CN117935859A - Multi-label music style feature selection method integrating feature similarity - Google Patents

Abstract

The invention belongs to the field of machine learning and pattern recognition, and relates to a multi-label music style feature selection method integrating feature similarity. The invention combines feature relativity to construct a multi-label music style feature selection method, which divides candidate music style labels into real music style labels and noise music style labels, then respectively utilizes feature similarity to limit coefficient matrix W mapped to the real labels and coefficient matrix S mapped to the noise labels, utilizes a multi-label classifier to limit sum matrix H of W and S, and finally, utilizes the W matrix to score the features to select related music style features. The problem difficulty is reduced, and the performance of the multi-label k nearest neighbor (MLKNN) model is improved.

Description

Multi-label music style feature selection method integrating feature similarity

Technical Field

The invention belongs to the field of machine learning and pattern recognition, and relates to a method for selecting multi-label music style features by integrating feature similarity.

Background technique

Music style classification is a way to summarize and distinguish musical works, dividing different types of music by specific musical elements and style characteristics. This classification system helps to understand and describe the diversity of music and helps listeners better choose the music they like. In the music style classification, there are many different genres and types, each with a unique sound and expression. Classical music is a form of music that expresses emotions and thoughts through complex structures and exquisite arrangements, while pop music focuses more on popular and easy-to-understand melodies and lyrics, and is the mainstream music loved by the public. In addition, jazz is known for its complex harmony and improvisation, and electronic music creates futuristic sounds through electronic devices and technology. Folk music is usually guitar-based and emphasizes storytelling and authenticity, while rock music features strong rhythms and guitar playing to express attitudes towards society and individuals. Overall, music style classification is a rich and colorful field that reflects the unique understanding and creation of music by different cultures, eras, and individuals. This classification system helps us better appreciate and understand the diversity of music, allowing us to find our favorite sounds in the vast world of music. Music style classification adds corresponding labels to music according to its style. Through music style labels, music platforms can better recommend music to interested users and improve the user experience. Multi-label classification is also commonly used in music style classification. In multi-label music style classification, a song may be labeled with styles such as "metal", "punk", "rock", and "pop". In the actual data set collection process, it is often impossible to obtain completely accurate style labels. Because most of the actual collected data is crawled from the Internet, and there are usually some unreliable labelers in the network, which inevitably leads to some labeling errors. This means that some labels crawled from the Internet during data collection have some human factors that cause a song to be labeled with a style that does not belong to it, and only the labels given by some of the labelers are valid. For example, a song or a piece of music may have seven labels on the Internet, namely "folk", "electronic", "pop", "independent", "jazz" and "classical". However, a professional who carefully identifies this piece of music can find that there are many errors in these labels. Among them, only "folk", "electronic", "pop" and "independent" are valid labels. We call classification in a multi-label music style classification dataset with incorrect labels as partial multi-label music style classification.

In the traditional multi-label learning task of music style classification, the accuracy of the learning algorithm can often be effectively improved by reducing the extracted music features using feature selection methods. Previous theories and practices have shown that using appropriate feature selection methods can effectively reduce the difficulty of learning tasks and improve the accuracy of the model. Multi-label feature selection technology reduces the time and hardware resources consumed by the learner by reducing redundant music features and removing music features that are irrelevant to the classification task. And because irrelevant and redundant features are eliminated, the influence of these features on the multi-label learner is removed, and the learning performance is improved.

In the scenario of multi-label music style classification, directly using the multi-label feature selection algorithm ignores the impact of noise labels on the feature selection results, resulting in the wrong selection of features related to noise music style labels. The features selected in this way lack credibility. There are few existing multi-label feature selection methods. One algorithm is to divide the candidate labels of multi-label into two parts, and respectively map the coefficient matrix from the feature space to the label space through the nuclear norm and l ₁ norm restrictions, and then use the manifold regularization to make similar instances have similar labels.

Summary of the invention

In view of the current technical deficiencies, the present invention proposes a multi-label music style feature selection method integrating feature similarity.

The specific steps of the present invention are as follows:

Step 1: Extract music features to obtain a multi-label music classification dataset M and specify the feature subset dimension K, where the set M contains n music samples, q labels, and d music features;

Step 2: Divide the multi-label music classification dataset M into a training sample set MT and a test sample set MP. Here, X represents the feature matrix of the training sample set, (X) _ij represents the jth eigenvalue of the i-th sample, Y represents the candidate label indicator matrix of the training sample, and (Y) _ij represents whether the candidate label of the j-th sample has the i-th label, which is 1 if it exists and 0 if it does not exist.

Step 3: Calculate the local Gram matrix of each sample _xg to represent feature similarity. The local Gram matrix _Wg∈Rdxd of sample _xg ^is calculated as follows:

in, δ is the threshold of neighborhood granularity, and Δ is the distance metric formula.

Step 4: Define a partial multi-label classifier with the following objective function

Among them, W and S are the mapping matrices mapped to the true label and the noise label, respectively.

Step 5: Restrict the feature similarity to the mapping matrices W and S respectively. Here, we use the idea of popular regularization to define the final objective function

Wherein, L _g = D _g + W _g , and D _g is a diagonal matrix formed by adding the elements of each row of W _g .

Step 7: Use the alternating solution method to minimize the objective function to solve W and S, calculate the bi-norm of each column vector of W to get the feature score, and then select the largest K features.

Step 8: Use the selected K features to reduce the dimensions of the training sample set MT and the test sample set MP, and obtain the reduced-dimensional training sample set MT′ and the reduced-dimensional test sample set MP′ respectively. Then, input the reduced-dimensional training sample set MT′ into the multi-label k-nearest neighbor (ML-KNN) model for training to obtain the trained multi-label k-nearest neighbor model (ML-KNN) model.

The beneficial effect of the present invention is as follows: the present invention integrates feature similarity to construct a partial multi-label music style feature selection method, including: preprocessing a multi-label data set, including missing value filling, data discretization, etc.; using the partial multi-label music style feature selection method to perform feature screening on the processed data set to obtain a screened feature set. The obtained feature data set is input into a multi-label k-nearest neighbor (MLKNN) model to obtain a multi-label k-nearest neighbor (MLKNN) model after the data set is optimized. The present invention divides the candidate music style labels into real music style labels and noise music style labels, and then uses feature similarity to restrict the coefficient matrix W mapped to the real label and the coefficient matrix S mapped to the noise label, respectively, and uses a partial multi-label classifier to restrict the sum matrix H of W and S, and finally, the W matrix is used to score the features and select relevant music style features. It reduces the difficulty of the problem and improves the performance of the multi-label k-nearest neighbor (MLKNN) model.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a flow chart of the present invention.

Detailed ways

The specific implementation modes of the present invention are described in detail below.

Step 1: Extract music features to obtain a multi-label music classification dataset M and specify the feature subset dimension K, where the set M contains n music style samples, q music style labels, and d music features;

Step 2: Divide the multi-label music style classification dataset M into a training sample set MT and a test sample set MP. Here, X represents the feature matrix of the training sample set MT, (X) _ij represents the jth eigenvalue of the i-th sample, Y represents the candidate label indicator matrix of the training sample, and (Y) _ij represents whether the candidate label of the j-th sample has the i-th label, 1 if it exists, and 0 if it does not exist;

Step 3: Calculate the local Gram matrix of each sample to represent feature similarity. The local Gram matrix W _g ∈ R ^dxd of sample x _g is calculated as follows:

in, δ is the threshold of neighborhood granularity, and Δ uses Euclidean distance.

Step 4: Define a partial multi-label classifier with the following objective function

Among them, W and S are the mapping matrices mapped to the true label and the noise label, respectively.

Step 5: Restrict the feature similarity to the mapping matrices W and S respectively. Here, we use the idea of popular regularization to define the final objective function

Where L _g = D _g + W _g , D _g is the diagonal matrix formed by adding the elements of each row of W _g . λ, β, γ, η ₁ and η ₂ are balance parameters. Use the alternating solution method to minimize the objective function to solve W, calculate the bi-norm of each column vector of W to obtain the feature score, and then select the largest K features.

Step 6: Use the selected K features to reduce the dimensions of the training sample set MT and the test sample set MP, and obtain the reduced-dimensional training sample set MT′ and the reduced-dimensional test sample set MP′ respectively. Then, input the reduced-dimensional training sample set MT′ into the multi-label k-nearest neighbor (ML-KNN) model for training to obtain the trained multi-label k-nearest neighbor model (ML-KNN) model.

The λ, β, γ, _η1 and _η2 parameters in step 5 are obtained by cross validation.

The specific solution process in step 5 is as follows: First, the objective function is converted into

Using the LADMAP method, the above formula is converted into

Fixed W, S, P, Q, optimized H

H _k+1 =(Y ^T X + μ ₁ S _k + μ ₁ W _k - Y ₁ )(XX ^T + λ I + μ ₁ I) ^-1

Fixed W, H, S, Q, optimized P

Fixed W, H, S, P, optimized Q

Fixed H, P, Q, optimized W and S

is a diagonal matrix, />

_Sk+1 = soft _2γ /( _μ1 + _μ2 )( _μ3Q + _Y3 + _μ1H - _μ1W + _Y1 ), where soft _ξ (x) = sign(x)max(|x|-ξ, 0)

Update Y ₁ , Y ₂ , Y ₃ , μ ₁ , μ ₂ , μ ₃

Wherein μ _max is the maximum value limited by μ ₁ , μ ₂ , and μ ₃ , and ρ is the growth coefficient.

Verification example:

In order to verify the effectiveness of this application, we verified it on the music_styles dataset, which is a dataset for partial label music style classification, which annotates and predicts the style of music. It includes 10 music style labels such as "classical", "jazz", and "rock", and consists of 6839 pieces of music. Each piece of music has 98 features and is marked as one of the 10 labels. The original labels collected by the network are candidate labels, and the real labels are obtained through manual fine resolution.

According to the steps of the present invention, λ is set to 1, β is set to 1, γ is set to 0.1, η ₁ is set to 0.01, and η ₂ is set to 0.01. At this time, the input set M is music_styles, and the input feature subset dimension K is 24. Finally, the selected feature numbers are {65, 66, 26, 4, 54, 64, 96, 81, 92, 27, 79, 52, 29, 49, 1628, 58, 1, 5, 95, 55, 57, 9, 93}, and then a new training set is created according to the selected feature set, and finally the new training set is used to train the MLKNN classifier model to obtain the model MLKNN-FS.

One Error, Ranking Loss, Coverage Error, and Average Precision are used as criteria for judging multi-label classification models. Next, in the comparative verification experiment, the MLKNN model is trained directly with the complete training set without feature selection to obtain the model MKLNN-ALL. Substitute it into the test set to obtain the four indicators of the MLKNN-ALL model. The above data is aggregated into the following table:

Table 1 Comparison of four indicators of prediction results of model MLKNN-ALL and model MLKNN-FS

The Average Precision index in Table 1 is as large as possible, while the Coverage Error, One Error and Ranking Loss indexes are as small as possible. The experimental results show that the MLKNN-FS classifier is better than the MLKNN-ALL classifier in many indicators. This shows that the present invention can effectively improve the performance of the multi-label classification model.

Claims (3)

1. A method for selecting multi-label music style features by integrating feature similarity, comprising the following steps: Step 1: Extract music features to obtain a multi-label music classification dataset M and specify the number of selected features K, where the set M contains n music samples, q labels, and d music features; Step 2: Divide the multi-label music classification dataset M into a training sample set MT and a test sample set MP. Here, X represents the feature matrix of the training sample set, (X) _ij represents the jth eigenvalue of the i-th sample, Y represents the candidate label indicator matrix of the training sample, and (Y) _ij represents whether the candidate label of the j-th sample has the i-th label, which is 1 if it exists and 0 if it does not exist. Step 3: Calculate the local Gram matrix of each sample _xg to represent feature similarity; the local Gram matrix _Wg∈Rdxd of sample _xg ^is calculated as follows: in, δ is the threshold of neighborhood granularity, Δ is the distance metric formula; Step 4: Define a partial multi-label classifier with the following objective function Among them, W and S are the mapping matrices mapped to the true label and the noise label respectively; Step 5: Restrict the feature similarity to the mapping matrices W and S respectively. Here, we use the idea of popular regularization to define the final objective function Where L _g = D _g + W _g , D _g is the diagonal matrix formed by adding the elements of each row of W _g ; λ, β, γ, η ₁ and η ₂ are balance parameters; the alternating solution method is used to minimize the objective function to solve W and S, the bi-norm of each column vector of W is calculated to obtain the feature score, and then the largest K features are selected; Step 6: Use the selected K features to reduce the dimensions of the training sample set MT and the test sample set MP, and obtain the reduced-dimensional training sample set MT′ and the reduced-dimensional test sample set MP′ respectively. Then, input the reduced-dimensional training sample set MT′ into the multi-label k-nearest neighbor (ML-KNN) model for training to obtain the trained multi-label k-nearest neighbor model (ML-KNN) model. 2. According to the method for selecting multi-label music style features by integrating feature similarity according to claim 1, it is characterized in that in step 5 First, the objective function is transformed into Using the LADMAP method, the above formula is converted into Fixed W, S, P, Q, optimized H H _k+1 =(Y ^T X + μ ₁ S _k + μ ₁ W _k - Y ₁ )(XX ^T + λ I + μ ₁ I) ^-1 Fixed W, H, S, Q, optimized P Fixed W, H, S, P, optimized Q Fixed H, P, Q, optimized W and S is a diagonal matrix, /> where soft _ξ (x) = sign(x)max(|x|-ξ, 0) Update Y ₁ , Y ₂ , Y ₃ , μ ₁ , μ ₂ , μ ₃ Wherein μ _max is the maximum value limited by μ ₁ , μ ₂ , and μ ₃ , and ρ is the growth coefficient. 3. According to the method for selecting multi-label music style features based on feature similarity fusion according to claim 1, it is characterized in that in step 6, the step of training the MLKNN classifier comprises: The newly generated feature subset is input into the MLKNN model. At this time, the number of parameters k of the MLKNN model is 10, and other parameters remain default, and finally the optimized MLKNN model is obtained.