Research And Application Of Music Recommendation Algorithms Based On Behavioral Sequences And Features

In recent years,with the speedy development of internet technology,the unspeakable quantity of information has brought unprecedented challenges and difficulties for people to obtain certain information.Such a phenomenon is considered as information redundancy.This problem was not effectively solved until the advent of the recommendation system.At present,the recommendation system is developing rapidly,and various recommendation algorithms emerge in endlessly.Especially with the wide application of deep learning in image processing,it has been naturally applied to recommender systems to more effectively alleviate the information overload problem and speed up information filtering.Recommendation system are now a hot topic of interest in the industry and are being used in a number of areas,including music.A good music recommendation system can better provide services for users,reduce user screening time,and create higher value for users and platforms.Therefore,many people at home and abroad have focused on music recommendation,conducted a lot of research on it,and produced a large number of results.However,compared with other types of recommendation,music recommendation with unique characteristics mainly obtains user preferences in an implicit way,which is more flexible.Due to its limitations,the effect of previous methods needs to be improved.Therefore,people also apply deep learning to music recommendation.Based on this,this paper improves the existing time series recommendation model based on deep learning from different angles,and applies it to the music recommendation system.The main work of this paper is as follows:（1）A DIAFN（Deep Interest and Attentional Factorization machines Network）model is proposed.Based on the Deep Interest Network,the model uses a two-way parallel structure and introduces an attention based factor decomposition machine model to enable the user behavior sequence to realize the effective combination between low-order features,realize the information that can extract both low-order features and high-order features,and capture multiple levels of interactive behavior,enable the model to better learn the user’s interest characteristics and obtain richer and more accurate behavior dependence information;the model also designs a new attention calculation method,which calculates the similarity weight by using the relationship between the cosine similarity and the absolute value of the module length difference of any two vectors,which can more effectively learn the relationship between any two vectors and improve the interpretability of the model.（2）A DICN（Deep Interest and Compressed interaction Network）model is proposed.The model similarly based on the Deep Interest Network,with compressed interactive network.Fully excavate the relationship between high-order features,highlight the importance of feature intersection,and make the model combine explicit vector level and implicit element level feature combination at the same time,so as to fully excavate the potential information of item combination features.On this basis,a new attention calculation method is used to verify its effectiveness again.（3）A number of commonly used public datasets and the private dataset in this paper are selected and compared with advanced models in extensive experiments respectively,and the experimental results are used to verify the effectiveness and practicality of the model.Finally,the proposed model is applied to a personalized Music Recommendation System which is designed and developed in conjunction with practical scenarios.In a word,this paper mainly combines the actual music recommendation scene,designs and implements the timing recommendation model based on deep learning from different angles,which effectively improves the model and is verified in the personalized system,which has a certain practical significance.