CN108595550A - A kind of music commending system and recommendation method based on convolutional neural networks - Google Patents

Abstract

The invention provides a music recommendation system and recommendation method based on a convolutional neural network, including a music user modeling module for collecting historical behavior data of music users and constructing a music user preference model; a music feature building module for obtaining a regression model Modular module; a recommendation algorithm module for finding music objects that match music user preferences through a regression model and recommending them to music users. The invention applies deep learning to the recommendation system, effectively compensates for the semantic difference between song features and audio signals, avoids problems such as "cold start" in collaborative filtering, and improves the accuracy of the recommendation system. Convolutional neural network is used to solve the contradiction between low training efficiency and high timeliness requirements, and the user's historical behavior information and audio acoustic features are added to the model, so that the recommendation results are more in line with the user's preference needs, and the user's awareness of the problem is enhanced. The user experience of the recommendation system.

Description

A music recommendation system and recommendation method based on convolutional neural network

technical field

The invention belongs to the technical field of intelligent recommendation, and in particular relates to a music recommendation system and a recommendation method based on a convolutional neural network.

Background technique

With the continuous development and application of digital multimedia technology, digital music has been appreciated and loved by the public, and people can conveniently obtain music resources through various methods such as online audition and online download. However, as the music library becomes larger and the music resources become more and more abundant, how to allow users to efficiently obtain the songs they are interested in in the vast music world has become a difficult problem. In response to this phenomenon, a personalized music recommendation system came into being.

At present, the commonly used music recommendation methods mainly include content-based recommendation, collaborative filtering recommendation, association rule-based recommendation, and hybrid recommendation. The most widely used is collaborative filtering recommendation, which includes two methods based on neighbors and based on models. Neighbor-based music recommendation methods are usually used to study the similarity calculation problem between users or music, that is, by calculating the interest similarity between users and neighbors, recommending music that neighbors have listened to to users, or by analyzing user listening behavior records. Music, calculate the similarity between different music based on user behavior. Whereas, model-based music recommendation methods try to convert user-music relationship evaluation data into different models and use these models to recommend music to users in unknown scenarios.

Although there are many types of existing music recommendation systems, the recommendation effects are uneven, and there are more or less typical problems, such as cold start, sparsity, and scalability. Therefore, the music recommendation system still needs further research and improvement to better adapt to the increasingly strong personalized application requirements. Convolutional neural network (CNN) is an efficient recognition method developed in recent years and has attracted widespread attention. It is widely used in different large-scale machine learning problems such as speech recognition, image recognition and natural speech processing, and has achieved comparative results. Traditional machine learning has better effect improvement.

Contents of the invention

In order to solve the deficiencies of the prior art, the present invention provides a music recommendation system and recommendation method based on a convolutional neural network.

The purpose of the present invention is achieved through the following technical solutions:

A music recommendation system based on convolutional neural network, including

Music user modeling module: used to collect historical behavior data of music users and build music user preference models;

Music feature modeling module: the music feature modeling module is electrically connected to the music user modeling module, and is used to train a deep convolutional neural network to obtain a regression model for training samples;

Recommendation algorithm module: the recommendation algorithm module is electrically connected to the music feature modeling module, and is used to find music objects that match music user preferences through a regression model and recommend them to music users.

Preferably, the above-mentioned recommendation method of a music recommendation system based on a convolutional neural network comprises the following steps:

S1. The music user modeling module acquires the behavior data of music users, constructs the actual scoring matrix R through quantitative standards, and obtains the user-latent factor matrix Q and latent factor-music matrix P through matrix decomposition;

S2, using the Mel Frequency Cepstral Coefficient (MFCC) method to perform time-frequency audio feature extraction on the music in the data set used, thereby obtaining the Mel Spectrum of the audio segment;

S3. Use the Mel acoustic spectrum as the network input of the convolutional neural network, use the music feature factor vector obtained by decomposition in S1 as the basic truth value of the training prediction model, and train the deep convolutional neural network through a series of training samples, and continuously Reduce the mean square error (MSE) between the latent factor vector and the audio feature prediction value, and finally form a regression model;

S4. The test sample i predicts the latent factor-music vector p _i ′ through the above regression model, and performs an inner product operation with the user-latent factor vector q _u decomposed in S1 to obtain the user u’s estimated score for song i After calculating the estimated scoring matrix composed of all test samples After that, the music with higher score is selected and recommended to the corresponding user.

Preferably, the user's behavior data in S1 is a record of the user's past operations on a certain piece of music obtained through a web crawler, and the operations include but are not limited to single track looping, sharing, favorites, skipping, black.

Preferably, the matrix decomposition in S1 is a weighted matrix decomposition method.

Preferably, the time-frequency audio feature extraction in the S2 includes steps such as pre-emphasis, framing, windowing, fast Fourier transform (FFT), Mel filter bank, discrete cosine transform (DCT), and the training set The parameters such as frame length, window function, frame shift and the number of Mel filters are the same as the test set.

Preferably, the frame length is 256 sampling points, the frame shift is half of the frame length, the window function is a Hamming window with a=0.46, and the number of Mel filters is 24.

Preferably, the deep convolutional neural network in S3 is composed of 4 convolutional layers, pooling layers and 1 fully connected layer.

The beneficial effects of the present invention are reflected in: 1. The present invention applies deep learning to the recommendation system, and predicts the potential factors of music audio by training the deep convolutional neural network, effectively making up for the semantic difference between song features and audio signals , while avoiding problems such as "cold start" in collaborative filtering, and improving the accuracy of the recommendation system.

2. Compared with other deep learning models, the convolutional neural network used in the present invention has fewer parameters and higher training efficiency, which solves the contradiction between low training efficiency and high timeliness requirements, and at the same time integrates user historical behavior information and audio acoustics The features are added to the model together to make the recommendation results more in line with the user's preferences and enhance the user's experience in using the recommendation system.

Description of drawings

Fig. 1: The structural block diagram of the music recommendation system of the present invention.

Figure 2: Flow chart of the training process of the convolutional neural network model of the present invention.

Detailed ways

The present invention discloses a music recommendation system and its recommendation method based on convolutional neural network. As shown in Fig. 1-Fig. A music user modeling module, a music feature modeling module electrically connected to the music user modeling module, and a recommendation algorithm module electrically connected to the music feature modeling module. The music feature modeling module is used to train a deep convolutional neural network on training samples to obtain a regression model. The recommendation algorithm module is used to find music objects that match the preferences of music users through a regression model, and recommend them to music users.

A kind of recommendation method based on the music recommendation system of convolutional neural network described above, comprises the following steps:

S1. The music user modeling module acquires the behavior data of music users, and constructs the actual scoring matrix R through quantitative standards, and obtains two low-dimensional user-latent factor matrices Q and latent factors through weighted matrix decomposition matrix decomposition -Music matrix P, which respectively represents the degree of preference of different users for different elements and the composition of various elements in each type of music. The user's behavior data is a record of the user's past operations on a certain piece of music obtained through a web crawler, and the operations include but are not limited to single track looping, sharing, favorites, skipping, and blacklisting. The preferred quantification standard in the present invention is set as: single cycle = 5, share = 4, favorite = 3, active play = 2, finish listening = 1, skip = -2, block = -5.

S2, utilize Mel Frequency Cepstral Coefficient (MFCC) method to carry out time-frequency audio feature extraction to the music in used data set, thereby obtain the Mel sound spectrum of audio clip; Described time-frequency audio feature extraction comprises pre-emphasis, framing , windowing, fast Fourier transform (FFT), Mel filter bank, discrete cosine transform (DCT) and other steps, and the training set and test set use the same frame length, window function, frame shift and Mel filter number and other parameters; the training set and test set are obtained by random distribution of music resource data sets in the system. Usually the audio signal is in the range of 10-30ms, and its characteristics remain basically unchanged. When dividing frames, it is necessary to select an appropriate frame length; at the same time, in order to avoid excessive changes in two adjacent frames, resulting in incoherent audio information, two adjacent frames There needs to be an overlapping area between them, that is, frame shift; each frame is multiplied by a suitable window function, which can increase the continuity of the left and right ends of the frame; the frequency spectrum can be smoothed and harmonics can be eliminated by using a filter bank. In the present invention, the frame length is preferably 256 sampling points, the frame shift is half of the frame length, the window function is a Hamming window of a=0.46, and the number of Mel filters is 24. The above parameters can be changed according to specific needs.

S3. Use the Mel acoustic spectrum as the network input of the convolutional neural network, use the music feature factor vector obtained by decomposition in S1 as the basic truth value of the training prediction model, and train the deep convolutional neural network through a series of training samples, and continuously Reduce the mean square error (MSE) between the latent factor vector and the audio feature prediction value until the required training conditions are reached, and finally form a regression model; after continuous adjustment and optimization, the deep convolutional neural network uses multi-layer convolution layer in order to be able to extract deeper feature information, and reduce the dimensionality of the extracted feature information through the pooling layer, which simplifies the complexity of network calculations and helps prevent overfitting. In the present invention, it is preferred to adopt a deep convolutional neural network consisting of 4 convolutional layers, pooling layers and 1 fully connected layer. To speed up model training, the input is the mel spectrogram of 1-second audio clips randomly clipped from the audio. For longer audio clips, it needs to be divided into several 1-second durations and the window prediction values should be averaged. And the learning rate is set to 0.001, and the ReLu activation function and Adam gradient descent optimization algorithm are used.

S4. The test sample i predicts the latent factor-music vector p _i ′ through the above regression model, and performs an inner product operation with the user-latent factor vector q _u decomposed in S1 to obtain the user u’s estimated score for song i After calculating the estimated scoring matrix composed of all test samples After that, the music with higher score is selected and recommended to the corresponding user.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it can still be Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent replacements are made to some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims (7)

1. a kind of music commending system based on convolutional neural networks, it is characterised in that：Including

Music user modeling module：For acquiring music user's history behavioral data, the preference pattern of music user is built；

Musical features modeling module：The musical features modeling module is electrically connected with the music user modeling module, is used for Depth convolutional neural networks are trained to training sample and obtain regression model；

Proposed algorithm module：The proposed algorithm module is electrically connected with the musical features modeling module, is returned for passing through Model finds the music object to match with music user preference, recommends music user.

2. a kind of recommendation method of the music commending system based on convolutional neural networks as described in claim 1, feature exist In：Include the following steps：

S1, music user modeling module obtain the behavioral data of music user, and construct reality by quantitative criteria Rating matrix R, and user-latent factor matrix Q and latent factor-music matrix P are obtained by matrix decomposition；

S2, T/F audio feature extraction is carried out to music in data set used using mel-frequency cepstrum coefficient method, to Obtain the Meier sound spectrum of audio fragment；

S3, using Meier sound spectrum as the network inputs of convolutional neural networks, will be decomposed in S1 the musical features factor of acquisition to Amount is used as the basic true value of training prediction model, trains depth convolutional neural networks by a series of training samples, constantly reduces Mean square deviation between latent factor vector and audio frequency characteristics predicted value, ultimately forms regression model；

S4, test sample i go out latent factor-music vector p by above-mentioned forecast of regression model_i', and obtained with being decomposed in S1 User-latent factor vector q_uInner product operation is carried out, user u is obtained and scores the estimation of song iIt is calculating The estimation rating matrix that all test samples are constitutedAfterwards, the selection higher music of score recommends corresponding user.

3. a kind of recommendation method of the music commending system based on convolutional neural networks as claimed in claim 2, feature exist In：The behavioral data of user carried out certain song for the user that is obtained by way of web crawlers in the past in the S1 Operation note, it is described operation include but not limited to single recycle, share, collecting, skipping, draw it is black.

4. a kind of recommendation method of the music commending system based on convolutional neural networks as claimed in claim 2, feature exist In：Matrix decomposition is weighting matrix decomposition method in the S1.

5. a kind of recommendation method of the music commending system based on convolutional neural networks as claimed in claim 2, feature exist In：In the S2 T/F audio feature extraction include but not limited to preemphasis, framing, adding window, Fast Fourier Transform (FFT), Meier filter group, discrete cosine transform step, and training sample and test sample use identical parameter, the parameter packet Include frame length, window function, frame shifting and Meier number of filter.

6. a kind of recommendation method of the music commending system based on convolutional neural networks as claimed in claim 5, feature exist In：The frame length is 256 sampled points, and frame moves the half for frame length, and window function is the Hamming window of a=0.46, Meier filtering The number of device is 24.

7. a kind of recommendation method of the music commending system based on convolutional neural networks as claimed in claim 2, feature exist In：Depth convolutional neural networks described in the S3 are made of 4 convolutional layers and pond layer and 1 full articulamentum.