An entity tag association prediction method and device and a computer readable storage medium

Abstract

The application provides an entity label association prediction method, device, system and computer readable storage medium, the method comprises: determining a plurality of dimensions of entity relationship network, hierarchical label relationship network and entity label association network, the entity label association network contains unknown entity label association relationship; according to the entity relationship network, hierarchical label relationship network and entity label association network, the entity similarity association matrix, the label similarity association matrix and the entity label association matrix are constructed; according to the entity similarity association matrix, the label similarity association matrix and the entity label association matrix, model solving is carried out, and a prediction result is obtained;Wherein, the objective function of the model introduces the regular term constraint of weight coefficient vector μ To facilitate the weight coefficient vector to have sparsity, the weight coefficient vector μ It is a parameter to be learned, indicates the contribution degree of each dimension of entity relationship network to the model. By using the method, the final prediction accuracy of the model can be effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of machine learning, and specifically relates to an entity label association prediction method, device and computer-readable storage medium. Background Technology

[0002] This section is intended to provide background or context for embodiments of the invention as set forth in the claims. The description herein is not an admission that it is prior art simply because it is included in this section.

[0003] With the advent of the big data era, massive amounts of data have emerged, and the semantic content of objects such as users, articles, and products is becoming increasingly rich, while the label information describing their characteristics is becoming increasingly complex. An e-commerce platform's tag library for describing user characteristics often reaches tens of thousands of entries, and the average user is labeled with thousands of tags. To meet the current demand for refined descriptions of sample objects, multi-label learning has gradually become a research hotspot in academia and industry, and has wide applications in many fields such as image recognition, text classification, information retrieval, and personalized recommendation.

[0004] However, due to the sheer size of the tag library and the tens of thousands of samples, the existing annotations for these samples are often incomplete, with numerous omissions. For example, a book recommendation and rating platform might categorize the book "Introduction to Bioinformatics: Algorithms and Applications for High-Performance Computing" as "Biological Sciences" and "Biotechnology," but omit annotations such as "Computer Technology" and "Automated Reasoning and Machine Learning," resulting in an inaccurate description of the book. This issue can lead to unsatisfactory recommendation results when the system makes accurate book recommendations to users.

[0005] Therefore, improving the completeness of sample data annotation is an urgent problem to be solved. Summary of the Invention

[0006] To address the problems existing in the prior art, an entity tag association prediction method, apparatus, and computer-readable storage medium are proposed. Using this method, apparatus, and computer-readable storage medium, the above problems can be solved.

[0007] The present invention provides the following solutions.

[0008] Firstly, a method for predicting entity label association is provided, including: determining... The system includes a multi-dimensional entity relationship network, a hierarchical tag relationship network, and an entity tag association network. The entity tag association network contains unknown entity tag association relationships. The integer is a positive integer greater than 1; an entity similarity association matrix is ​​constructed based on the entity relationship network, a label similarity association matrix is ​​constructed based on the hierarchical label relationship network, and an entity label association matrix is ​​constructed based on the entity label association network; the model is solved based on the entity similarity association matrix, the label similarity association matrix, and the entity label association matrix to obtain the prediction result; the objective function of the model introduces the weight coefficient vector. The regularization term constraint is applied to make the weight coefficient vector sparse. These are the parameters to be learned, used to indicate the... The contribution of the entity relationship network of each dimension to the model.

[0009] In one implementation, model solving further includes: obtaining the predicted values ​​of the initialized entity label association score matrix. and weight coefficient vector The following problem is iteratively solved using the alternating least squares method until convergence to obtain the prediction result: Solve the first optimization problem derived from the objective function decomposition, and calculate and update the predicted value. Solving the objective function decomposes into a second optimization problem, and the weight coefficient vector is then calculated and updated. .

[0010] In one implementation, the first term of the objective function is used to predict the entity label association score matrix for elements on which the entity label relationships are explicitly known. Entity label association matrix Approaching.

[0011] In one implementation, the second term of the objective function corresponds to the predicted value. Apply smoothing constraints to the entity relationship network and use a weight coefficient vector. right The constraint terms on the entity relationship network are linearly summed; the third term of the objective function affects the predicted value. Apply smoothing constraints to the hierarchical label relationship network.

[0012] In one implementation, the fourth term of the objective function is used to adjust the weight coefficient vector. The regularization term that penalizes the modulus is used to constrain the weight coefficient vector to satisfy sparsity.

[0013] In one implementation, the method further includes: determining the objective function of the prediction model according to the following formula:

[0014]

[0015] in, The predicted values ​​for the entity label association score matrix. This is an entity label association matrix. Denotes the Frobenius norm of the matrix; where, Entity Relationship Network The normalized symmetric graph Laplace matrix, For a diagonal matrix, the first... diagonal elements for No. The line of lines, This is the entity similarity association matrix; where, For hierarchical label relationship network The graph Laplace matrix, diagonal matrix The Middle diagonal elements for No. The line of lines, Let be the tag similarity association matrix; where is the weight coefficient vector. For the parameters to be learned, their linear integration The degree to which the entity relationship network contributes to the model. Representing vectors Norm; , and These are hyperparameters used to balance the weights of each term in the objective function.

[0016] In one implementation, it further includes: fixing the weight coefficient vector in the objective function. Find the predicted value of the entity label association score matrix that minimizes the objective function. As the first optimization problem; and / or, fixing the predicted values ​​in the objective function. Find the weight coefficient vector that minimizes the objective function. This is the second optimization problem.

[0017] In one implementation, the first optimization problem is the problem of solving the Sylvester equation.

[0018] In one implementation, the second optimization problem is a quadratic programming problem.

[0019] In one implementation, separating the first optimization problem from the objective function further includes: fixing the weight coefficient vector in the objective function. To obtain the predicted value sub-objective function Regarding the predicted value sub-objective function By taking the derivative and setting it to zero, we obtain the Sylvester equation to be solved.

[0020] In one implementation, solving the first optimization problem further includes: solving the Sylvester equation using a fast solution algorithm based on symmetric matrix tridiagonalization and tridiagonal matrix spectral decomposition.

[0021] In one implementation, the Sylvester equation is: ,in, , Graph Laplace matrix and It is a symmetric matrix. and All are symmetric matrices; solving the Sylvester equations includes: using the Householder transformation algorithm to transform the coefficient matrix... and Transform into a symmetric tridiagonal matrix and ,in, , ; Calculation of symmetric tridiagonal matrices based on the arrow divide-and-conquer method and symmetric tridiagonal matrix spectral decomposition ; Translate the tridiagonal matrix The spectral decomposition is substituted into the Sylvester equation and solved recursively.

[0022] The spectral decomposition of the symmetric tridiagonal matrix is ​​calculated based on the arrow divide-and-conquer method, including: The symmetric tridiagonal matrix Decomposed into two sub-matrices and Recursively find the two submatrices and Spectral decomposition, and then the symmetric tridiagonal matrix Disassembled into ,in Given a symmetric arrow matrix; determine the symmetric arrow matrix using the characteristic equation method. eigenvalues According to the symmetric arrow matrix eigenvalues and the symmetric arrow matrix diagonal matrix structure According to the above eigenvectors To approximate the above eigenvectors According to the above eigenvectors Derive the symmetric tridiagonal matrix eigenvectors Output the symmetric tridiagonal matrix. eigenvectors and the symmetric arrow matrix eigenvalues The spectral decomposition of the symmetric tridiagonal matrix is ​​obtained as follows: .

[0023] In one implementation, separating a second optimization problem from the objective function further includes: fixing the predicted values ​​in the objective function. To obtain the weight coefficient vector The second optimization problem is:

[0024] in, , .

[0025] In one implementation, solving the second optimization problem further includes: constructing a Lagrange function based on the second optimization problem; and using the KKT conditions of the Lagrange function to solve the optimal solution of the second optimization problem; wherein the KKT conditions are necessary conditions for the optimal solution of the nonlinear programming problem.

[0026] In one implementation, a Lagrangian function is constructed based on the second optimization problem:

[0027] in, and It is a Lagrangian multiplier.

[0028] In one implementation, the convergence condition for the iteration is that the absolute value of the difference between the objective function values ​​of two consecutive iterations is less than a set threshold. .

[0029] In one implementation, determining a multi-dimensional entity relationship network includes: determining a set of entities... An entity relationship network with multiple entities as nodes. ;in, As a dimension, , The total number of dimensions, For the total number of entities, , is used to indicate the relationships between the entities in an entity set.

[0030] In one implementation, determining a hierarchical tag relationship network includes: determining a set of tags A label relationship network of nodes with multiple labels in the data. ;in, For the total number of tags, This is used to indicate the relationship between the tags in a tag set.

[0031] In one implementation, the tag set is a hierarchical tag set, and the tag relationship network is formed as a directed acyclic graph.

[0032] In one implementation, an entity tag association network is generated. This includes: determining known associations between entities and labels. ;in, Representing entities With tags There is a positive correlation between them. represent and The relationship between them is unknown.

[0033] In one implementation, the method further includes: updating existing labeling information according to the hierarchical relationship between labels; wherein, if a specified entity is labeled with a specified label, then all ancestor labels of the specified label are added to the specified entity according to the label relationship network.

[0034] In one implementation, an entity similarity association matrix is ​​constructed based on the entity relationship network. It also includes: identifying entity relationship networks. adjacency matrix ;in, For dimension, in the first Adjacency matrix of the interaction network in each dimension The Middle Each element represents an entity. and The closeness of the relationship between them.

[0035] In one implementation, a tag similarity association matrix is ​​constructed based on a hierarchical tag relationship network. It also includes: based on currently known entity label association networks Calculate the information content of each tag Using the information content of each tag, calculate the similarity measure between a pair of tags. The generated tag similarity association matrix is ​​as follows: .

[0036] In one implementation, a similarity metric is determined. ,include: ,in, yes and The Most Informative Common Ancestor (MICA) is the common ancestor label node that has the highest information content between these two labels.

[0037] In one implementation, the information content of each tag is calculated using the following formula. : ,in, It is the total number of entities. It is known that it is tagged The number of entities marked.

[0038] In one implementation, an entity tag association matrix is ​​constructed based on the entity tag association network. It also includes: determining the entity label association matrix as follows: ;in, The An element equal to 1 represents an entity. With tags There must be a positive correlation between them, and a value of 0 indicates that the relationship between them is not yet clear.

[0039] In a second aspect, an entity tag association prediction apparatus is provided, configured to perform the method of the first aspect, comprising: a network determination module for determining... The system includes a multi-dimensional entity relationship network, a hierarchical tag relationship network, and an entity tag association network. The entity tag association network contains unknown entity tag association relationships. The integer is a positive integer greater than 1; the matrix construction module is used to construct an entity similarity association matrix based on the entity relationship network, a label similarity association matrix based on the hierarchical label relationship network, and an entity label association matrix based on the entity label association network; the model solving module is used to solve the model based on the entity similarity association matrix, the label similarity association matrix, and the entity label association matrix to obtain the prediction result; the objective function of the model introduces the weight coefficient vector. The regularization term constraint is applied to make the weight coefficient vector sparse. These are the parameters to be learned, used to indicate the... The contribution of the entity relationship network of each dimension to the model.

[0040] Thirdly, an entity tag association prediction apparatus is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform: the method of the first aspect.

[0041] Fourthly, a computer-readable storage medium is provided, which stores a program that, when executed by a multi-core processor, causes the multi-core processor to perform the method of the first aspect.

[0042] The at least one technical solution adopted in this application embodiment can achieve the following beneficial effects: In this embodiment, multiple entity relationship networks are specifically constructed based on the various connection relationships between entities. The model can more comprehensively capture the characteristics of the association relationships implied between entities. Furthermore, a regularization term constraint on the weight coefficient vector of the entity similarity network is added to the model's optimization objective, forcing the weight coefficient vector to have sparse characteristics, i.e., some weight values ​​are zero, thereby filtering out less relevant entity relationship networks and highlighting the contribution of high-quality information. Thus, the model can filter out valuable dimensional information, reduce the interference of noise information in the prediction results, and more effectively improve the final prediction accuracy of the model. In addition, it can achieve label filling for some already labeled samples, improving the completeness of entity label annotation.

[0043] It should be understood that the above description is merely an overview of the technical solution of the present invention, so as to enable a clearer understanding of the technical means of the present invention, and thus enable its implementation in accordance with the contents of the specification. To make the above and other objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are illustrated below. Attached Figure Description

[0044] By reading the detailed description of the exemplary embodiments below, those skilled in the art will understand the advantages and benefits described herein, as well as other advantages and benefits. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart illustrating an entity tag association prediction method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an entity relationship network according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a tag relationship network according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an entity tag association network according to an embodiment of the present invention; Figure 5 This is a flowchart illustrating an entity tag association prediction method according to another embodiment of the present invention; Figure 6 This is a schematic diagram of the tag relationship network in a document management website scenario according to an embodiment of the present invention; Figure 7This is a schematic diagram of the structure of an entity tag association prediction device according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of an entity tag association prediction device according to an embodiment of the present invention.

[0045] In the accompanying drawings, the same or corresponding reference numerals indicate the same or corresponding parts. Detailed Implementation

[0046] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0047] In the description of the embodiments of this application, it should be understood that terms such as “comprising” or “having” are intended to indicate the presence of features, numbers, steps, behaviors, components, portions or combinations thereof disclosed in this specification, and are not intended to exclude the possibility of the presence of one or more other features, numbers, steps, behaviors, components, portions or combinations thereof.

[0048] Unless otherwise stated, " / " means "or". For example, A / B can mean A or B. In this article, "and / or" is merely a way of describing the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone.

[0049] The terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of embodiments of this application, unless otherwise stated, "a plurality of" means two or more.

[0050] It should also be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0051] Figure 1 This is a flowchart illustrating an entity tag association prediction method according to an embodiment of this application, used to predict the association relationship between all entities and tags. In this process, from the perspective of the device, the executing entity can be one or more electronic devices; from the perspective of the program, the executing entity can be a program mounted on these electronic devices.

[0052] Figure 2 This is a schematic diagram of the algorithm architecture of an exemplary entity label association prediction model according to an embodiment of this application.

[0053] The following is for reference Figure 1 and Figure 2 The embodiments of this application will be described in detail.

[0054] like Figure 1 As shown, the method provided in this embodiment may include the following steps: S10, Confirm Entity relationship networks of various dimensions, hierarchical tag relationship networks, and entity tag association networks; Among them, the entity tag association network contains unknown entity tag association relationships; It is a positive integer greater than 1.

[0055] refer to Figure 2 Entity relationship network is based on entities As nodes, as entities An entity relationship network is a network of edges that indicate the relationships between entities. It is an undirected weighted network. For example, in a transaction system, users can be considered as entities, and relationships between entities can be established based on factors such as historical transactions between users, user locations, and user attributes.

[0056] For example, in real-world scenarios such as social media, users do not exist independently but rather within certain social relationships, thus they can be modeled as entity relationship networks. Since users exhibit a variety of behaviors, these entity relationship networks are often not singular. For instance, behaviors such as following, sharing, liking, commenting, and mentioning can all construct corresponding relationship networks. Each network reflects, from one perspective, the closeness of the social relationships between users; modeling based on only one type of relationship network will lead to a biased portrayal.

[0057] Based on this, in this embodiment, due to the existence of multiple connections between entities, a special modeling design was carried out to construct multiple entity relationship networks. Each network describes the similarity between entities from one dimension. For example, in Figure 2 There exist multiple dimensions of entity relationship networks: 1, ..., entity relationship networks. v In this way, the model can more comprehensively capture the characteristics of the relationships between entities, thus achieving superior predictive performance compared to using only the similarity of a single entity.

[0058] refer to Figure 3A hierarchical tag relationship network is a network with tags as nodes and the relationships between tags as edges, used to indicate the relationships between tags. In this embodiment, the tags are organized in the form of a Directed Acyclic Graph (DAG) to form a hierarchical structure, i.e., a hierarchical tag relationship network. The topmost tag is the "root tag," which has no actual meaning. Each intermediate tag below it has several child tags and several parent tags. The child tags and parent tags have a parent-child inheritance relationship (i.e., an "is-a" relationship), meaning that a child tag is a refinement of a parent tag, and a parent tag is a generalization of a child tag. For example, the tag "furniture" and the tag "bed" have a parent-child relationship, and the tag "bed" also has a parent-child relationship with the tags "double bed" and "single bed," thus forming a hierarchical tag relationship network.

[0059] refer to Figure 4 The entity label association network contains at least some of the associations between entities and labels. It is understood that due to the enormous size of the label library and the tens of thousands of samples, the existing annotations of the samples are often incomplete, with a large number of omissions. Therefore, in the existing entity label association network of S10, only some entity-label associations exist; a large number of unknown entity-label associations need to be predicted in subsequent models. Thus, the model in this embodiment aims to predict the labels of entities that are missing; in other words, to predict the degree of association between each entity and each label.

[0060] S20. Construct an entity similarity association matrix based on the entity relationship network, construct a tag similarity association matrix based on the hierarchical tag relationship network, and construct an entity tag association matrix based on the entity tag association network. Specifically, since the entity relationship network is an undirected weighted network, it already possesses the characteristics of a similarity graph. The entity similarity association matrix is ​​a homogeneous similarity graph constructed based on the multidimensional entity relationship network. For example, the entity similarity association matrix of each dimension can be determined directly from the adjacency matrix of the entity relationship network in each dimension, and is used to record the closeness of the relationship between each pair of entities.

[0061] Similarly, based on the hierarchical label relationship network and the entity label association network, the degree of closeness of the relationship between each pair of labels can be determined, thereby constructing the label similarity association matrix. Based on the entity label association network, the degree of closeness of the relationship between each entity and each label can be determined, thereby constructing the entity label association matrix.

[0062] S30. Solve the model based on the entity similarity association matrix, the label similarity association matrix, and the entity label association matrix to obtain the prediction results; The objective function of this model introduces the weight coefficient vector. The regularization term is used to ensure the sparsity of the weight coefficient vector, which consists of parameters to be learned and is used to indicate... The degree to which the entity relationship network contributes to the model. For example, The parameter representing the contribution of the first-dimensional entity relationship network to the model is... Representing the The contribution parameters of the dimensional entity relationship network to the model can be deduced similarly.

[0063] It's understandable that the quality of information in various entity relationship networks varies greatly. Some of this information may not be of substantial help in the final prediction, and some may even be noise, hindering the model's predictive performance. Take social media user networks as an example: to identify "AA" accounts, relationship networks such as follows, likes, and comments are likely more important because "AA" is often used to inflate these metrics. However, mentions in shares and posts may be less relevant. If the model indiscriminately includes all information and simply assigns a non-zero weight to the similarity information between each entity, irrelevant information or even noise may be introduced into the model, interfering with the prediction and ultimately harming its performance.

[0064] To address this, this embodiment adds a regularization constraint to the weight coefficient vector of the entity similarity network in the optimization objective. This forces the weight coefficient vector to have sparse characteristics, meaning some weights are zero, thereby filtering out less relevant entity relationship networks and highlighting the contribution of high-quality information. As a result, the model can filter out valuable dimensional information, reducing the interference of noise in the prediction results and more effectively improving the final prediction accuracy.

[0065] The following text provides a detailed explanation of the specific implementation of S10, with examples.

[0066] In one implementation, to make construction simpler The entity relationship network of various dimensions, S10 can specifically include: obtaining the entity set. ， in This represents the total number of entities; thus, the relationships between these entities can be obtained. Different dimensions of relationships are used to further determine the entity set. An entity relationship network in which multiple entities are nodes:

[0067] in, As a dimension, The total number of dimensions, , Used to indicate the relationships between the entities in an entity set.

[0068] Furthermore, this entity relationship network In the The adjacency matrix of the interaction network in each dimension is: , of which Each element represents an entity. and The degree of closeness of the relationship between two edges: the closer the edge weight is to 1, the closer the relationship between the two edges; the closer the edge weight is to 0, the more distant the relationship between the two edges; an edge weight of 0 indicates that there is no direct relationship between the two edges.

[0069] In one implementation, to more easily and effectively determine the hierarchical label relationship network, the above-mentioned S10 may specifically include: determining the label set The label relationship network of multiple nodes:

[0070] in, For the total number of tags, This is used to indicate the relationship between the tags in a tag set.

[0071] In one implementation, the tag set is a hierarchical tag set, and the tag relationship network is formed as a directed acyclic graph.

[0072] like Figure 3 As shown, the multiple labels in the above label set follow a hierarchical structure. Thus, the label relationship network actually constitutes a directed acyclic graph, that is, there are no directed cycles in the network, and only the root node has an in-degree of 0.

[0073] In one implementation, to generate an entity label association network, S10 may specifically include: determining each entity based on existing labeling information. and each tag The relationship between Therefore, the known entity-label association network between entities and labels is determined as follows:

[0074] in, Representing entities With tags There is a positive correlation between them. represent and The relationship between them is unknown.

[0075] In one implementation, based on the hierarchical tag set described above, the method further includes: updating existing annotation information according to the hierarchical relationship between tags; wherein, if a specified entity is annotated with a specified tag, then all ancestor tags of the specified tag are added to the specified entity according to the hierarchical relationship between tags.

[0076] Specifically, the known association between an entity and a label is denoted as ,in Representing entities With tags There must be a positive correlation between them, and Does not represent and There is no direct relationship between them; their relationship is simply currently unknown. Because tags follow a hierarchical structure, the relationship between entities and tags satisfies the "true path rule." If an entity is labeled by a certain tag, then it is labeled by all ancestor tags of that tag, meaning that the label can propagate upwards along the tag hierarchy.

[0077] It should be understood that S10 in this embodiment is not limited to the above implementation method, and can be implemented using any general relational network construction method.

[0078] The following text provides a detailed explanation of the specific implementation of S20 with examples.

[0079] In one implementation, in order to construct an entity similarity association matrix S20 may also specifically include: determining the entity similarity association matrix as an entity relationship network. Adjacency matrix:

[0080] in, For dimension, the value is... In the Adjacency matrix of the interaction network in each dimension The Middle Each element represents an entity. and The degree of closeness of the relationship between two edges is determined by the edge weight. The closer the edge weight is to 1, the closer the relationship between the two edges; the closer the edge weight is to 0, the more distant the relationship between the two edges; and an edge weight of 0 indicates that there is no direct relationship between the two edges.

[0081] In one implementation, a tag similarity association matrix is ​​constructed based on a hierarchical tag relationship network. S20 may also specifically include: (1) Based on the currently known entity label association network Calculate the information content of each tag ; Furthermore, the information content of each label can be determined using the following formula. :

[0082] in, It is the total number of entities. It is known that it is tagged The number of entities marked.

[0083] Of course, other formulas or methods can also be used to calculate the amount of label information for each label. Any formula or method that can indicate the number of entities labeled by the label can be used. This application does not impose any specific restrictions on this.

[0084] (2) Calculate the similarity measure between a pair of tags using the information content of each tag. Generate a tag similarity association matrix. .

[0085] Furthermore, the similarity metric can be determined using the following formula. :

[0086] in, yes and The Most Informative Common Ancestor (MICA) is the common ancestor label node that has the highest information content between these two labels.

[0087] Of course, other formulas or methods can also be used to calculate tag similarity, such as by judging... The determination is made based on whether there is a common ancestor tag and the distance between the two tags from the common ancestor tag. It is understood that any method or formula that can indicate the implicit closeness between tags can be used, and this application does not impose specific restrictions on it.

[0088] Through the above steps (1) and (2), the implicit similarity measure between tags was discovered, and a hierarchical tag relationship network that can express the similarity between tags was constructed, thereby improving the prediction accuracy of the prediction model.

[0089] In one implementation, in order to construct an entity label association matrix S20 may also include: determining the known entity label association matrix as follows:

[0090] in, The An element equal to 1 represents an entity. With tags There must be a positive correlation between them, and a value of 0 indicates that the relationship between them is not yet clear.

[0091] The optimization objective used in this embodiment, namely the objective function in S30 above, will be described in detail below with specific examples.

[0092] In short, the core idea of ​​the model solving algorithm in this embodiment is to make the predicted value approach the target value on elements with clearly known entity label relationships, while imposing smoothing constraints on the entity similarity association matrix and the label similarity association matrix, and adaptively adjusting the contribution of multiple dimensions of entity similarity association matrices to the smoothing constraint terms through regularization terms.

[0093] In one implementation, the objective function of the model can adopt any one or more of the following rules: (1) The first term of the objective function is used to predict the entity label association score matrix for elements whose entity label relationships are clearly known. Entity label association matrix Approaching.

[0094] (2) The second term of the objective function is used to predict the value. Apply a smoothing constraint to the entity similarity association matrix and use a weight coefficient vector. right The constraint terms of the entity similarity association matrix are linearly summed; the third term of the objective function affects the predicted value. Apply a smoothing constraint to the label similarity association matrix.

[0095] Thus, this embodiment adds label-side smoothing constraints to the optimization objective, explicitly introducing the association between labels. Utilizing the explicit hierarchical structure and semantic information between labels, it encourages similar labels to annotate the same entities. Through the coupling of smoothing constraints on both the entity and label sides, information is fully exchanged between the two types of nodes, which helps improve the final prediction performance. That is, similar entities can be labeled with the same label, similar labels can label the same entity, and the contribution of each entity similarity association matrix to the model's prediction can be dynamically adjusted.

[0096] (3) The fourth term of the objective function is used to adjust the weight coefficient vector. The regularization term that penalizes the modulus is used to constrain the weight coefficient vector to satisfy sparsity.

[0097] In this way, the contribution of the similarity association matrix of important entities is highlighted more effectively, while the interference of the similarity association matrix of irrelevant entities is suppressed.

[0098] It is understood that the terms of the objective function indicated in (1), (2), and (3) above can be combined to form the objective function, or they can be combined individually with other terms not shown to form the objective function.

[0099] In one implementation, in S30, the objective function of the model can be designed as follows:

[0100]

[0101] in, The predicted values ​​for the entity label association score matrix. Given the known entity label association matrix, Denotes the Frobenius norm of the matrix; where, Entity Relationship Network The normalized symmetric graph Laplace matrix, For a diagonal matrix, the first... diagonal elements for No. The line of lines, This is the entity similarity association matrix; where, For hierarchical label relationship network The graph Laplace matrix, diagonal matrix The Middle diagonal elements for No. The line of lines, Let be the tag similarity association matrix; where is the weight coefficient vector. For the parameters to be learned, their linear integration The degree to which the entity relationship network contributes to the model. Representing vectors Norm; , and These are hyperparameters used to balance the weights of each term in the objective function.

[0102] Figure 5 The optimized solution scheme for S30 above is shown below. Figure 5 The optimization scheme for solving the model is further explained in detail with specific examples.

[0103] refer to Figure 5 The model solution further includes the following steps: S301. Initialize the predicted values ​​of the entity label association score matrix. and weight coefficient vector ; The following S302 and S303 are iteratively solved repeatedly until convergence is achieved, yielding the prediction result. : S302. Solve the Sylvester equation and calculate the new predicted value. .

[0104] In one implementation, it is necessary to decompose the first optimization problem from the objective function, which can be done by fixing the weight coefficient vector in the objective function. Find the predicted value of the entity label association score matrix that minimizes the objective function. This is the first optimization problem.

[0105] For example, by combining the objective function described above, the weight coefficient vector in the objective function can be fixed. To obtain the predicted value sub-objective function :

[0106] Regarding the predicted value sub-objective function Taking the derivative, we get:

[0107] Setting the derivative to 0, we obtain the first optimization problem to be solved:

[0108] The above equation can be simplified as follows: ,in , Graph Laplace matrix and Both are symmetric matrices, therefore and Both are symmetric matrices. It can be seen that this is a typical Sylvester equation, and both coefficient matrices of this Sylvester equation are symmetric.

[0109] Based on the properties of the graph Laplacian matrix, for a normalized symmetric Laplacian matrix... Its eigenvalues ​​are all real numbers, and satisfy:

[0110] That is, all eigenvalues ​​are non-negative. And:

[0111] Therefore, it can be concluded that:

[0112] That is, matrix All eigenvalues ​​of the matrix are greater than or equal to 1. Similarly, it can be seen that the matrix... All eigenvalues ​​are non-negative, that is:

[0113] Therefore, for any and ,have

[0114] Therefore, according to the unique solution criterion theorem for the Sylvester equation, the first optimization problem to be solved must have a unique solution.

[0115] In one implementation, to improve the solution speed, S302 further includes the following (1)-(3): (1) Based on the Householder transformation algorithm, the coefficient matrix is ​​transformed. and Transform into a symmetric tridiagonal matrix and ,in, , ; For example, the following gives the coefficient matrix and Specific examples of tridiagonalization: First, we introduce the Householder transformation. The Householder matrix is ​​defined as follows:

[0116] Where, vector This is called the Householder vector, and the vector is usually... The corresponding matrix Recorded as .

[0117] Given a non-zero vector There must exist a Householder matrix. Make ,in, (or ), Let the vector be a unit vector. This theorem states that a vector can be reduced to zero by using a Householder matrix, with all elements except the first element set to zero. Generally, vectors are... Called The corresponding Householder vector.

[0118] The following is for finding a given vector Corresponding Householder vector , making ,in Algorithm 1 flow.

[0119]

[0120] So, to make the matrix Tridiagonalization allows us to sequentially transform the matrix... Take out the elements below the main diagonal of each column, find the Householder matrix corresponding to the vector, and after matrix transformation, retain only the elements on the second diagonal of the column, while all the elements below it are reduced to zero.

[0121] The following is the matrix Algorithm 2 for converting to a tridiagonal matrix.

[0122]

[0123] Algorithm 2 can be used to obtain the coefficient matrix of the equation to be solved. and Transform it into a symmetric tridiagonal matrix, that is , .

[0124] (2) Calculation of symmetric tridiagonal matrices based on the arrow divide-and-conquer method and symmetric tridiagonal matrix spectral decomposition ; For example, the following uses a matrix For example, a specific illustrative explanation of calculating the spectral decomposition of a symmetric tridiagonal matrix is ​​given, for the matrix... The calculation methods are exactly the same.

[0125] For a symmetric tridiagonal matrix For ease of representation, let's remember...

[0126] The eigenvalue problem of a symmetric tridiagonal matrix is ​​to find the eigenvalues ​​of the matrix. Spectral decomposition ,in It is a diagonal matrix, and its diagonal elements are... eigenvalues. It is an orthogonal matrix, and each column is a corresponding eigenvector.

[0127] The core idea of ​​the arrowhead divide-and-conquer (ADC) method is to break down a large-scale matrix into smaller submatrices and then use recursion to find eigenvalues ​​and eigenvectors.

[0128] Specifically, the matrix Break it down into two subproblems

[0129] in, , and They are respectively of order and Order submatrix, Represents the first dimension under the corresponding dimension A unit vector. Typically, Pick If the result has already been obtained recursively... spectral decomposition Substituting into the above formula, we get

[0130] in, express The last line, express The first line, This is called a symmetric arrowhead matrix. Therefore, we only need to find... spectral decomposition Then we can find out eigenvalues eigenvectors .

[0131] The following section describes how to calculate symmetric arrow matrices. The method of spectral decomposition. For ease of representation, let's denote it as...

[0132] in, for A diagonal matrix of order 1, without loss of generality. . For length The vector. It is a scalar.

[0133] (I) First step, find eigenvalues .

[0134] According to the Wilkinson-O'Leary-Stewart theorem eigenvalues Satisfying the alternation property

[0135] And it is the characteristic equation (secular equation).

[0136] The root.

[0137] Below is about the root The solution is discussed by classifying and solving different problems.

[0138] ① Consider ,in The situation.

[0139] (i) If ,but Therefore, we only need to find the equation.

[0140] exist The root You can then base it on Calculate the eigenvalues.

[0141] (ii) If ,but Therefore, we only need to find the equation.

[0142] exist The root You can then base it on Calculate the eigenvalues.

[0143] ② Consider beg The situation.

[0144] Only need to find the equation

[0145] exist The root You can then base it on Calculate the eigenvalues.

[0146] ③ Consider beg The situation.

[0147] Only need to find the equation

[0148] exist The root You can then base it on Calculate the eigenvalues.

[0149] Generally, the above equations can be solved using Newton's iteration method, which usually only requires a few iterations to converge.

[0150] Thus, the answer has been found. eigenvalues , that is eigenvalues .

[0151] (II) Second step, according to eigenvalues Find the eigenvectors .

[0152] First, according to the Boley-Golub theorem, given the eigenvalues ​​calculated in the previous step... and diagonal matrix Another symmetric arrow matrix can be constructed.

[0153] Its eigenvalue is exactly .here, Depend on

[0154] Give, The symbol can be chosen arbitrarily; here it is specified that... Scalar Depend on

[0155] Provided.

[0156] Next, according to the Wilkinson-O'Leary-Stewart theorem, the constructed matrix is ​​obtained. eigenvectors. For eigenvalues Its corresponding feature vector is

[0157] Finally, take eigenvectors Approximate representation eigenvectors .

[0158] Thus, the answer has been found. eigenvectors And from this, it can be derived eigenvectors .

[0159] The following is the algorithm flow of the arrow divide-and-conquer method:

[0160] Algorithm 3 can be used to calculate the symmetric tridiagonal matrix in the previous step. and Spectral decomposition into and .

[0161] (3) Transform the tridiagonal matrix The spectral decomposition is substituted into the Sylvester equation and solved recursively.

[0162] Furthermore, according to the spectral decomposition theorem for real symmetric matrices, there exist orthogonal matrices. and ,make and ,in They are respectively based on and The eigenvalues ​​are diagonal matrices with diagonal elements. It is by and The orthogonal matrix formed by the eigenvectors. Based on the aforementioned tridiagonalization and spectral decomposition algorithms, it can be seen that... , .

[0163] Therefore, the above Sylvester equation can be interpreted as follows: Convert to

[0164] in, , .

[0165] Decompose the equation into a recursive form, i.e.

[0166] In the formula for The first column vector, For all columns except the first column 1-th order matrix for The first column vector, For all columns except the first column 1-th order matrix for The element in the first row and first column, for The rest The order submatrix. This formula can be equivalently expressed as:

[0167] It can be seen that the above equation has been transformed into a linear homogeneous matrix equation, which can be solved recursively. Therefore, it can be deduced that...

[0168] Therefore, the solution to the original equation is obtained. .

[0169] The following is a summary of the algorithm for solving the Sylvester equation in this embodiment:

[0170] S303. Solve the quadratic programming problem and calculate the new weight coefficient vector. .

[0171] In one implementation, a second optimization problem needs to be derived from the objective function, specifically by fixing the predicted values ​​in the objective function. Find the weight coefficient vector that minimizes the objective function. This is the second optimization problem.

[0172] For example, by combining the objective function described above, the predicted values ​​in the objective function can be fixed. This yields the vector of weight coefficients. The second optimization problem is:

[0173] in, , .

[0174] As can be seen, the second optimization problem mentioned above is a typical convex quadratic programming (QP) problem, which can be solved using any general-purpose QP solver.

[0175] Furthermore, to construct a simpler algorithm for finding the optimal solution, the following algorithm can also be used to solve the second optimization problem in S303: (1) Construct the Lagrangian function based on the second optimization problem; For example, construct the following Lagrange function:

[0176] in, and It is a Lagrangian multiplier.

[0177] (2) Use the KKT (Karush-Kuhn-Tucker conditions) of the Lagrange function to solve the optimal solution of the second optimization problem; where the KKT conditions are necessary conditions for the optimal solution of the nonlinear programming.

[0178] For example, the optimal solution of the above Lagrange function should satisfy the following KKT conditions:

[0179] Regarding the stability condition, we know that: Solving for the problem yields: .

[0180] The following discussion will categorize the cases. ① If Due to dual feasibility conditions ,Know Therefore, based on the condition of complementary slackness... , can be obtained ②If Due to primal feasibility conditions , there must be Therefore, based on the complementary relaxation condition... ,have to ③If Due to complementary relaxation conditions It can be known that .

[0181] Based on the above three situations, we can conclude that The optimal solution satisfies the following rule:

[0182] Without loss of generality, it can be applied to Sort in ascending order to get ,in And record The elements in the original vector The index sequence formed by the corresponding subscripts in is For example, suppose The result after sorting it in ascending order is: index sequence .

[0183] Therefore, the process of finding the optimal solution is transformed into: finding the index , making and satisfy .

[0184] Then, calculate the following formula:

[0185] Finally, According to the index sequence Rearrange to obtain a vector And satisfy .then, This is the optimal solution required.

[0186] The following is about finding the optimal solution. A summary of the algorithm process.

[0187]

[0188] At this point, the prediction of the value has been completed. Subproblems and about the weight coefficient vector Solving subproblems.

[0189] Subsequently, S302 and S303 are iteratively solved using the alternating least squares method until convergence, and the prediction results are obtained.

[0190] For example, the following is a summary of the algorithm flow for solving the problem using the alternating least squares method.

[0191]

[0192] In one implementation, the convergence condition for the iteration is that the absolute value of the difference between the objective function values ​​of two consecutive iterations is less than a set threshold. .

[0193] Figure 6 A schematic diagram of an exemplary model algorithm architecture is shown. In short, in the above embodiment, the alternating least squares method is used to solve the model, decomposing the objective function into a first optimization problem: solving the Sylvester equation, and a second optimization problem: solving a quadratic programming problem. Furthermore, for the first optimization problem, a fast solution algorithm is proposed using matrix transformations such as tridiagonalization and spectral decomposition, significantly improving computational speed compared to existing general-purpose algorithms. For the second optimization problem, starting from the KKT conditions of nonlinear programming, a simplified solution algorithm is derived, which can directly obtain the optimal solution, avoiding a large amount of tedious iterative calculations in traditional methods and greatly reducing computation time.

[0194] Therefore, the time complexity of this embodiment is significantly reduced. In addition, the optimization variables of this algorithm are only two: the entity label association matrix to be predicted and the weight parameter vector of the entity relationship network. The space complexity is also significantly reduced, and less storage resources are consumed.

[0195] The entity tag association prediction method described in the above embodiments can be applied to multiple practical application scenarios. Taking a document management website as an example, operators often need to classify articles according to their type.

[0196] Figure 6 The document provides an excerpt illustrating the hierarchical relationship between article categories. According to this classification rule, a technical and applied science article can be categorized into sections such as "General Technology," "Computational Technology," and "Engineering Technology." "Engineering Technology" can be further subdivided into categories such as "Materials Science" and "Chemical Engineering," while "Chemical Engineering" can be further divided into categories such as "Organic Chemistry" and "Inorganic Chemistry." Specifically, an article can be classified into multiple categories across multiple levels. For example, the article "Design and Synthesis of Multi-element Heterogeneous Structures Accelerated by Machine Learning" could be tagged with "Technology and Applied Science," "Computational Technology," "Engineering Technology," and "Materials Science," among others.

[0197] Typically, an article includes a title, author, abstract, keywords, and citations. Based on this information, various relationship networks between articles can be constructed.

[0198] For example, for titles, popular deep text embedding extraction models such as Doc2Vec, Transformer, BERT, and GPT-3 can be used to generate dense vector representations of the titles, thus constructing an article association network based on title semantic similarity. For abstracts and keywords, the same method can be used to construct two other types of article association networks. For authors, an author-based article relationship network can be established, using articles as nodes and author participation relationships as edges. Since cited articles generally have a strong connection to the cited article, an article citation relationship network can be established using inter-article citation relationships as edges. Thus, various inter-article relationship networks are constructed from different perspectives.

[0199] Based on this, the missing annotations in the article can be predicted and filled using the above S10-S30.

[0200] First, in S10, the input consists of a network of relationships between various articles, a maintained hierarchical structure of text categories, and an existing set of text labels.

[0201] Subsequently, in S20, according to the association matrix construction method proposed in this embodiment, an article similarity association matrix, a tag similarity association matrix, and an entity tag association matrix are generated.

[0202] Finally, in S30, the constructed correlation matrix is ​​input into the model, and the model is solved iteratively until convergence, thus obtaining the predicted missing text label annotations.

[0203] In the description of this specification, the references to terms such as "some possible implementations," "some embodiments," "examples," "specific examples," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0204] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0205] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0206] Regarding the method flowcharts of embodiments of this application, certain operations are described as different steps performed in a certain order. Such flowcharts are illustrative and not restrictive. Some steps described herein may be grouped together and performed in a single operation, some steps may be divided into multiple sub-steps, and some steps may be performed in an order different from that shown herein. The various steps shown in the flowcharts may be implemented in any way by any circuit structure and / or tangible mechanism (e.g., software running on a computer device, hardware (e.g., logic functions implemented by a processor or chip), and / or any combination thereof).

[0207] Based on the same technical concept, embodiments of the present invention also provide an entity label association prediction device for executing the entity label association prediction method provided in any of the above embodiments. Figure 7 This is a schematic diagram of an entity tag association prediction device provided in an embodiment of the present invention.

[0208] The network determination module 701 is used to determine multi-dimensional entity relationship networks, hierarchical label relationship networks, and entity label association networks. The entity label association network includes unknown entity label association relationships. It is a positive integer greater than 1; The matrix construction module 702 is used to construct an entity similarity association matrix based on the entity relationship network, construct a label similarity association matrix based on the hierarchical label relationship network, and construct an entity label association matrix based on the entity label association network. Model solving module 703 is used to solve the model based on the entity similarity association matrix, label similarity association matrix, and entity label association matrix to obtain the prediction result; wherein, the objective function of the model introduces the weight coefficient vector The regularization term constraint is applied to ensure that the weight coefficient vector is sparse. These are the parameters to be learned, used to indicate the contribution of the entity relationship network in each dimension to the model.

[0209] It should be noted that the apparatus in the embodiments of this application can implement the various processes of the aforementioned method embodiments and achieve the same effects and functions, which will not be repeated here.

[0210] Figure 8 This is a schematic diagram of the structure of an entity tag association prediction device according to an embodiment of this application, used to perform... Figure 1 The method shown includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method described in the above embodiments.

[0211] According to some embodiments of this application, a non-volatile computer storage medium is provided for an entity label association prediction method, having stored thereon computer-executable instructions configured to execute, when run by a processor, the method described in the above embodiments.

[0212] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the descriptions of the apparatus, device, and computer-readable storage medium embodiments are simplified because they are substantially similar to the method embodiments; relevant details can be found in the descriptions of the method embodiments.

[0213] The apparatus, device, and computer-readable storage medium provided in the embodiments of this application are one-to-one corresponding to the method. Therefore, the apparatus, device, and computer-readable storage medium also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the apparatus, device, and computer-readable storage medium will not be repeated here.

[0214] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0215] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0216] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0217] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0218] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0219] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0220] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device. Furthermore, although the operations of the method of the invention are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0221] While the spirit and principles of the invention have been described with reference to several specific embodiments, it should be understood that the invention is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for ease of description. The invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A method for predicting entity label association, characterized in that, include: A multi-dimensional entity relationship network, a hierarchical tag relationship network, and an entity tag association network are defined. The entity relationship network uses users as entities and can be constructed based on user behaviors such as following, sharing, liking, commenting, and mentioning. The hierarchical tag relationship network forms a directed acyclic graph: if an entity is labeled with a tag, then the entity is labeled with all ancestor tags of that tag. The entity tag association network contains unknown entity tag association relationships, and the tags are used to describe user characteristics. An entity similarity association matrix is ​​constructed based on the entity relationship network; a tag similarity association matrix is ​​constructed based on the hierarchical tag relationship network; and an entity tag association matrix is ​​constructed based on the entity tag association network. The model is solved based on the entity similarity association matrix, the label similarity association matrix, and the entity label association matrix to obtain the prediction result; The objective function of the model introduces a weight coefficient vector. The regularization term constraint is applied to make the weight coefficient vector sparse. These are the parameters to be learned, used to indicate the contribution of the entity relationship network in each dimension to the model.

2. The method according to claim 1, characterized in that, Solving the model also includes: Obtain the predicted values ​​of the initialized entity label association score matrix. and weight coefficient vector ; The following problem is solved iteratively using the alternating least squares method until convergence, and the prediction results are obtained: Solve the first optimization problem derived from the objective function, and calculate and update the predicted value. ; Solving the objective function decomposes into a second optimization problem, and the weight coefficient vector is then calculated and updated. .

3. The method according to claim 2, characterized in that, The first term of the objective function is used to predict the value of the entity label association score matrix for elements whose entity label relationships are explicitly known. Entity label association matrix Approaching.

4. The method according to claim 2, characterized in that, The second term of the objective function relates to the predicted value. Apply smoothing constraints to the entity relationship network, and use the weight coefficient vector. right The constraint terms on the entity-relationship network of each dimension are linearly summed. The third term of the objective function relates to the predicted value. Apply smoothing constraints to the hierarchical label relationship network.

5. The method according to claim 2, characterized in that, The fourth term of the objective function is used to adjust the weight coefficient vector. The modulus is used to apply a regularization term to constrain the weight coefficient vector to satisfy sparsity.

6. The method according to claim 1, characterized in that, Also includes: The objective function of the model is determined according to the following formula: in, The predicted values ​​for the entity label association score matrix. This is an entity label association matrix. For the total number of entities, For the total number of tags, Denotes the Frobenius norm of a matrix; in, Entity Relationship Network The normalized symmetric graph Laplace matrix, For a diagonal matrix, the first... diagonal elements for No. The line of lines, This is an entity similarity association matrix. As a dimension, the The The total number of dimensions; in, For hierarchical label relationship network The graph Laplace matrix, diagonal matrix The Middle diagonal elements for No. The line of lines, This is a tag similarity correlation matrix; Wherein, the weight coefficient vector For the parameters to be learned, their linear integration The degree to which the entity relationship network contributes to the model. Representing vectors Norm; , and These are hyperparameters used to balance the weights of each term in the objective function.

7. The method according to claim 2, characterized in that, Also includes: Fix the weight coefficient vector in the objective function Find the predicted value of the entity label association score matrix that minimizes the objective function. This serves as the first optimization problem; And / or, Fix the predicted value in the objective function Find the weight coefficient vector that minimizes the objective function. This serves as the second optimization problem.

8. The method according to claim 2, wherein, The first optimization problem is the problem of solving the Sylvester equation.

9. The method according to claim 2, wherein, The second optimization problem is a quadratic programming problem.

10. The method according to claim 7, further comprising decomposing the first optimization problem from the objective function: Fix the weight coefficient vector in the objective function To obtain the predicted value sub-objective function : Regarding the predicted value sub-objective function By taking the derivative and setting it to zero, we obtain the Sylvester equation to be solved.

11. The method according to claim 8, characterized in that, Solving the first optimization problem also includes: The Sylvester equation is solved using a fast solution algorithm based on tridiagonalization of symmetric matrices and spectral decomposition of tridiagonal matrices.

12. The method according to claim 10, wherein the Sylvester equation is: ,in, , , and For hyperparameters, As a dimension, the The The total number of dimensions; the Graph Laplacian matrix and It is a symmetric matrix. and All are symmetric matrices; Solving the Sylvester equations includes: Based on the Householder transformation algorithm, the coefficient matrix is... and Transform into a symmetric tridiagonal matrix and ,in, , ; The symmetric tridiagonal matrix is ​​calculated based on the arrow divide-and-conquer method. and the symmetric tridiagonal matrix spectral decomposition ; They are respectively based on and The eigenvalues ​​are diagonal matrices with diagonal elements; The symmetric tridiagonal matrix The spectral decomposition is substituted into the Sylvester equation and solved recursively.

13. The method according to claim 12, wherein calculating the spectral decomposition of the symmetric tridiagonal matrix based on the arrow divide-and-conquer method comprises: Will The symmetric tridiagonal matrix Decomposed into two sub-matrices and Recursively find the two submatrices and Spectral decomposition, and then the symmetric tridiagonal matrix Disassembled into ,in It is a symmetric arrow matrix; Determining symmetric arrow matrix using the characteristic equation method eigenvalues ; According to the symmetric arrow matrix eigenvalues and the symmetric arrow matrix diagonal matrix structure According to the above eigenvectors To approximate the above eigenvectors ; According to the above eigenvectors Derive the symmetric tridiagonal matrix eigenvectors ; Output the symmetric tridiagonal matrix eigenvectors and the symmetric arrow matrix eigenvalues The spectral decomposition of the symmetric tridiagonal matrix is ​​obtained as follows: .

14. The method according to claim 7, characterized in that, Decomposing the second optimization problem from the objective function also includes: Fix the predicted value in the objective function To obtain the weight coefficient vector The second optimization problem is: in, , The The total number of dimensions; This represents the predicted value in the objective function; For hyperparameters, Representing vectors Norm.

15. The method according to claim 13, characterized in that, Solving the second optimization problem also includes: Construct a Lagrangian function based on the second optimization problem; The optimal solution to the second optimization problem is obtained by using the KKT conditions of the Lagrange function; wherein the KKT conditions are necessary conditions for the optimal solution of the nonlinear programming problem.

16. The method according to claim 14, characterized in that, Construct the Lagrange function based on the second optimization problem: in, and It is a Lagrangian multiplier.

17. The method according to claim 2, characterized in that, The convergence condition for the iteration is that the absolute value of the difference between the objective function values ​​of two consecutive iterations is less than a set threshold. .

18. The method according to claim 1, characterized in that, Determine a multi-dimensional network of entity relationships, including: Determine the entity set An entity relationship network with multiple entities as nodes. ; in, As a dimension, the The The total number of dimensions, For the total number of entities, the This is used to indicate the relationships between the various entities in the entity set.

19. The method according to claim 1, characterized in that, Determining the hierarchical label relationship network includes: Determine the set of tags A label relationship network of nodes with multiple labels in the data. ; in, For the total number of tags, the This is used to indicate the relationship between the various tags in the tag set.

20. The method according to claim 19, characterized in that, The tag set is a hierarchical tag set, and the tag relationship network is formed as a directed acyclic graph.

21. The method according to claim 1, characterized in that, Generate entity label association network ,include: Determine the known associations between entities and labels ; For the total number of entities, Total number of tags; in, Representing entities With tags There is a positive correlation between them. represent and The relationship between them is unknown.

22. The method according to claim 19, characterized in that, The method further includes: Update the existing labeling information according to the hierarchical relationship between the labels; If a specified entity is labeled with a specified tag, then all ancestor tags of the specified tag are added to the specified entity according to the tag relationship network.

23. The method according to claim 1, characterized in that, Construct an entity similarity association matrix based on the entity relationship network. It also includes: Determine the entity relationship network The adjacency matrix is ​​used as the adjacency matrix of the above. ; The total number of entities; in, For dimension, in the first Adjacency matrix of the interaction network in each dimension The Middle Each element represents an entity. and The closeness of the relationship between them.

24. The method according to claim 19, characterized in that, A tag similarity association matrix is ​​constructed based on the hierarchical tag relationship network. It also includes: Based on currently known entity label association networks Calculate the information content of each tag ; Calculate the similarity measure between a pair of tags using the information content of each tag. ; The generated tag similarity association matrix is ​​as follows .

25. The method of claim 24, wherein the similarity metric is determined. ,include: in, yes and The most informative common ancestor, that is, the common ancestor label node with the highest information content between these two labels.

26. The method of claim 25, wherein the information content of each tag is calculated using the following formula. : in, It is the total number of entities. It is known that it is tagged The number of entities marked.

27. The method according to claim 19, characterized in that, Construct an entity label association matrix based on the entity label association network. It also includes: Determine the entity label association matrix as follows ; For the total number of entities, Total number of tags; in, The An element equal to 1 represents an entity. With tags There must be a positive correlation between them, and a value of 0 indicates that the relationship between them is not yet clear.

28. An entity tag association prediction device, characterized in that, Configured to perform the method as described in any one of claims 1-27, comprising: The network determination module is used to determine multi-dimensional entity relationship networks, hierarchical tag relationship networks, and entity tag association networks. The entity relationship network uses users as entities and can be constructed based on user behaviors such as following, sharing, liking, commenting, and mentioning. The hierarchical tag relationship network is formed as a directed acyclic graph: if an entity is labeled with a tag, then the entity is labeled with all ancestor tags of that tag. The entity tag association network contains unknown entity tag association relationships, and the tags are used to describe user characteristics. The matrix construction module is used to construct an entity similarity association matrix based on the entity relationship network, construct a tag similarity association matrix based on the hierarchical tag relationship network, and construct an entity tag association matrix based on the entity tag association network. The model solving module is used to solve the model based on the entity similarity association matrix, the label similarity association matrix, and the entity label association matrix to obtain the prediction result; wherein, the objective function of the model introduces a weight coefficient vector. The regularization term constraint is applied to make the weight coefficient vector sparse. These are the parameters to be learned, used to indicate the contribution of the entity relationship network in each dimension to the model.

29. An entity tag association prediction device, characterized in that, include: At least one processor; And a memory communicatively connected to at least one processor; wherein the memory stores instructions executable by at least one processor, the instructions being executed by at least one processor to enable at least one processor to perform: the method as described in any one of claims 1-27.

30. A computer-readable storage medium storing a program that, when executed by a multi-core processor, causes the multi-core processor to perform the method as described in any one of claims 1-27.

Images