Multi-mark learning method and system

Abstract

The invention provides a partial multi-mark learning method. The partial multi-mark learning method comprises the following steps: constructing an example matrix and a mark matrix containing candidate marks; constructing a first objective function based on the example and the tag correlation hypothesis; constructing a second objective function based on the generic attribute and the example correlation hypothesis; adding the first objective function and the second objective function to obtain a final objective function, and obtaining an optimization problem of the objective function; solving an optimal solution of an optimization problem of a matrix related to the example and the mark by using an alternating minimization algorithm; according to the optimal solution, the mark vector of the new example is obtained, the noise marks in the candidate marks can be screened out through the method, and the prediction performance of the partial multi-mark learning method is improved.

Description

Technical Field

[0001] This specification relates to the field of machine learning and computer applications, and in particular to a multi-label learning method and system. Background Art

[0002] In the multi-label learning framework, each object in the real world is described by an example, which is associated with multiple tags. The goal of multi-label learning is to accurately predict all tags associated with a given new example. Due to the uneven professionalism of the annotators, some annotators may mistakenly label some irrelevant tags as relevant tags. Based on this reality, the partial multi-label learning framework came into being. In this framework, each object is described by an example, which is associated with a candidate tag set containing an unknown number of relevant tags and some irrelevant tags (i.e., noise tags). How to learn a multi-label classifier from training data containing noise tags to accurately predict all tags associated with a given new example is the main problem to be solved by partial multi-label learning. From the perspective of examples, two similar examples tend to have similar tags. For example, two images of mountain views are likely to have the tags "mountain" and "tree"; from the perspective of tags, highly related tags tend to appear at the same time. For example, if an image has the tags "theme park" and "Mickey Mouse", it is likely to also have the tag "Disney", but the tag may be an incorrect tag. How to eliminate irrelevant noise labels in examples to improve the label prediction performance of multi-label learning methods remains an open problem. Summary of the invention

[0003] One of the embodiments of this specification provides a multi-label learning method. The multi-label learning method includes:

[0004] Step 1: Construct the sample matrix in, is the example of the i-th sample, d is the dimension of the example, and n is the number of samples;

[0005] Step 2: Construct a label matrix Y = [y1, y2, ..., y n ]∈{0,1} q×n , where y ji = 1 means that the i-th example sample is labeled with the candidate label of the j-th label, y ji =0 means that the i-th sample is not labeled with the j-th label, and q is the number of labels;

[0006] Step 3: Based on the correlation assumption between the example and the label, construct a first objective function f1(U, V, Z), where: is the labeled low-rank matrix, which is the matrix composed of the low-rank representation of all labeled vectors. is an example low-rank matrix, which is a matrix consisting of low-rank representations of all the examples, is a matrix consisting of the probability of selecting each representative marker by all the markers;

[0007] Step 4: Based on the class attributes and the example correlation hypothesis, construct a second objective function f2(W), where: A linear mapping relationship from the example space to the label space constructed based on the example matrix X and the label matrix Y;

[0008] Step 5: Add the first objective function and the second objective function to obtain the final objective function f(U, V, Z, W), and obtain the optimization problem of the objective function;

[0009] Step 6: Use the alternating minimization algorithm to find the optimal solution for the four matrices U, V, Z, and W involved in the optimization problem.

[0010] Step 7: According to the optimal solution New Example Substitute into the formula Get the label vector

[0011] One of the embodiments of this specification provides a partial multi-label learning system, which includes a data processing unit, an optimization target modeling unit, a model training unit and a label prediction unit;

[0012] The data processing unit is used to construct an example matrix in, is the example of the i-th sample, d is the dimension of the representation, and n is the number of samples;

[0013] The data processing unit is also used to construct a candidate marker matrix Y=[y1, y2, ..., y n ]∈{0,1} q×n , where y ji = 1 means that the i-th sample is labeled with the j-th tag, y ji =0 means that the i-th sample is not labeled with the j-th label, and q is the number of labels;

[0014] The optimization target modeling unit is used to construct a first target function f1(U, V, Z) based on the correlation hypothesis between the example and the candidate tag, where: is a labeled low-rank matrix, which is a matrix composed of low-rank representations of all the labeled vectors. is an example low-rank matrix, which is a matrix consisting of low-rank representations of all the examples, is a matrix consisting of the probability of selecting each representative marker by all the markers;

[0015] Based on the generic attributes and the example correlation hypothesis, a second objective function f2(W) is constructed, where: A linear mapping relationship from a feature space to a label space constructed based on the example matrix X and the label matrix Y;

[0016] Adding the first objective function to the second objective function obtains a final objective function f(U, V, Z, W), and obtains an optimization problem of the objective function;

[0017] The model training unit is used to use the alternating minimization algorithm to obtain the optimal solution for the four matrices U, V, Z, and W involved in the optimization problem.

[0018] The label prediction unit is used to predict the optimal solution New Example Substitute into the formula Get the label vector

[0019] One of the embodiments of the present specification provides a partial multi-label learning device, including a processor, wherein the processor is used to execute a partial multi-label learning method.

[0020] One of the embodiments of the present specification provides a computer-readable storage medium, wherein the storage medium stores computer instructions. When a computer reads the computer instructions in the storage medium, the computer executes a partial multi-label learning method.

[0021] The beneficial effects of the present invention include: the noise labels in the candidate labels can be screened out by the above method, thereby improving the prediction performance of the multi-label learning method. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] This specification will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:

[0023] Figure 1 is an exemplary flow chart of a partial multi-label learning method according to some embodiments of this specification;

[0024] Figure 2 It is an exemplary structural diagram of a multi-label learning system according to some embodiments of this specification. DETAILED DESCRIPTION

[0025] In order to more clearly illustrate the technical solutions of the embodiments of this specification, the following is a brief introduction to the drawings required for the description of the embodiments. Obviously, the drawings described below are only some examples or embodiments of this specification. For ordinary technicians in this field, this specification can also be applied to other similar scenarios based on these drawings without creative work. Unless it is obvious from the language environment or otherwise explained, the same reference numerals in the figures represent the same structure or operation.

[0026] It should be understood that the "system", "device", "unit" and / or "module" used herein are a method for distinguishing different components, elements, parts, portions or assemblies at different levels. However, if other words can achieve the same purpose, the words can be replaced by other expressions.

[0027] As shown in this specification and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "comprise" and "include" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0028] Flowcharts are used in this specification to illustrate the operations performed by the system according to the embodiments of this specification. It should be understood that the preceding or following operations are not necessarily performed precisely in order. Instead, the steps may be processed in reverse order or simultaneously. At the same time, other operations may also be added to these processes, or one or more operations may be removed from these processes.

[0029] Figure 1 is an exemplary flow chart of a partial multi-label learning method according to some embodiments of this specification. Figure 1 As shown, the method flow includes the following steps:

[0030] Step 1: Construct the sample matrix in, is the example of the i-th sample, d is the dimension of the example, and n is the number of samples.

[0031] For example, assuming d=5, n=4, the example matrix X has the following form:

[0032]

[0033] That is, the dimension of the example is 5, the number of samples is 4, and each column corresponds to an example of a sample.

[0034] Step 2: Construct a label matrix Y = [y1, y2, ..., y n]∈{0,1} q×n , where y ji = 1 means that the i-th sample is labeled with the candidate label of the j-th label, y ji =0 means that the i-th sample is not labeled with the j-th label, and q is the number of labels.

[0035] In some embodiments, a candidate label may represent an example containing irrelevant label errors, ie, a noise label.

[0036] Specifically, assuming q=6, the labeling matrix Y has the following form:

[0037]

[0038] That is, the number of labels is 6, and each column corresponds to the label vector of a sample.

[0039] Step 3: Based on the correlation hypothesis between the example and the candidate tag, construct a first objective function f1(U, V, Z), where: is a labeled low-rank matrix, which is a matrix composed of low-rank representations of all the labeled vectors; is an example low-rank matrix, which is a matrix composed of low-rank representations of all the examples; is a matrix consisting of the probability of selecting each representative marker among all the markers.

[0040] Specifically, since the labels are correlated, the product of the label low-rank matrix U and the example low-rank matrix V is approximately obtained Among them, ||·|| F represents the Frobenius norm.

[0041] The sample samples are grouped using the K-Means mean clustering method. 11 Based on (U,V), we introduce the row sparse term ‖UV m ‖ 1,2 ,get Among them, V m is a matrix consisting of low-rank representations of examples belonging to the mth group of samples, α m It is the influence parameter, which is used to control the influence of the row sparse item on the objective function.

[0042] In order to eliminate noise labels based on sample correlation, the K-Means mean clustering method is used to divide the samples so that the samples belonging to the same group are as similar as possible, and the samples belonging to different groups are as dissimilar as possible. Since samples with similar examples tend to have similar label vectors, the label matrix composed of the label vectors of samples in the same group often has only a few rows containing non-zero elements. Based on this inference, the row sparse term ‖UV is introduced m ‖ 1,2 .

[0043] Minimizing the objective function can ensure that the label matrix composed of the label vectors of each group of samples has only a few rows containing elements with values ​​other than 0, thereby achieving the goal of eliminating noise labels based on example correlation.

[0044] Furthermore, since the low-rank representations of two tags with high correlation should also be similar, the regularization term of the candidate tag correlation and the row sparseness term of the probability matrix are introduced. jj′ represents the probability that the j'th candidate marker selects the jth candidate marker as its representative marker. Then the probability of all markers selecting each representative marker can form a probability matrix Z, and then the regularization term of the candidate marker correlation is introduced. and the row sparse items of the probability matrix ‖Z‖ 1,2 , and obtain the first objective function:

[0045]

[0046] Among them, β1 and β2 are parameters used to control the influence of the regularization term of the candidate label correlation associated with the matrix and the row sparse term of the probability matrix on the first objective function.

[0047] In some embodiments, step 3 may also introduce a constraint term, setting 0≤Z≤1, so that the probability of each tag selecting any tag as its representative tag is between 0 and 1. Introducing the constraint Make the sum of the probabilities of each marker selecting the 1st to qth markers as its representative markers equal to 1;

[0048] On this basis, the optimization problem can be obtained

[0049] Step 4: Based on the class attributes and the example correlation hypothesis, construct a second objective function f2(W), where: It is a linear mapping relationship from the feature space to the label space constructed based on the example matrix X and the candidate label matrix Y.

[0050] Specifically, a linear mapping W from the example space to the label space is introduced, and a function about W is constructed

[0051] Based on the class attributes, the linear mapping W is regularized to obtain the class attribute regularization term ||W||1, and the function f is introduced 21 get: Among them, λ1 is the influence parameter, which is used to control the influence of the category attribute regularization term on the objective function.

[0052] Minimizing this objective function can make the matrix W sparse, thereby achieving the goal of improving the label prediction performance based on the category attributes.

[0053] According to the example correlation, an example correlation matrix is ​​constructed by an exponential function. in,

[0054] exp(·) represents the exponential function, It refers to a set of multiple example numbers similar to the i-th example.

[0055] Furthermore, since samples with similar feature vectors often have similar label vectors, a regularization term of the example correlation matrix is ​​introduced. The example correlation matrix is ​​regularized to obtain the regularization term of the example correlation matrix Where X ：i Denote the i-th column of the matrix X and denote it equivalently as i, where tr(·) denotes the trace of the matrix, L = DS, D = diag(S1 n ), is a column vector consisting of n ones. diag(a) returns a diagonal matrix with the elements in vector a as diagonal elements, and then the second objective function is obtained:

[0056]

[0057] Among them, λ2 is the influence parameter, which is used to control the influence of the last item on the second objective function.

[0058] In some embodiments, step 4 can also be based on the second objective function to obtain the optimization problem

[0059] Step 5: Add the first objective function and the second objective function to obtain the optimized objective function f(U, V, Z, W), and obtain the optimization problem of the optimized objective function.

[0060] Specifically, the first objective function and the second objective function are added to obtain the optimized objective function f(U, V, Z, W):

[0061]

[0062] Introduce a constraint term, set 0≤Z≤1, so that the probability of each candidate marker selecting any candidate marker as its representative marker is between 0 and 1, and set Make the sum of the probabilities of each candidate marker selecting the 1st to qth markers as its representative markers equal to 1;

[0063] Minimize the objective function to get the optimization problem

[0064] Step 6: Use the alternating minimization algorithm to find the optimal solution of the four matrices U, V, Z, and W involved in the optimization problem. Specifically, the optimal solution is an estimated value of the optimal solution obtained by using an alternating minimization algorithm.

[0065] Step 7: According to the optimal solution New Example Bring in Get the label vector

[0066] In some embodiments, the above-mentioned multi-label learning method can help the system screen out some irrelevant noise labels in the fields of picture classification and image recognition, thereby improving the performance of example label prediction.

[0067] Figure 2 2 is a schematic diagram of the structure of a partial multi-label learning system provided by an embodiment of the present invention, the system includes: a data processing unit 210, an optimization target modeling unit 220, a model training unit 230 and a label prediction unit 240. Specifically, it includes:

[0068] Data processing unit 210: used to construct an example matrix in, is the example of the i-th sample, d is the dimension of the example, and n is the number of samples;

[0069] The data processing unit is also used to construct a marker matrix Y=[y1, y2, ..., y n ]∈{0,1} q×n , where y ji = 1 means that the i-th sample is labeled with the candidate label of the j-th label, y ji =0 means that the i-th sample is not labeled with the j-th label, and q is the number of labels;

[0070] The optimization target modeling unit 220 is used to construct a first target function f1(U, V, Z) based on the correlation hypothesis of the example and the label, wherein: is a labeled low-rank matrix, which is a matrix composed of low-rank representations of all the labeled vectors. is an example low-rank matrix, which is a matrix consisting of low-rank representations of all the examples, is a matrix consisting of the probability of selecting each representative marker by all the markers;

[0071] Based on the generic attributes and the example correlation hypothesis, a second objective function f2(W) is constructed, where: A linear mapping relationship from a feature space to a label space constructed based on the example matrix X and the label matrix Y;

[0072] Adding the first objective function to the second objective function obtains a final objective function f(U, V, Z, W), and obtains an optimization problem of the objective function;

[0073] Model training unit 230: used to use the alternating minimization algorithm to solve the optimal solution of the four matrices U, V, Z, and W involved in the optimization problem

[0074] Tag prediction unit 240: used to predict the optimal solution New Example Substitute into the formula Get the label vector

[0075] An embodiment of the present invention further provides a partial multi-label learning device, comprising a processor, wherein the processor is used to execute the partial multi-label learning method described above.

[0076] An embodiment of the present invention provides a computer-readable storage medium, wherein the storage medium stores computer instructions. When a computer reads the computer instructions in the storage medium, the computer executes any of the above-described partial multi-label learning methods.

[0077] The basic concepts have been described above. Obviously, for those skilled in the art, the above detailed disclosure is only for example and does not constitute a limitation of this specification. Although not explicitly stated here, those skilled in the art may make various modifications, improvements and corrections to this specification. Such modifications, improvements and corrections are suggested in this specification, so such modifications, improvements and corrections still belong to the spirit and scope of the exemplary embodiments of this specification.

[0078] At the same time, this specification uses specific words to describe the embodiments of this specification. For example, "one embodiment", "an embodiment", and / or "some embodiments" refer to a certain feature, structure or characteristic related to at least one embodiment of this specification. Therefore, it should be emphasized and noted that "one embodiment" or "an embodiment" or "an alternative embodiment" mentioned twice or more in different positions in this specification does not necessarily refer to the same embodiment. In addition, certain features, structures or characteristics in one or more embodiments of this specification can be appropriately combined.

[0079] In addition, unless explicitly stated in the claims, the order of the processing elements and sequences described in this specification, the use of alphanumeric characters, or the use of other names are not intended to limit the order of the processes and methods of this specification. Although the above disclosure discusses some invention embodiments that are currently considered useful through various examples, it should be understood that such details are only for illustrative purposes, and the attached claims are not limited to the disclosed embodiments. On the contrary, the claims are intended to cover all modifications and equivalent combinations that are consistent with the essence and scope of the embodiments of this specification. For example, although the system components described above can be implemented by hardware devices, they can also be implemented only by software solutions, such as installing the described system on an existing server or mobile device.

[0080] Similarly, it should be noted that in order to simplify the description disclosed in this specification and thus help understand one or more embodiments of the invention, in the above description of the embodiments of this specification, multiple features are sometimes combined into one embodiment, figure or description thereof. However, this disclosure method does not mean that the features required by the subject matter of this specification are more than the features mentioned in the claims. In fact, the features of the embodiments are less than all the features of the single embodiment disclosed above.

[0081] In some embodiments, numbers describing the number of components and attributes are used. It should be understood that such numbers used in the description of the embodiments are modified by the modifiers "about", "approximately" or "substantially" in some examples. Unless otherwise specified, "about", "approximately" or "substantially" indicate that the numbers are allowed to vary by ±20%. Accordingly, in some embodiments, the numerical parameters used in the specification and claims are approximate values, which may change according to the required features of individual embodiments. In some embodiments, the numerical parameters should take into account the specified significant digits and adopt the general method of retaining digits. Although the numerical domains and parameters used to confirm the breadth of their range in some embodiments of this specification are approximate values, in specific embodiments, the setting of such numerical values ​​is as accurate as possible within the feasible range.

[0082] Each patent, patent application, patent application publication, and other materials, such as articles, books, specifications, publications, documents, etc., cited in this specification are hereby incorporated by reference in their entirety. Except for application history documents that are inconsistent with or conflicting with the contents of this specification, documents that limit the broadest scope of the claims of this specification (currently or later attached to this specification) are also excluded. It should be noted that if the descriptions, definitions, and / or use of terms in the materials attached to this specification are inconsistent or conflicting with the contents described in this specification, the descriptions, definitions, and / or use of terms in this specification shall prevail.

[0083] Finally, it should be understood that the embodiments described in this specification are only used to illustrate the principles of the embodiments of this specification. Other variations may also fall within the scope of this specification. Therefore, as an example and not a limitation, alternative configurations of the embodiments of this specification may be considered consistent with the teachings of this specification. Accordingly, the embodiments of this specification are not limited to the embodiments explicitly introduced and described in this specification.

Claims

1. A partial multi-label learning method, characterized in that: include: Step 1: Construct the sample matrix in, is the example of the i-th sample, d is the dimension of the example, and n is the number of samples; Step 2: Construct a label matrix Y = [y1, y2, ..., y n ]∈{0,1} q×n , where y ji = 1 means that the i-th sample is labeled with the candidate label of the j-th label, y ji =0 means that the i-th sample is not labeled with the j-th label, and q is the number of labels; Step 3: Based on the correlation assumption between the example and the label, construct a first objective function f1(U, V, Z), where: is the labeled low-rank matrix, which is the matrix composed of the low-rank representation of all labeled vectors. is an example low-rank matrix, which is a matrix consisting of low-rank representations of all the examples, is a matrix consisting of the probability of selecting each representative marker by all the markers; Step 4: Based on the class attributes and the example correlation hypothesis, construct a second objective function f2(W), where: A linear mapping relationship from a feature space to a label space constructed based on the example matrix X and the label matrix Y; Step 5: Add the first objective function and the second objective function to obtain the final objective function f(U, V, Z, W), and obtain the optimization problem of the objective function; Step 6: Use the alternating minimization algorithm to find the optimal solution of the four matrices U, V, Z, and W involved in the optimization problem. Step 7: According to the optimal solution New Example Substitute into the formula Get the label vector 2. The multi-label learning method according to claim 1, characterized in that: The step three comprises: Introduce the labeled low-rank matrix U and the example low-rank matrix V, and construct a function about U and V Among them, ||·|| F represents the Frobenius norm; The sample samples are grouped using the K-Means mean clustering method. 11 (U,V) introduces row sparse term ||UV m || 1，2 ,get Among them, V m is a matrix consisting of low-rank representations of examples belonging to the mth group, α m is the influence parameter, which is used to control the influence of the row sparse item on the objective function; Let z jj′ represents the probability that the j′th marker selects the jth marker as its representative marker. Then the probability of all markers selecting each representative marker can form a probability matrix Z, and then the regularization term of marker correlation is introduced. and the row sparse entries of the probability matrix ||Z|| 1，2 , and obtain the first objective function: Among them, β1 and β2 are parameters used to control the influence of the regularization term of label correlation and the row sparse term of the probability matrix on the first objective function.

3. The multi-label learning method according to claim 2, characterized in that: The fourth step comprises: Introduce a linear mapping W from the example space to the label space and construct a function about W Based on the generic attributes, the linear mapping W is regularized to obtain the generic attribute regularization term ||W||1, and the function g is introduced 21 get: Among them, λ1 is the influence parameter, which is used to control the influence of the category attribute regularization term on the objective function; According to the example correlation, an example correlation matrix is ​​constructed by an exponential function. in, exp(·) represents the exponential function, It refers to a set of multiple example numbers similar to the i-th example; The example correlation matrix is ​​regularized to obtain the regularization term of the example correlation matrix Where X ：i represents the i-th column of matrix X, and it is equivalently expressed as where tr(·) represents the trace of the matrix, L = DS, D = diag(S1 n ), is a column vector consisting of n ones. diag(a) returns a diagonal matrix with the elements in vector a as diagonal elements, and then the second objective function is obtained: Among them, λ2 is the influence parameter, which is used to control the influence of the last item on the second objective function.

4. The multi-label learning method according to claim 3, characterized in that: The step five comprises: The first objective function is added to the second objective function to obtain the final objective function f(U, V, Z, W): Introduce a constraint term, set 0≤Z≤1, so that the probability of each tag selecting any candidate tag as its representative tag is between 0 and 1, and set Make the sum of the probabilities of each marker selecting the 1st to qth markers as its representative markers equal to 1; Minimize the objective function to get the optimization problem 5. A partial multi-label learning system, characterized in that: include: Data processing unit, optimization target modeling unit, model training unit and label prediction unit; The data processing unit is used to construct an example matrix in, is the example of the i-th sample, d is the dimension of the example, and n is the number of samples; The data processing unit is also used to construct a marker matrix Y=[y1, y2, ..., y n ]∈{0,1} q×n , where y ji = 1 means that the i-th sample is labeled with the candidate label of the j-th label, y ji =0 means that the i-th sample is not labeled with the j-th label, and q is the number of labels; The optimization target modeling unit is used to construct a first target function f1(U, V, Z) based on the correlation assumption between the example and the label, where: is the labeled low-rank matrix, which is the matrix composed of the low-rank representations of all labels. is an example low-rank matrix, which is a matrix consisting of low-rank representations of all the examples, is a matrix consisting of the probability of selecting each representative marker by all the markers; Based on the generic attributes and the example correlation hypothesis, a second objective function f2(W) is constructed, where: A linear mapping relationship from a feature space to a label space constructed based on the example matrix X and the label matrix Y; Adding the first objective function to the second objective function obtains a final objective function f(U, V, Z, W), and obtains an optimization problem of the objective function; The model training unit is used to use the alternating minimization algorithm to obtain the optimal solution for the four matrices U, V, Z, and W involved in the optimization problem. The label prediction unit is used to predict the optimal solution New Example Substitute into the formula Get the label vector 6. A partially multi-label learning device, comprising a processor, wherein the processor is used to execute the partially multi-label learning method according to any one of claims 1 to 4.

7. A computer-readable storage medium, wherein the storage medium stores computer instructions. When a computer reads the computer instructions in the storage medium, the computer executes the multi-label learning method as described in any one of claims 1 to 4.

Images