Digital Music Recommendation Technology for Music Teaching Based on Deep Learning

With the rapid development of music streaming media service industry, users can easily hear any song on mobile devices. Internet has become a huge music storage platform. With the development of network and large-scale digital music industry, the acquisition and listening of music are presented to users in a more convenient way. How to ﬁ nd the music loved by users from the massive Internet digital music data has become the key problem and main goal to be solved in the ﬁ eld of music information retrieval. Personalized music recommendation system can accurately ﬁ nd and push songs that users may be interested in from tens of millions of huge music libraries according to users ’ information under the condition that users only have vague demand for listening to songs. Relying on the traditional search method to ﬁ nd the music that you are interested in can no longer meet the needs of users, so the current music recommendation system needs to dig out the music that has no clear needs in the long tail to help people ﬁ nd their favorite songs.


Introduction
The 21st century is a big data era with the rapid development of the Internet and the rapid dissemination of information [1].The Internet has become the main way for people to obtain multimedia resources, including books, movies, and music.In contemporary society, the Internet and electronic information technology have made great progress [2].At the hardware level, electronic chips are becoming smaller and smaller, but their functions are becoming more and more powerful, CPU computing power is rising, and the processing capacity of personal electronic devices is becoming stronger and stronger, which means that it is easier for everyone to create and share new information, such as pictures, videos, and text information [3].Semiconductor storage capacity is becoming stronger and stronger.Server clusters store a large amount of user information and item information.Listening to music, as one of people's main recreational activities, has always played an important role in people's life [4].Ancient sages attached great impor-tance to the important role of music in the development of social civilization and the maintenance of social order [5].They believed that music could help maintain social harmony, that is, the ideal realm of social development of "no resentment when music comes, no struggle when etiquette comes."Due to the rapid growth of music streaming media service industry and the rapid progress of portable device technology, thousands of music has become accessible, but it is becoming more and more difficult to find the music you want [6].The network has brought a lot of information to people, met people's demand for information in the information age, and benefited from it, but it has also brought the problem of information overload.Both consumers and producers of information have encountered great challenges [7].Because of the great variety and amount of information contained in the data, people cannot directly find the information they want.According to this, the predecessors also proposed the use of search engine, which is suitable for such scenarios when users have clear requirements and the requirements can be described by keywords.How to enable users to quickly find the items they are interested in from a large amount of information is becoming more and more important and valuable [8].
The function of recommendation system is to help people find out what they are potentially interested in from the vast amount of information.At present, the scale of the song library of large-scale music portal websites often contains tens of millions of songs, which are divided into different languages, genres, years, themes, moods, and scenes [9] and contain abundant information, and there is a serious information overload.Music retrieval and music recommendation, as the development products of the era of big data, have gradually entered people's daily life and been widely used.After many years of development and improvement, recommendation technology has been widely used in many fields, such as short video, news, stock, and e-commerce.Because of the necessity of music recommendation system in the current society, it is a research hotspot in both industry and academia.In the past, people could only search music by keywords such as music name, singer, and classification, and the search results did not only take into account the differences of users but also led to the phenomenon of long tail of music [10].However, the music recommendation system can help users find the music they want to listen to according to their past behaviors, provide users with a series of song lists, and at the same time increase the sales of digital music.The task of the recommendation system is to act as the link between users and items [11].First, it can make it easier for users to accurately and quickly find items that users are potentially interested in from a large number of items.Second, it can make more items in the item library have the opportunity to be exposed in front of users, so that unpopular items can be explored more.Music retrieval and music recommendation, as the development products of the era of big data, have gradually entered people's daily life and been widely used.With the development of music recommendation algorithm, people put forward higher requirements [12].On the one hand, the problems of traditional recommendation methods such as "cold start" in collaborative filtering need to be solved urgently, and the original recommendation algorithm needs to be upgraded.On the other hand, with the development of machine learning and deep learning, the emergence of new computing technologies is helpful to fully tap the potential preferences of users and improve the performance of recommendation systems [13].
The emergence of recommendation system has a great impact on traditional information retrieval services and traditional Internet services.It can accurately customize and recommend the songs that users like, enhance the retention of users, enhance users' stickiness to the platform, and lay the foundation for the next paid products, thus making the music platform profitable [14].Because of the penetrating power of recommendation algorithm, it is an active mining of the content and information existing in the Internet.With the establishment of this connection, music information retrieval, as a new technology, has brought surprises to the society [15].Although its history is only twenty or thirty years old, as a research field, it has always stood at the forefront of technological development.Different from other recommendation fields, music recommendation has its own characteristics.For example, book recommendation, users want to spend a long time reading a book, and many people will not read it repeatedly, but people choose to listen to a song repeatedly because it takes a short time.Music recommendation system should be able to integrate various factors for real-time adjustment and realize personalized recommendation for different needs of users, which is more complicated than general recommendation system.At present, the related research of music recommendation system mainly focuses on the improvement of recommendation strategies and algorithms, and there are some problems in the recommendation results, such as low accuracy and coverage and lack of personalization.This paper studies music recommendation technology based on deep learning digitalization, optimizes the algorithm of recommendation technology, and provides more possibilities for recommendation system.

Literature Review
Literature [16] proposes that the music retrieval model is people's perception of music works, especially the feeling of music similarity, which is mainly affected by the lyrics, rhythm, the performance of players, or the mental state of users at that time.Literature [17] music recommendation system is regarded as the cross field of music information retrieval and recommendation system.The main goal of music information retrieval technology is to extract the characteristics of different levels of music, so as to retrieve music from all aspects.These features can be audio signal, song name, album name or singer name, etc.The content recommendation of document [18] needs to count the items that users like and then learn the features of these items, extract the user's preferences, and then obtain the user portrait.By calculating the matching degree between the user portrait features and the content features of the items, the users are recommended the items that may be of interest according to the matching degree.Literature [19] through various music labels, the music retrieval system can match the user's needs with the songs in the database, and the music recommendation system can connect the songs in the database, to recommend according to the user's listening history.Literature [20] due to the leading technology, the commercial recommendation system was first successfully deployed by e-commerce giant Amazon, setting off a trend of applying recommendation algorithms in the field of ecommerce.Then, Netflix, a film rental website, applied the recommendation system in the field of film rental with great success.In literature [21] a personalized navigation system which lets us browse based on cooperation is proposed, which marks the start of global personalized service.As an efficient data mining technology, recommendation system has been studied and used in both academic and commercial circles, and various research results are emerging.In literature [22], the characteristics of music content are contained in music signals, such as musical form structure, melody, and rhythm.Music context contains information that cannot be extracted directly from music signals.These 2 Wireless Communications and Mobile Computing information comes from music clips, artists, or players, such as artists' cultural and political background, semantic tags, and album titles.Literature [23] says that with the maturity of machine learning technology and the popularity of deep learning technology, many large companies begin to use machine learning or deep learning related technologies to build recommendation systems, and the traditional collaborative filtering and various rule-based recommendation systems are gradually eliminated.Literature music metadata has many forms, such as manual annotation, social tag, and annotation automatically mined from the web using text retrieval technology.At present, there are many online editing metadata database communities established by music experts or lovers, and the annotation contents are genre, rhythm, emotion, age, and emotion.Literature [24] studies the personalized music recommendation system using content-based recommendation, obtains the user's preference through the rhythm and melody of the user's favorite songs, classifies the candidate songs through the melody preference classifier, and then, recommends the classified songs with similar rhythm and melody to the user, to realize the personalized recommendation of music.
To sum up, it can be seen that most music researches are based on data mining, database establishment, and personalized recommendation system research according to the height of unique performance values.Moreover, recommendation systems are now being applied in various regions.At present, the recommendation system has made great progress in theory and application research, but it has not yet reached a very mature stage.As a frontier exploration field, there are still many problems worthy of in-depth analysis and further discussion.

Related Technology of Music Recommendation System
Music content refers to the information contained in music works, which can represent the music itself.The goal of music content description technology is to automatically extract meaningful features from music.Music recommendation is a special research field in the recommendation system.Its typical application is network music radio.It learns users' preferences from users' historical behaviors such as playing, collecting, and downloading, so as to generate a song playlist for users and push songs that users are satisfied with.Recommendation system is a kind of software or technology that can provide suggestions for users when purchasing and using certain items.Generally speaking, recommendation system is to connect items with users through technical means.Music semantics is the experience of music works after they are listened to by people.Retrieval behavior is a process in which people get relevant music actively or passively, with people's subjective consciousness.Content-based music retrieval is an objective existence starting from the characteristics of music, but when people get the music they really "want," there is an insurmountable "semantic gap," that is to say, it is difficult to directly obtain semantic-related results from music content analysis.The description level of music content is mainly the abstract level, which extracts high-level semantic features from the low-level signal feature description; the second is his timing layer.The content description is related to a certain time range, which may be short-term or calculated by frame, finally, the presentation layer of music, melody, rhythm, chord, music/instrument, rhythm, structure, etc.At the same time, it is also the core basis to distinguish different recommendation system types.
3.1.Convolutional Neural Network.Convolutional neural network (CNN) is a feedforward neural network with convolution calculation and depth structure.Specifically, first, according to the user's previous behavior, obtain the items interacted with it, such as the items that the user has selected or rated, then calculate its preference by extracting the characteristics of these items, then calculate the similarity with each item to be recommended, and finally, recommend it to the user according to the similarity, to recommend the items that may be of interest to it.Music recommendation is different from other recommendations.By studying the characteristics of his recommendation, we can push it more accurately.The trial production of music works is relatively short, so music is likely to be consumed at will, and other data generated by users' hang-up behavior can be eliminated in the recommendation process.Also, users have different ways of tracking users' preferences, and the production cost of movies is high.However, because of its own characteristics, music recommendation does not design to collect explicit feedback, and music recommendation can only acquire users' preferences implicitly.Also, the listening environment of music recommends suitable songs according to the situation; the last one is the emotional connotation of songs.Music can arouse strong personal feelings.By grasping users' recent listening styles, customers can get emotional catharsis.
Because convolutional neural network (CNN) has strong nonlinear fitting ability and has achieved good results in many fields, more and more scholars apply deep learning to the extraction of music recommendation technology.The automatic description of music content is based on computable time-frequency domain signal feature extraction.In order to realize the core function of recommendation system, it is necessary to find useful items, and the recommendation system must be able to predict its recommendation value.Low-level features are extracted directly or indirectly from the frequency representation of music signals.They are easy to be mined by computer system, but they are of little significance to users.Low-level features are the basis of advanced feature analysis, so they should provide an appropriate expression for the studied sound object.CNN is used to predict the implicit features of music, obtain the low dimensional vector representation of music features, and then combine with the implicit representation of user preferences to finally generate reasonable top n recommendations for relevant users.The characteristics of neural network and the connected feature plane used in CNN can allow the image to be used as input directly; it avoids the process of rebuilding the model in the traditional recognition algorithm.The structural law of CNN model is shown in Figure 1.

Wireless Communications and Mobile Computing
Slightly different from convolution in the field of signal processing, the operation in CNN is more like linear weighting operation.For the input image x, convolution is performed using the convolution kernel K of size, and finally, the characteristic image y is output, which can be expressed as: In the formula, b is the bias, p * q is the size of the output feature map, CNN has remarkable advantages in image recognition and voice analysis.The fundamental reason is that the weight sharing network structure reduces the number of parameters and simplifies the complex network structure.
Pool layer, pool operation is another basic operation in CNN.The redundant information extracted in convolution layer is removed to save the most basic and important information, so as to achieve the purpose of dimension reduction.Pooling operation first needs to divide the input convolution feature map to get a number of small local receptive fields and then assign values to the local receptive fields according to the pooled function.For the Jth output feature map a l−1 j in the l-1 layer, the feature map obtained by pooling operation can be expressed as: where b l j b i j is the offset and down ð:Þ is the pooling function.Assuming maximum pooling, the characteristic value of the local receptive field in the graph is the maximum value in the local area of the characteristic graph, while average pooling and summation pooling are the average or summation of the characteristic points in the pooling domain, and the pooling process is similar to this.The amplitude function in traditional segmentation algorithm is defined by the following formula: where AðxÞ represents waveform amplitude function; aðWÞ is the amplitude of the wth sampling point; N is the window length; X is a frame of the input signal, x ∈ ð0, MÞ; and M is the number of input signal frames.Then, the amplitude difference function is: The dividing line of single note with D A ðxÞ is more obvious than that with AðxÞ alone, which is convenient for subsequent processing.

Recommendation Algorithm Based on Collaborative
Filtering.User-based collaborative filtering recommendation is somewhat similar to the principle of "grouping people."The principle of user-based collaborative filtering recommendation algorithm is shown in Figure 2.
According to whether the type of user rating information is direct or indirect, the recommendation system based on collaborative filtering can be divided into two different types of problems.
The first type is direct user rating information, which is called recommendation.Direct user scoring includes digital scoring, sequential scoring, and binary scoring.Domainbased collaborative filtering is a heuristic recommendation method.It is the most basic and core method in the recommendation system.It is also the focus of research and application in academia and business.In this case, we first learn an objective function, which can predict the user's new use and then recommend the items with the highest score for the user.
The second category is indirect user scoring information, which mainly refers to single value scoring types.Wireless Communications and Mobile Computing Collaborative recommendation strategy was born and developed.Its main purpose is to mine the relevance of users or the relevance of the project itself according to the user's preference for the project and then recommend based on these relevance.This kind of problem is called top-N recommendation.Due to the lack of displayed user scores, this kind of recommendation cannot be evaluated correctly.The commonly used evaluation standards are two standards in the field of information retrieval-accuracy rate and recall rate, such as formulas ( 6) and (7).
Among them, the item set is divided into training set and test set, the training set is used to calculate L, and the intersection of the test set and the item set scored by the user constitutes T, namely, TðuÞ ⊂ I u ∩ I tst .LðuÞ is the recommendation list of user u.
Item-based collaborative filtering recommendation is to calculate the similarity between any two items in the system according to the scores of all users and then recommend the items with higher similarity to the items in the list to the target user according to the user's historical preference list of items.The principle of item-based collaborative filtering recommendation algorithm is shown in Figure 3.
(1) Calculate the similarity between any two items according to all user preference behavior information in the system (2) Combining the user's historical behavior and making recommendations to the user according to the similarity of the items

The Effectiveness of Music Recommendation Method
In the research scheme of this paper, the first is through feature extraction.Firstly, each label in the label set is retrieved, a single label is sorted according to the decision value of the label, and the evaluation indexes of all labels are averaged.For the music in the test set, the semantic vector is obtained through the convolution model, and the return value is obtained in the marked corpus set.If the labeling accuracy of the algorithm is high, the original songs manually labeled in the corpus can be returned after retrieval.The accuracy of CNN data set is shown in Figure 4.
It can be seen from the figure that the proportion of the first 10 bits that can return the original song is greater than 92%, indicating that the labeling result of the algorithm is better in the whole song, and can be close to the performance of manual labeling in semantics.However, there are more than 400000 song sets in the music database, so we need to use the user-based collaborative algorithm to collect customer information and select some songs that users have not heard as the total song set.The number of users in the dataset-the distribution of playing songs is shown in Figure 5.
The figure shows that the number of songs of most users is concentrated between 0 and 1000.We try to randomly select users from users with more universal behavior.Users with too few songs have limited data, but very few users have too many songs, and it may be that they do not close the 5 Wireless Communications and Mobile Computing music playback software in time.In order to reduce noise, these user data are not used.I process the experimental data objectively.Using the optimization of the algorithm, the algorithm can calculate the number of features, which can better combine the user's feature recommendation.We use the ROC curve and AUC value of the evaluation index to classify the classification results on the validation set of the three experimental groups, as shown in Figure 6.
It can be seen from the figure that with the increase of the number of our features, the effect is better.It shows that adding the statistical features of users and the audio features of music is conducive to the classifier to judge users' preferences.At the same time, it can help the model find out the potential reasons why users will like the music.
The recommendation algorithm based on the combination of item collaborative filtering and interest tag proposed above mainly uses the user's behavior to calculate the song similarity offline for preference and then quickly and accurately find the candidate similar song collection that the user is most likely to be interested in from the ten million level song library according to the song similarity.The experimental data set is the similar candidate song set calculated by hybrid recommendation, and the data is assembled according to the input format required by the deep neural network.In general, user-based collaborative filtering recommendation is more social.The recommended items are popular items in the neighborhood of the target user, while item-based collaborative filtering recommendation is more personalized, because the recommended items generally meet their own interests and preferences.Using the above formula, we tested the two software.We focus different L values on 0.5-0.7,as shown in Figures 7 and 8.
It can be seen from the figure that when L = 0:5, the data obtained is small, and the recommendation is more accurate; when L = 0:6 and L = 0:7, the data can also be more accurate, but compared with the value of 0.5, too many candidate songs will be filtered out, resulting in a short candidate list.
Different recommendation models will have different superior performance.CNN and collaborative filtering algorithms mentioned in this paper are the most commonly used algorithms at present, besides which there are different recommendation algorithm models such as Frunk-SVD, User-CF, and CB.Frunk-SVD is the most basic matrix decomposition method of implicit semantic model.Random gradient descent method is used to find the optimal solution.Finally, by completing the matrix, the user's rating of items can be predicted, so as to achieve recommendation.CB is to recommend items that are similar to those that users liked in the past according to the items that users liked in the past.In order to make the structure of the article more rigorous, the accuracy of three kinds of data is tested as shown in Figure 9.
The data show that the two algorithms mentioned in this paper have higher accuracy than frunk SVD and CB.CNN and collaborative filtering algorithm can improve the cold  6 Wireless Communications and Mobile Computing start of traditional recommendation algorithms, supplement available information sources for music recommendation system, and improve the performance of recommendation system as a whole.

Conclusions
Music is a popular artistic expression.The development of digital music industry puts forward higher requirements for music information retrieval.People cannot live without music.Music contains a huge library of songs, so it is impossible for users to listen to every song in the library.Moreover, in many cases, users' vague demand for music is just "one or several nice songs."In the era of information overload, recommendation system can act as the link between users and items and can help users find items that may be of interest without clear requirements.Music recommendation system is a research topic with broad application scenarios and practical application value.Improving the effect of music recommendation can not only promote users' experience but also greatly improve the profits of music streaming media service providers.The traditional music information retrieval based on text metadata has been unable to meet people's increasing retrieval requirements.
Considering the particularity of the music recommendation system, this paper studies the content-based music recommendation system, tries different methods for two important links of the content-based music recommendation system, and improves the method of extracting music audio features based on deep learning, which has been effectively promoted.In the recommendation link, the traditional way of calculating the similarity of songs or the similarity between songs and users is not used for recommendation, and the effectiveness of the recommendation is proved from objective and subjective perspectives.Due to the time problem, the content is not very comprehensive, and the follow-up research will continue.

Figure 6 : 7 Figure 7 :
Figure 6: ROC curves of three groups of experimental results.