CN107845424B - Method and system for diagnostic information processing analysis - Google Patents

Abstract

The invention discloses a method and a system for processing and analyzing diagnostic information, and relates to the technical field of medical systems. The method for processing and analyzing diagnostic information includes the following steps: establishing a feature space and a label space by using a plurality of sample features and label information, and establishing a feature vector and a first label vector of the sample; calculating the number of occurrences of each label information; calculating every two The number of times that each label information appears in a sample at the same time; calculate the similarity of each two label information, and establish a label-label similarity matrix; reconstruct the first label vector to obtain the second label vector, and calculate according to the second label vector The occurrence probability of the target sample label information; a preset number of label information is selected as the recommended label information of the target sample. The invention can effectively utilize the correlation between the disease and its complications to discover more potential diseases and improve the accuracy of diagnosis decision support by using the features corresponding to the diseases of the patients to form the feature space and label information.

Description

Method and system for processing and analyzing diagnostic information

technical field

The present invention relates to the technical field of medical systems, in particular to a method and system for processing and analyzing diagnostic information.

Background technique

In the era of big data, people have gradually embraced the use of health big data to assist in diagnosis and treatment. With the success of many industrial applications in using big data, the health service industry has also begun to use medical big data to improve service efficiency and quality.

Health information tools and machine learning techniques have been successfully used to help doctors diagnose diseases and develop treatment plans more efficiently. Clinical decision support applications include systems that provide diagnosis, personalized drug evaluation, treatment plans, and relevant medical knowledge. Clinical decision support applications aim to provide medical staff with professional knowledge, patient information and intelligent means to improve the efficiency and effectiveness of medical staff decision-making. Through clinical decision support applications, medical errors can be reduced and the quality of medical services can be improved. In the medical field, there is an increasing demand for high-quality clinical decision support systems.

Clinicians use their experience and knowledge to differentiate and diagnose patients. Therefore, if clinicians do not have rich experience and accurate judgment, it will cause inevitable medical errors. The goal of establishing a clinical decision support system is to improve the accuracy and efficiency of clinicians through machine learning technology. The system can extract patient characteristics from personal health records, such as physiological data, electronic medical records, 3D images, radiological images, genome sequencing, clinical and billing data, etc., classify patients according to their characteristics, and provide corresponding clinical recommendations to doctor. Scoring criteria in medical scenarios and the complexity of the medical field are difficult problems for clinical decision support systems. Currently, many clinical decision support systems have been developed in the market to assist physicians.

Because some diseases have complications, it is common for a patient to have multiple diseases at the same time. In order to estimate the reference disease to the clinician, the clinical support decision system needs to be more complex. Analysis of real clinical diagnoses revealed that the number of patients with multiple diseases at the same time accounted for a large proportion of all patients. Therefore, the clinical decision support system needs to recommend multiple reference diseases to clinicians. Therefore, recommending diseases is transformed into the problem of multi-label classification of reference diseases.

Due to the simple steps and outstanding effects of ML-kNN (lazy multi-label classification method), this algorithm has been widely used and studied. However, this algorithm ignores the association between labels by estimating the likelihood of each label independently. However, in the actual diagnosis of diseases, many labels are related. For the application scenarios where the labels are related, the ML-kNN method lacks effectiveness.

SUMMARY OF THE INVENTION

The main purpose of the present invention is to provide a method and system for processing and analyzing diagnostic information, aiming at effectively improving the accuracy of diagnosing diseases by utilizing the correlation between diseases.

To achieve the above object, the present invention provides a method for processing and analyzing diagnostic information, comprising the following steps:

A feature space and a label space of multi-label information are established by multiple sample features and label information of multiple samples, and a feature vector and a first label vector of the sample are established according to each sample feature and each label information;

Calculate the number of times of occurrence of each label information; calculate the number of times that every two label information appears in a sample at the same time;

Calculate the similarity of each two tag information, and establish a tag-tag similarity matrix;

Reconstructing the first label vector of the sample through the label-label similarity matrix to obtain a second label vector, and calculating the occurrence probability of label information of the target sample according to the second label vector;

Sort the probability of occurrence of the target sample label information in descending order, and select a preset number of label information as the recommended label information of the target sample.

Preferably, the feature space and label space for establishing multi-label information by using multiple sample features and label information of multiple samples further include:

The preset F={f ₁ , f ₂ , ... f _b } is the b-dimensional feature space of multiple label information, and the preset L={l ₁ , l ₂ , ... l _q } is the multiple label information q-dimensional label space;

为所述样本的b维特征向量；则The preset T={(X ₁ ,Y ₁ ),(X ₂ ,Y ₂ ),...,(X _n ,Y _n )} is a set of multiple label information, the preset

is the b-dimensional feature vector of the sample; then

为所述样本X_i对应的标签向量；若特征向量X_i有标签空间l_j，则标签向量

否则

is the label vector corresponding to the sample X _i ; if the feature vector X _i has a label space l _j , then the label vector

otherwise

Preferably, the calculating the number of occurrences of each label information includes: calculating the sample corresponding to the label information according to the number of times the label information appears in each sample, and setting the value as r _ij ; if the feature vector If there is a label space _li in X _k , then ri _ij =1, otherwise ri _ij =0.

Preferably, the calculation of the similarity of every two tags and the establishment of a tag-tag similarity matrix further include:

Use the cosine similarity calculation method to calculate the similarity of every two tags.

Preferably, calculating the similarity of every two tags by using the cosine similarity calculation method includes:

计算每两个标签的相似度，by calculation formula

Calculate the similarity of every two labels,

为标签空间l_i与标签空间l_j的同时出现在样本X_k中的次数；

和

分别为标签空间l_i出现的总次数和标签空间l_j出现的总次数。Among them, P _ij is a set including both label space _{li and label space l j ,}

is the number of times the label space l _i and the label space l _j appear in the sample X _k at the same time;

and

are the total number of occurrences of label space _li and the total number of occurrences of label space _lj , respectively.

Preferably, calculating the similarity of each two tag information, and establishing a tag-tag similarity matrix includes:

The label-label similarity matrix is

Wherein, the element S _ij =sim(I _i , I _j ) in the matrix represents the similarity between the label l _i and the label l _j .

Preferably, the label information matrix of the sample reconstructed by the label-label similarity matrix is: Y=g(X),

where g(X) is:

Preferably, the calculating the occurrence probability of the label information of the target sample according to the second label vector includes:

According to the second label vector, a lazy multi-label classification algorithm is used to calculate the occurrence probability of each label information in the target sample.

The present invention also provides a diagnostic information processing and analysis system, the system comprising:

A module for establishing a feature space and a label space of multi-label information according to multiple sample features and label information of multiple samples;

A module for establishing a feature vector and a first label vector of the sample according to each sample feature and each label information;

A module for calculating the number of occurrences of each label information;

A module for calculating the number of times that each two label information appears in a sample at the same time;

A module for calculating the similarity of each two tag information and establishing a tag-tag similarity matrix;

A module for reconstructing the label information matrix of the sample by the label-label similarity matrix, reconstructing the first label vector of the sample to obtain a second label vector, and for calculating the target sample according to the second label vector The module of the probability of occurrence of label information;

A module used to sort the probability of occurrence of the target sample label information in descending order'

A module for selecting a preset number of label information as the recommended label information for the target sample.

Preferably, the system further includes: the module for calculating the similarity of each two tag information, and establishing a tag-tag similarity matrix is a cosine similarity calculation module.

The technical scheme of the present invention uses the characteristics corresponding to the disease of the patient to form the feature space in the multi-label learning algorithm, and uses the disease as the label information in the multi-label learning algorithm, and calculates and analyzes the label information to obtain each label information. The number of times the two label information co-occurs on a patient, and the label matrix is used to reconstruct the label matrix to update the label vector corresponding to each patient, and then the multi-label learning algorithm is used to predict possible diseases for the target patient. Effectively use the association of diseases with their complications to discover more potential diseases and improve the accuracy of diagnostic decision support.

Description of drawings

FIG. 1 is a schematic flowchart of a method for processing and analyzing diagnostic information according to the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. The components of the embodiments of the invention generally described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations. Thus, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but is merely representative of selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present invention.

As shown in Figure 1, the present invention provides a method for processing and analyzing diagnostic information, comprising the following steps:

A feature space and a label space of multi-label information are established by multiple sample features and label information of multiple samples, and a feature vector and a first label vector of the sample are established according to each sample feature and each label information; calculate the The number of occurrences of each label information; calculate the number of times that every two label information appears in a sample at the same time; calculate the similarity of every two label information, and establish a label-label similarity matrix; through the label-label similarity matrix Reconstruct the first label vector of the sample to obtain a second label vector, and calculate the probability of occurrence of label information of the target sample according to the second label vector; sort the probability of occurrence of label information of the target sample in descending order, and select a preset number of labels information as the recommended label information of the target sample.

The principle of the present invention is that a complication refers to another disease or symptom caused during the development of one disease. When a doctor diagnoses a patient, if a disease is diagnosed, the doctor will consider whether the patient also suffers from complications of the disease to ensure that more diseases can be found in the patient. At the same time, a disease can reflect the relationship between itself and other diseases through the number of co-occurrences with other diseases. Therefore, the co-occurrence frequency between two diseases can be used to reflect the similarity between diseases. Using the calculated disease similarities combined with a lazy multi-label classification method, information about the diseases that the patient may suffer from is finally recommended.

In a specific embodiment, a multi-label data set (Basic multiplelabels data set, denoted as BMLDS) is preset according to the patient's diagnosis record; at the same time, the feature set and label set of the sample are separated in the BMLDS, denoted as F and L respectively , and divide the BMLDS into a training set and a test set according to the ratio of 7:3, the method of the present invention includes:

S1. Establish a feature space and a label space of multi-label information through multiple sample features and label information of multiple samples:

The preset F={f ₁ , f ₂ , ... f _b } is the b-dimensional feature space of multiple label information, and the preset L={l ₁ , l ₂ , ... l _q } is the multiple label information q-dimensional label space;

为所述样本的b维特征向量；则The preset T={(X ₁ ,Y ₁ ),(X ₂ ,Y ₂ ),...,(X _n ,Y _n )} is a set of multiple label information, the preset

is the b-dimensional feature vector of the sample; then

为所述样本X_i对应的标签向量；若特征向量X_i有标签空间l_j，则标签向量

否则

is the label vector corresponding to the sample X _i ; if the feature vector X _i has a label space l _j , then the label vector

otherwise

S2, calculate the number of times of occurrence of each label information: according to the number of times that the label information appears in each sample, calculate the sample corresponding to this label information, and set the value to be _{r ij} ; Label space _{li, then r ij =} 1, otherwise r _ij =0.

S3. Calculate the similarity of each two tags, and establish a tag-tag similarity matrix:

Use the cosine similarity calculation method to calculate the similarity of every two tags.

There are three similarity calculation methods, the cosine similarity calculation method, the Pearson correlation coefficient calculation method and the Jaccard similarity coefficient calculation method.

Specifically, using the cosine similarity calculation method to calculate the similarity of every two tags includes:

计算每两个标签的相似度，其中，P_ij是同时包括标签空间l_i与标签空间l_j的集合，

为标签空间l_i与标签空间l_j的同时出现在样本X_k中的次数；

和

分别为标签空间l_i出现的总次数和标签空间l_j出现的总次数；by calculation formula

Calculate the similarity of every two labels, where P _ij is a set including both label space _{li and label space l j ,}

is the number of times the label space l _i and the label space l _j appear in the sample X _k at the same time;

and

are the total number of occurrences of label space l _i and the total number of occurrences of label space l _j ;

The Pearson correlation coefficient is calculated as:

Among them, I and J are the sample vectors corresponding to the labels l _i and l _j respectively, and the Pearson correlation coefficient is the cosine angle of the space vector calculated after the overall normalization of the vector I and the vector J itself.

The calculation method of the Jaccard similarity coefficient is:

where I and J are the sample vectors corresponding to labels l _i and l _j , respectively.

S4. Calculate the similarity of each two tag information, and establish a tag-tag similarity matrix:

The label-label similarity matrix is

Wherein, the element S _ij =sim(I _i , I _j ) in the matrix represents the similarity between the label l _i and the label l _j .

S5. Reconstruct the label information matrix of the sample through the label-label similarity matrix

Y=g(X), where g(X) is:

S6. Calculate the occurrence probability of the label information of the target sample according to the second label vector:

According to the second label vector, a lazy multi-label classification algorithm is used to calculate the occurrence probability of each label information in the target sample. Specifically, the probability value of each label information appearing in the target sample ranges between [0, 1].

The classification function of the lazy multi-label classification algorithm is as follows:

表示第i个标签信息取值为b的概率，其中b为0或者1，b为0表示标签不出现，b为1表示标签出现。事件

表示k个近邻中恰好有j个l_i标签，通过计算等式值的大小来确定标签l是否出现在样本x中。First, count the number of occurrences of each label in the k nearest neighbor (kNN) samples of each sample, and use the method of maximizing the posterior probability to estimate the labels that may appear in the unlabeled samples. For a sample space X containing m samples, its label space is denoted as L. event

Indicates the probability that the i-th label information takes the value of b, where b is 0 or 1, b is 0 means the label does not appear, and b is 1 means the label appears. event

Indicates that there are exactly _{j li} labels in the k nearest neighbors, and whether the label l appears in the sample x is determined by calculating the magnitude of the equation value.

由公式

计算，in,

by formula

calculate,

y _j (li _i ) indicates whether the sample _j has a label li , s∈(0,1). therefore,

的计算需要在遍历训练样本集时，统计每个样本的k近邻的样本中包含标签l_i的情况。数组C[j]统计每个样本标签l_i取值为1时，该样本的k近邻样本中包含标签l_i的个数；数组C'[j]标签l_i取值为0时，该样本的k近邻样本中包含标签l_i的个数。p表示标签的个数。条件概率由下面的等式计算：Conditional Probability

The calculation of , requires that when traversing the training sample set, count the cases where the labels l _i are included in the k-nearest neighbors of each sample. Array C[j] counts the number of labels l _i included in the _k -nearest neighbor samples of each sample when the value of each sample label l _i is 1; The number of labels li included in the _k -nearest neighbor samples of . p represents the number of labels. The conditional probability is calculated by the following equation:

S7. Sort the probability of occurrence of the target sample label information in descending order, and select a preset number of label information as the recommended label information of the target sample.

According to the probability value of the label information appearing in the target sample, the labels are arranged from large to small, and the N label information before rehearsal is selected as the recommended label of the target sample. When treating a patient, the top N possible complication label information with the highest probability value is recommended, and the specific value of N can be preset according to different diseases.

In a specific embodiment, the patient is a sample in the multi-label learning algorithm, the features corresponding to the disease of each patient constitute the feature space in the multi-label learning algorithm, and the diseases of all patients are used as the label information in the multi-label learning algorithm . Through the analysis of the label information, the co-occurrence times of each two label information are obtained; the more times two different label information appear in the same sample at the same time, the greater the correlation between the two label information; according to the cosine similarity Calculate the similarity between each two label information, and recalculate the label vector corresponding to the sample according to the similarity of the label. If the vector value corresponding to a label is equal to or exceeds 0.5, the vector value of the label is reset to 1, otherwise the vector value of the label is still 0; by reconstructing the label vector corresponding to the sample, the potential possible labels of the training sample are found, the labels of the training samples are associated, and the label matrix in the label space is reconstructed by using the label similarity matrix to Update the label vector corresponding to each sample, and finally use the multi-label learning algorithm to predict possible labels for the target sample.

In specific examples, 9 common diseases (including type 2 diabetes, hyperlipidemia, fatty liver, hyperkalemia, hypoalbuminemia, diabetic nephropathy, cerebral infarction, coronary heart disease and osteoporosis) were selected. One or more patients suffering from these diseases were selected from the hospital as experimental samples; the laboratory reports and basic information of the patients were collected as sample characteristics, and 459 patient samples containing 5 kinds of basic information of patients and 278 kinds of test items were obtained. Gender, age, body temperature, height and weight are extracted from the patient sample as the basic attributes of the patient. The value of gender is binary, such as 0 for male and 1 for female. However, the values of age, body temperature, height and weight are numerical, and their actual values are retained; for items whose laboratory values are numerical, if their laboratory values are within the normal reference range, their values are set to 1; If its assay value is out of the normal range, its value is set as the actual assay value. For an item whose assay value is in the form of a text description, collect the different text description values of the item and arrange them using an array. If the text description value of the item is equal to the reference value, the value is 0; if the text description value of the item is not equal to the reference value, the value is set to the arrangement value of the text description in the array.

As shown in Tables 1 and 2, statistics of characteristics and statistics of diseases are listed, respectively. Overall, 60% of patients were male and 40% were female. The mean age, body temperature, height and weight of the patients were 64.64, 36.5, 167.81 and 67.75, respectively. From the statistics of diseases, among these 9 diseases, type 2 diabetes and cerebral infarction are the two most common diseases. In fact, these diseases are also the most common diseases in the elderly population. We randomly select 70% of patients as training samples and the remaining 30% as test samples.

Table 1

Table 2

The evaluation criteria for multi-label classification problems are divided into two categories: a. Evaluation criteria based on ranking: the goal is to rank relevant examples before irrelevant examples. b. Binary prediction evaluation: The goal is to make a strict yes/no classification for each target sample. Hamming Loss, precision, recall and F1-score (F1 score) are used to evaluate the effect of the present invention.

Hamming Loss evaluates the average difference between the recommended label of the test sample and its actual label:

Among them, h( _xi ) is the recommended label set of the test sample _xi ; p is the number of test samples; Y _i is the actual label set of the test sample _xi ; Δ is the symmetric difference.

The accuracy rate is defined as the ratio of the number of tags hit in the tag recommendation list to the total number of tag recommendation lists. That is, the accuracy rate represents the probability that the test sample has the accuracy rate of the recommended label. The formula for the accuracy rate is as follows:

Recall is defined as the ratio of the number of hit labels to the true labels of the test sample. In other words, recall represents the probability that the true label is recommended with accuracy. The formula for recall is as follows:

F1-score considers both precision and recall, and its formula is as follows:

The effect of the method is analyzed by comparing the effect of the present invention with the effect of two classical multi-label classification methods. Two classic multi-label classification methods are a lazy multi-label classification method (ML-kNN) and a multi-label classification method combining BR method and kNN method (BR-kNN). In Table 3, it is mainly compared with the classical multi-label classification algorithm. According to the number of nearest neighbors with good stability and high accuracy, the number of nearest neighbors is set to 10. The smoothing factor of the present invention and the smoothing factor of ML-kNN are both Set to 1. The recommended number of labels is 2 in all methods. Experiments are performed using 10-fold cross-validation, and the final result is the average of these experimental results. In Table 3, for the present invention, its precision is 0.236, recall is 0.3793 and F1-score is 0.2915, which are better than the other two methods. Compared with ML-kNN, which ranks second in the experimental results, the precision, recall and F1-score of the present invention are improved by 11%, 13% and 12%, respectively. The HammingLoss of the present invention is 0.2117, which is also better than the other two methods. Therefore, the performance of the present invention is superior to the other two methods.

table 3

↓: The smaller the value, the better the effect ↑: The larger the value, the better the effect

The present invention also provides a diagnostic information processing and analysis system, the system comprising:

A module for establishing a feature space and a label space of multi-label information according to multiple sample features and label information of multiple samples;

A module for establishing a feature vector and a first label vector of the sample according to each sample feature and each label information;

A module for calculating the number of occurrences of each label information;

A module for calculating the number of times that each two label information appears in a sample at the same time;

A module for calculating the similarity of each two tag information and establishing a tag-tag similarity matrix;

A module for reconstructing the label information matrix of the sample by the label-label similarity matrix, reconstructing the first label vector of the sample to obtain a second label vector, and for calculating the target sample according to the second label vector The module of the probability of occurrence of label information;

A module used to sort the probability of occurrence of the target sample label information in descending order'

A module for selecting a preset number of label information as the recommended label information for the target sample.

Preferably, the system further includes: the module for calculating the similarity of each two tag information, and the module for establishing a tag-tag similarity matrix is a cosine similarity calculation module.

The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention. should be included within the protection scope of the present invention.

Claims (9)

1. A method of diagnostic information processing analysis, comprising the steps of:

establishing a feature space and a label space of multi-label information through a plurality of sample features and label information of a plurality of samples, and establishing a feature vector and a first label vector of each sample according to each sample feature and each label information; the sample characteristics are the test report and basic information of the disease of the patient;

calculating the occurrence frequency of each label information; calculating the times that every two label information simultaneously appear in one sample; calculating the similarity of every two label information; establishing a label-label similarity matrix;

reconstructing a first label vector of the sample through the label-label similarity matrix to obtain a second label vector, and calculating the label information occurrence probability of the target sample according to the second label vector;

sorting the probability of the target sample label information in a descending order, and selecting a preset number of label information as the recommended label information of the target sample;

the establishing of the feature space and the label space of the multi-label information by the plurality of sample features and the label information of the plurality of samples further comprises:

preset F ═ F₁,f₂...f_bB-dimensional feature space of a plurality of label information, and preset L ═ L₁,l₂,...l_qQ-dimensional label space of the plurality of label information;

predetermined T { (X)₁,Y₁),(X₂,Y₂),...,(X_n,Y_n) Is a set of multiple label information, preset

B-dimensional feature vectors for the samples; then

Is the sample X_iA corresponding tag vector; if the feature vector X_iSpace with label l_jThen label vector

Otherwise

2. The method of claim 1, wherein said calculating the number of occurrences of said each tag information comprises: calculating the sample corresponding to the label information according to the occurrence frequency of the label information in each sample, and setting the value as r_ij(ii) a If the feature vector X_kIn which there is a label space l_iThen r is_ij1, otherwise r_ij＝0。

3. The method of claim 2, wherein the calculating the similarity between every two tags and establishing the tag-to-tag similarity matrix further comprises:

and calculating the similarity of every two labels by using a cosine similarity calculation method.

4. The method of claim 3, wherein the calculating the similarity of each two tags by using the cosine similarity calculation method comprises:

by calculation of formula

The similarity of every two labels is calculated,

wherein, P_ijIs to include label space l at the same time_iAnd the label space l_jThe set of (a) and (b),

is a label space l_iAnd the label space l_jWhile appearing in sample X_kThe number of times of (1);

and

respectively a label space l_iTotal number of occurrences and label space l_jTotal number of occurrences.

5. The method of claim 4, wherein the calculating the similarity between every two label information and establishing the label-to-label similarity matrix comprises:

the label-to-label similarity matrix is

Wherein the element S in the matrix_ij＝sim(I_i,I_j) Presentation label l_iAnd a label l_jThe similarity of (c).

6. The method of claim 5, wherein reconstructing the label information matrix of the sample by the label-to-label similarity matrix is: y ═ g (x),

wherein g (X) is:

7. the method of claim 1, wherein the calculating the probability of occurrence of the tag information for the target sample according to the second tag vector comprises:

and calculating the occurrence probability of each label information in the target sample by using a lazy multi-label classification algorithm according to the second label vector.

8. A diagnostic information processing and analysis system, the system comprising:

a module for establishing a feature space and a label space of multi-label information according to the plurality of sample features and label information of the plurality of samples; the sample characteristics are the test report and basic information of the disease of the patient;

means for establishing a feature vector and a first label vector for the sample based on each sample feature and each label information;

a module for calculating the number of occurrences of each of the tag information;

a module for counting the number of times that every two tag information simultaneously appear in one sample;

a module for calculating the similarity of every two label information and establishing a label-label similarity matrix;

the module is used for reconstructing a label information matrix of the sample through the label-label similarity matrix, reconstructing a first label vector of the sample to obtain a second label vector, and calculating the occurrence probability of label information of a target sample according to the second label vector;

means for sorting the probabilities of occurrence of the target sample label information in descending order;

a module for selecting a preset number of label information as the recommended label information of the target sample;

the establishing of the feature space and the label space of the multi-label information according to the plurality of sample features and the label information of the plurality of samples further comprises:

preset F ═ F₁,f₂...f_bB-dimensional feature space of a plurality of label information, and preset L ═ L₁,l₂,...l_qQ-dimensional label space of the plurality of label information;

predetermined T { (X)₁,Y₁),(X₂,Y₂),...,(X_n,Y_n) Is a set of multiple label information, preset

B-dimensional feature vectors for the samples; then

Is the sample X_iA corresponding tag vector; if the feature vector X_iSpace with label l_jThen label vector

Otherwise

9. The system of claim 8, further comprising: the module for calculating the similarity of every two label information and establishing the label-label similarity matrix is a module for calculating cosine similarity.

Images