metric ranking learning recommendation model based on content representation(CRMRL)is proposed.The main innovation:the current mainstream solution is to map users and music to the same dense semantic space,and then use the inner product of users and music in this space to excavate the user's preference relationship for music.To this end,this paper proposes to use observed and unobserved behavior data to construct a relative partial order relationship so that the model can be fully trained.At the same time,the audio feature extraction sub-model related to the recommendation task is constructed to further alleviate the data sparsity problem.Finally,the metric learning is used to mine Fine-grained and global relationship between users and songs.Compared with the existing recommendation algorithm,the CRMRL algorithm solves the recommendation problem under extremely sparse data scenarios,and can fully excavate the user's music preference characteristics.The research results can be used to increase the recommendation exposure rate and marketing investment of new songs\non-popular songs,which stimulates the creative enthusiasm of music artists.(2)In view of the differences and dynamics of users' music preferences,a music recommendation algorithm based on multi-layer attention representation(HARM)is proposed.The main innovation:the current industry methods mainly predicts the next song that the user most likely prefers based on the list of songs listened by the user and the timing relationship.This kind of methods can only excavate the timing relationship in the general sense,ignoring that the focus and degree of difference of different users towards the multi-dimensional features of the same song are different.To this end,this paper uses technologies such as self-attention networks and recurrent neural networks to learn embedded representations of music from multiple dimensions in user attributes and song content,and then identify differences in users'preferences for the same song and the user's interest degree for different songs in the listening sessions.Compared with the existing recommendation algorithm,the HARM algorithm realizes a multi-dimensional capture method of users'music preferences and further improves the prediction accuracy of the next song.The research results can be used to predict the next song and other recommendation tasks,bringing users with a more personalized music listening service,and then increases users' satisfaction and paying willingness towards the music platform.(3)Aiming at the problem that there exists a lot of noise data in implicit feedback behavior,an attention-mechanism-based sequential recommendation(ASR)algorithm is proposed.The main innovation:The current industry methods mainly use the attention mechanism to mitigate the impact of noise data,but most models only focus on the research of the attention mechanism at the global level,without considering recognizing the local noise data of the music from a fine-grained level.To this end,this article first characterizes the local feature information of music,uses the convolutional neural network to extract the high-level semantic features of music,and then reconstructs the high-level semantic features of the song from the previous layer into a conversation sequence based on the user's listening time sequence.The constructed two-way recurrent neural network model based on attention mechanism enables it to reduce the influence of noise data and learn strong dependencies between songs.Compared with the existing recommendation algorithm,the robustness of the ASR algorithm under different session lengths is significantly enhanced,which effectively improves the recommendation quality of the model under the influence of noisy data.The research results can be used for tasks such as daily playlist recommendation\automatic playlists,etc.,and enhances the user's perceived value for listening to this music theme.