Method and device for predicting author cooperation link of citation network based on multi-layer prediction sampling

Abstract

The invention discloses a quotation network author cooperation link prediction method and device based on multilayer prediction sampling. The method comprises the following steps: obtaining known data of an original quotation network; based on known data of the original citation network, constructing a citation network author relation tensor and a hidden feature matrix of a multi-layer citation network author link; based on the citation network author relation tensor, constructing a target loss function of citation network author relation prediction, and adopting the target loss function to train and iteratively optimize the hidden feature matrix of each layer of citation network author link; and based on the trained hidden feature matrix of each layer of citation network author link, obtaining a prediction result of the citation network author relation link. The method is specially applied to tensor author relation data, and author relation link prediction which accords with a statistical rule and is high in accuracy can be carried out. The method can be widely applied to the fields of information retrieval, network analysis and the like.

Description

Technical Field

[0001] The present invention relates to the field of information retrieval and network analysis, and in particular to a method and device for predicting author collaboration links in a citation network based on multi-layer prediction sampling. Background Art

[0002] With the rapid development of various networks, citation network embedding has gradually become a hot topic of research, aiming to retain the structure and semantic information of the network while learning node representation. Citation networks contain various types of nodes and diverse relationships between them, which are widely present in the real world. As an important part of the citation network, there are many associations between author nodes, for example: two authors co-author a paper, or their papers contain the same keywords, etc. Through the citation network embedding representation, we can predict the cooperative links between authors. Therefore, how to effectively use the various node relationships in the citation network to predict author cooperative links has become a key issue that needs to be solved urgently.

[0003] Currently, many researchers usually solve the link prediction problem by extracting features and training models to calculate the weight probability of links. However, these methods often rely on the selection of meta-paths, which largely depends on domain knowledge. In addition, the training of neural network models usually requires a lot of computing resources and data. When dealing with large-scale networks, the limitations of computing power and data acquisition become significant bottlenecks. Citation network data in reality is often very sparse, which affects the performance and prediction accuracy of the model. Therefore, how to effectively predict author collaboration links in citation networks and improve the accuracy and reliability of predictions has become a focus of researchers. Summary of the invention

[0004] Based on this, the technical problem to be solved by the present invention is to provide a method and device for predicting author collaboration links in a citation network based on multi-layer predictive sampling.

[0005] In a first aspect, an embodiment of the present application provides a method for predicting author collaboration links in a citation network based on multi-layer prediction sampling, comprising the following steps:

[0006] S1: Obtain known data of the original citation network;

[0007] S2: Based on the known data of the original citation network, construct the citation network author relationship tensor and the latent feature matrix of the multi-layer citation network author links;

[0008] S3: Based on the citation network author relationship tensor, construct a target loss function for citation network author relationship prediction, use the target loss function to train the latent feature matrix of each layer of citation network author links and iteratively optimize the target loss function;

[0009] S4: Based on the latent feature matrix of the author links in each layer of the citation network after training, the prediction results of the author relationship links in the citation network are obtained.

[0010] Preferably, the known data of the citation network is stored in the form of a quaternary group, and the quaternary group is represented by W=(a 1 ,a 2 ,rel,l);

[0011] Among them, a 1 and a 2 They represent different authors in the citation network, and rel represents author a 1 and author a 2 The meta-path relationship between them, l represents the author a 1 and author a 2 The relation link value of the collaboration by the author of the meta path relation rel.

[0012] Preferably, step S2 comprises:

[0013] S21: All the quadruples W in the known data of the citation network are 1 ,a 2 ,rel,l) Divide the data into K time periods according to the meta-path relationship rel;

[0014] S22: Quadruple W based on rel=1 (1) =(a 1 ,a 2 ,1,l) to get W (1) Data (a 1 ,a 2 ,l), and according to W (1) Construct a slice matrix T with meta-path relation rel=1 (1) ;

[0015] S23: According to the remaining element path relations rel, based on the above steps, from the four-tuple W=(a 1 ,a 2 ,rel,l) in turn select the four-tuple W (2) , W (3) , …, W (K) , respectively construct the slice matrix T (2) , T (3) ,…,T (K) ;

[0016] S24: Arrange the K slice matrices in order from small to large according to the order of the divided meta-path relationships, and construct a citation network author relationship tensor Y;

[0017] S25: Based on the known data of the citation network, construct the latent feature matrices G, D, Q of the author links of the multi-layer citation network;

[0018] Among them, the slice matrix T (1) The size of the slice matrix T is I rows and J columns, where I and J are the number of authors, K is a positive integer, and (1) The element T in (1) ij represents the relationship link value between the i-th author and the j-th author under the first relationship, 1≤i≤I, 1≤j≤J, Y∈R I×J×K , R represents a set of real numbers; among them, the size of the latent feature matrix G is I rows and R columns, the size of the latent feature matrix D is J rows and R columns, and the size of the latent feature matrix Q is K rows and R columns.

[0019] Preferably, step S3 comprises:

[0020] S31: Initialize the process parameters involved in the prediction of author relationship links in the citation network;

[0021] S32: Construct the target loss function for citation network author relationship link prediction based on the citation network author relationship tensor Y and the latent feature matrices G, D, and Q of the multi-layer citation network author links;

[0022] S33: using the target loss function to train the latent feature matrix of the author links of each layer of the citation network and iteratively optimize the target loss function;

[0023] S34: Determine whether the target loss function satisfies the iteration termination condition. If so, terminate the iteration and output the target loss function as the target loss function after iterative optimization; if not, continue the iteration.

[0024] Preferably, the process parameters include: citation network author relationship tensor Y, latent feature matrices G, D, Q; rank R that determines the dimension of the latent feature space of the latent feature matrices G, D, Q; number of multi-layer sampling layers N; maximum number of training iterations max Round; iterative round number control variable round during training; and convergence termination threshold τ.

[0025] Preferably, step S32 includes:

[0026] S321: traverse the multi-layer sampling layer number n∈{1,…,N}, for each layer of the constructed citation network author relationship tensor Y and the process parameters, use the objective function to train and obtain the latent feature matrix G of each layer of the citation network author connection n ,D n ,Q n ;

[0027] S322: Iteratively optimize the objective function using the Euclidean distance as the optimization target.

[0028] Preferably, step S22 specifically uses the SGD algorithm to calculate the latent feature matrix G of the nth layer. n ,D n ,Q n Training iterations, including:

[0029] (1) Fixed D n and Q n , train G through formula (1) n :

[0030]

[0031] (2) Fixed G n and Q n , train D through formula (2) n :

[0032]

[0033] (3) Fixed G n and D n , train Q by formula (3) n :

[0034]

[0035] Among them, ← is the assignment operation, G n ,D n ,Q n Represent the hidden feature matrix corresponding to the nth layer, Represents the hidden feature matrix G of the nth layer respectively n ,D n ,Q n The corresponding single element, η is the learning rate, n is the number of sampling layers; y ijk represents the relationship link value of the collaboration between the i-th author and the j-th author on the k-th meta-path relationship in the citation network author relationship tensor Y; represents the final predicted value of the citation network generated data;

[0036] Where ε is the target loss function, which is expressed by the following formula (4):

[0037]

[0038] Among them, the target loss function ε is the instantaneous loss of a single element in the case of known data in the citation network and data generated by the citation network at the nth layer by the following formula:

[0039]

[0040] Among them, α and β represent weights, For unknown authorg i In the corresponding meta-path relation q k Link author d j The relationship link prediction value is expressed by the following formula:

[0041]

[0042] Preferably, the iteration termination condition of step S34 is that the iteration round control variable round reaches the maximum training iteration round max Round or the difference between the prediction result of the citation network in the previous round and the prediction result of the citation network after the current round is less than the convergence termination threshold τ.

[0043] Preferably, the step S4 comprises:

[0044] The K slice matrices into which the citation network author relationship tensor Y is divided are sequentially traversed through each row, a blank position is randomly selected from the known data of the two citation networks, unknown author relationships are predicted and relationship connection prediction values ​​are obtained, and then the relationship link prediction values ​​are mapped through a nonlinear activation function to obtain a final prediction value;

[0045] Continuously select new unknown author relationships by slice and row order and repeat the above steps until the preset conditions are met, and all the final prediction values ​​and their unknown author relationships are used to generate citation network data;

[0046] Among them, the relationship connection prediction value is mapped by a nonlinear activation function as follows:

[0047]

[0048] in, is the final prediction value; the preset condition is that the amount of data generated by the citation network is the same as the amount of data known in the citation network.

[0049] In a second aspect, an embodiment of the present application provides a citation network author collaboration link prediction device based on multi-layer prediction sampling, comprising:

[0050] A citation network known data acquisition module is used to acquire original citation network known data;

[0051] The citation network author relationship tensor construction module is used to construct the citation network author relationship tensor and the latent feature matrix of multi-layer citation network author links based on the known data of the original citation network;

[0052] A citation network training module, for constructing a target loss function for citation network author relationship prediction based on the citation network author relationship tensor, using the target loss function to train the latent feature matrix of each layer of citation network author links and iteratively optimizing the target loss function;

[0053] The citation network author relationship link prediction module is used to obtain the prediction results of the citation network author relationship link based on the latent feature matrix of each layer of the citation network author link after training.

[0054] Compared with the prior art, the present invention has the following beneficial effects: the present invention uses known data of the citation network to construct a citation network author relationship tensor and a latent feature matrix of multi-layer citation network author links, constructs a target loss function based on the citation network author relationship tensor, and trains and iteratively optimizes the latent feature matrix of multi-layer citation network author links through the target loss function to generate citation network generation data, and jointly optimizes the prediction results of the citation network author relationship links through known data and unknown data. The present invention specifically acts on tensor author relationship data and can perform author relationship link prediction that conforms to statistical laws and has high accuracy. It can be widely used in information retrieval, network analysis and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] By referring to the following drawings, the exemplary embodiments of the present invention can be more completely understood. The drawings are used to provide a further understanding of the embodiments of the present application, and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present invention and do not constitute a limitation of the present invention. In the drawings, the same reference numerals generally represent the same parts or steps.

[0056] Figure 1 A flowchart of a method for predicting author collaboration links in a citation network based on multi-layer prediction sampling according to an exemplary embodiment of the present application;

[0057] Figure 2 A schematic diagram of the structure of a citation network author collaboration link prediction device based on multi-layer prediction sampling according to an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0058] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0059] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in this application should have the common meanings understood by technicians in the field to which this application belongs.

[0060] In addition, the terms "first" and "second" etc. are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or devices.

[0061] Example 1

[0062] The present application embodiment provides a citation network author collaboration link prediction method based on multi-layer prediction sampling, comprising the following steps:

[0063] S1: Obtain known data of the original citation network;

[0064] Specifically, data collection may be performed upon receiving an instruction, and the server may send instructions to the citation network by periodically sending instructions, or based on notification from the device, or based on notification from a server, requesting the sending of instructions for known data from the original citation network.

[0065] Specifically, the known data of the citation network is stored in the form of a quadruple, which is represented by W = (a 1 ,a 2 ,rel,l); where a 1 and a 2 They represent different authors in the citation network, and rel represents author a 1 and author a 2 The meta-path relationship between them, l represents the author a 1 and author a 2 The relationship link value of the author collaboration through the meta-path relationship rel. In this embodiment, it is necessary to pre-initialize and set the multi-layer sampling layer number N=20. Then, based on the known data of the original citation network, the citation network author relationship tensor and the latent feature matrix of the multi-layer citation network author link are constructed.

[0066] Specifically, step S2 includes:

[0067] S21: All the quadruples W in the known data of the citation network are 1 ,a 2 ,rel,l) Divide the data into K time periods according to the meta-path relationship rel;

[0068] S22: Quadruple W based on rel=1 (1) =(a 1 ,a 2 ,1,l) to get W (1) Data (a 1 ,a 2 ,l), and according to W (1) Construct a slice matrix T with meta-path relation rel=1 (1) ;

[0069] S23: According to the remaining element path relations rel, based on the above steps, from the four-tuple W=(a 1 ,a 2 ,rel,l) in turn select the four-tuple W (2) , W (3) , …, W (K) , respectively construct the slice matrix T (2) , T (3) ,…,T (K) ;

[0070] S24: Arrange the K slice matrices in order from small to large according to the order of the divided meta-path relationships, and construct the citation network author relationship tensor Y;

[0071] S25: Based on the known data of the citation network, construct the latent feature matrix G, D, Q of the author link of the multi-layer citation network;

[0072] Among them, the slice matrix T (1) The size of the slice matrix T is I rows and J columns, where I and J are the number of authors, K is a positive integer, and (1) The element T in (1) ij represents the relationship link value between the i-th author and the j-th author under the first relationship, 1≤i≤I, 1≤j≤J, Y∈R I×J×K , R represents a set of real numbers; among them, the size of the latent feature matrix G is I rows and R columns, the size of the latent feature matrix D is J rows and R columns, and the size of the latent feature matrix Q is K rows and R columns.

[0073] S3: Based on the citation network author relationship tensor, the target loss function for citation network author relationship prediction is constructed. The target loss function is used to train the latent feature matrix of each layer of citation network author links and iteratively optimize the target loss function.

[0074] Specifically, step S3 includes:

[0075] S31: Initialize the process parameters involved in the prediction of author relationship links in the citation network;

[0076] S32: Construct the target loss function for citation network author relationship link prediction based on the citation network author relationship tensor Y and the latent feature matrices G, D, and Q of the multi-layer citation network author links;

[0077] S33: Use the target loss function to train the latent feature matrix of the author links in each layer of the citation network and iteratively optimize the target loss function;

[0078] S34: Determine whether the target loss function satisfies the iteration termination condition. If so, terminate the iteration and output the target loss function as the target loss function after iterative optimization; if not, continue the iteration.

[0079] In this embodiment, the process parameters include: the citation network author relationship tensor Y, latent feature matrices G, D, and Q, where G is a latent feature matrix with I rows and R columns, D is a latent feature matrix with J rows and R columns, and Q is a latent feature matrix with K rows and R columns; the rank R that determines the dimension of the latent feature space of the latent feature matrices G, D, and Q; the number of layers N; the maximum number of training iterations max Round; the iteration number control variable round during the training process; and the convergence termination threshold τ.

[0080] In this embodiment, the specific operations of initializing each parameter are as follows:

[0081] (1) The rank R determines the dimension of the latent feature space of the latent feature matrices G, D, and Q and is initialized to a positive integer;

[0082] (2) The sizes of the latent feature matrices G, D, and Q are determined by the values ​​of each dimension and the rank R of the corresponding citation network author relationship tensor Y, i.e., G is a latent feature matrix with I rows and R columns, D is a latent feature matrix with J rows and R columns, and Q is a latent feature matrix with K rows and R columns. The three latent feature matrices are initialized with small random positive numbers.

[0083] (3) The maximum number of training iterations max Round is a variable that controls the upper limit of the iteration process and is initialized to a large positive integer; the iteration round control variable round is initialized to 0;

[0084] (4) The convergence termination threshold τ is a parameter used to determine whether the iterative process has converged and is initialized with 1e-5.

[0085] Specifically, step S32 includes:

[0086] S321: Traverse the multi-layer sampling layer number n∈{1,…,N}, for each layer of the constructed citation network author relationship tensor Y and process parameters, use the objective function to train and obtain the latent feature matrix G of each layer of citation network author connection n ,D n ,Q n ;

[0087] Based on the known data set Ψ and the generated data set Ω, the citation network tensor Y is constructed, and the target loss function ε is constructed, which is expressed by the following formula:

[0088]

[0089] Where n represents the number of layers; G n ,D n ,Q n Respectively represent the hidden feature matrix corresponding to the nth layer; represents the single element value corresponding to the hidden feature matrix of the nth layer; α and β represent weights; y ijk represents the known entities in the author relationship tensor Y of the citation network, i.e., the value of the cooperation relationship link between the i-th author and the j-th author on the k-th relationship; It represents the prediction value corresponding to the generated data selected layer by layer and the final prediction value obtained through the non-linear activation function (the citation network generated data unit of the citation network author relationship link prediction module will describe the selection strategy of the generated data in detail, and the citation network author relationship link prediction value calculation unit will introduce the non-linear activation function in detail).

[0090] S322: Iteratively optimize the objective function using the Euclidean distance as the optimization target.

[0091] Specifically, the target loss function uses the Euclidean distance as the optimization target. The first part of the above formula refers to the loss function on the known data set, and the second part refers to the loss function on the generated data set. Then, considering the two cases of the nth layer respectively, the instantaneous loss of a single element is expressed by the following formula:

[0092]

[0093] Among them, α and β represent weights, For unknown authorg i In the corresponding relationship q k Link author d j The author relationship link prediction value is expressed by the following formula:

[0094]

[0095] Specifically, the target loss function ε is iteratively optimized to minimize the value of the target loss function ε. Based on the known data set Ψ of the citation network and the data set Ω generated by the citation network, the SGD algorithm is used to optimize the latent feature matrix G of the nth layer. n ,D n ,Q n The formula for training iterations is as follows:

[0096] (1) Fixed D n and Q n , train G through formula (1) n :

[0097]

[0098] (2) Fixed G n and Q n , train D through formula (2) n :

[0099]

[0100] (3) Fixed G n and D n , train Q by formula (3) n :

[0101]

[0102] Among them, ← is the assignment operation, G n ,D n ,Q n Represent the hidden feature matrix corresponding to the nth layer, Represents the hidden feature matrix G of the nth layer respectively n ,D n ,Q n The corresponding single element, η is the learning rate, n is the number of sampling layers; y ijk represents the relationship link value of the collaboration between the i-th author and the j-th author on the k-th meta-path relationship in the citation network author relationship tensor Y; Represents the final predicted value of the citation network generated data.

[0103] In this embodiment, there are two situations in which the ε training iteration process reaches the termination condition. The first is that in each iteration, the value of the training iteration round control variable round is increased by 1, and when the value of round reaches the maximum training iteration round max Round, the training is stopped; the second is that during the training process, when the difference between the prediction evaluation index value of the previous round after the end of this round of iteration and the prediction evaluation index value calculated after the end of this round of iteration is less than the convergence termination threshold τ=1e-5, the training is stopped.

[0104] S4: Based on the latent feature matrix of the author links in each layer of the citation network after training, the prediction results of the author relationship links in the citation network are obtained.

[0105] Specifically, step S4 includes:

[0106] The K slice matrices into which the citation network author relationship tensor Y is divided are sequentially traversed in each row, a blank position is randomly selected from the known data of the two citation networks, and a prediction of the unknown author relationship is performed to obtain a prediction value, and the prediction value is mapped through a nonlinear activation function to obtain a final prediction value;

[0107] Continuously select new unknown author relationships by slice and row order and repeat the above steps until the preset conditions are met, and all the final prediction values ​​and their unknown author relationships are used to generate citation network data;

[0108] Among them, the relationship connection prediction value is mapped by a nonlinear activation function as follows:

[0109]

[0110] in, is the final prediction value; the preset condition is that the amount of data generated by the citation network is the same as the amount of data known in the citation network.

[0111] Specifically, it is determined whether the number of generated data meets the preset condition (consistent with the number of known data). If not, the generated data is selected and predicted, and then the final predicted value is obtained by mapping through the activation function. Each generated data is added to the known data and stored together in the form of a four-tuple. (Note that the first layer only has known data, no generated data. After the hidden feature matrix is ​​trained based on the known data in the first layer, the generated data is selected, so the first layer generated data set Ω is empty).

[0112] Example 2

[0113] In the above embodiment, a method is provided, and correspondingly, the present application also provides a device. The device provided in the embodiment of the present application can implement the above method, and the device can be implemented by software, hardware, or a combination of software and hardware. For example, the device may include integrated or separate functional modules or units to perform the corresponding steps in the above methods.

[0114] In some implementations of the embodiments of the present application, the device 10 provided in the embodiments of the present application is based on the same inventive concept as the method provided in the aforementioned embodiments of the present application and has the same beneficial effects.

[0115] Please refer to Figure 2 , which shows a schematic diagram of a device provided by some embodiments of the present application. Since the device embodiment is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described below is only illustrative.

[0116] like Figure 2As shown, the apparatus 10 may include:

[0117] A citation network known data acquisition module 110 is used to acquire original citation network known data;

[0118] A citation network author relationship tensor construction module 120 is used to construct a citation network author relationship tensor and a latent feature matrix of multi-layer citation network author links based on the original citation network known data;

[0119] A citation network training module 130 is used to construct a target loss function for citation network author relationship prediction based on the citation network author relationship tensor, use the target loss function to train the latent feature matrix of each layer of citation network author links, and iteratively optimize the target loss function;

[0120] The citation network author relationship link prediction module 140 is used to obtain the prediction result of the citation network author relationship link based on the latent feature matrix of each layer of the citation network author link after training.

[0121] Among them, the citation network training module 130 includes an initialization unit 131 and a training unit 132; the initialization unit 131 is used to initialize the process parameters involved in the citation network author relationship link prediction process; the training unit 132 is used to combine the constructed citation network author relationship tensor and the relevant parameters involved in the citation network author relationship link prediction process initialized by the initialization unit 131, and perform each layer of training to calculate the latent feature matrix of the historical citation network author relationship link.

[0122] Among them, the citation network author relationship link prediction module 140 includes a citation network generation data unit 141, a citation network author relationship link prediction value calculation unit 142, and a citation network generation data storage unit 143. The citation network author relationship link prediction value calculation unit 140 is used to select new citation network generation data according to the latent feature matrix of the historical citation network author relationship link obtained by the training unit 132, according to the generation data unit 141, and use the citation network author relationship link prediction value calculation unit 142 to calculate the citation network author relationship link prediction value, and perform nonlinear activation on the value. Finally, the prediction value and the corresponding author pair are stored in the citation network generation data storage unit 142.

[0123] The iterative optimization process and parameter-related definitions involved in this embodiment have been explained in the previous embodiments and will not be repeated here.

[0124] It should be noted that the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the systems, methods and computer program products according to multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of a code contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of the boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0125] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0126] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0127] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0128] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0129] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application, and they should all be included in the scope of the claims and specification of the present application.

Claims

1. A citation network author collaboration link prediction method based on multi-layer prediction sampling, characterized in that: The steps include: S1: Obtain known data of the original citation network; S2: Based on the known data of the original citation network, construct the citation network author relationship tensor and the latent feature matrix of the multi-layer citation network author links; S3: Based on the citation network author relationship tensor, construct a target loss function for citation network author relationship prediction, use the target loss function to train the latent feature matrix of each layer of citation network author links and iteratively optimize the target loss function; S4: Based on the latent feature matrix of the author links in each layer of the citation network after training, the prediction results of the author relationship links in the citation network are obtained.

2. The method according to claim 1, characterized in that The known data of the citation network is stored in the form of a quaternary group, and the quaternary group is represented by W=(a1, a2, rel, l); Among them, a1 and a2 represent different authors in the citation network respectively, rel represents the meta-path relationship between author a1 and author a2, and l represents the relationship link value of the author cooperation between author a1 and author a2 through the meta-path relationship rel.

3. The method according to claim 2, characterized in that The step S2 comprises: S21: Divide all the four-tuples W=(a1, a2, rel, l) in the known data of the citation network into K time period data according to the meta-path relationship rel; S22: Quadruple W based on rel=1 (1) =(a1,a2,1,l) to get W (1) The data (a1, a2, l) and according to W (1) Construct a slice matrix T with meta-path relation rel=1 (1) ; S23: Select four tuples W from the four tuple W = (a1, a2, rel, l) in turn according to the remaining meta-path relations rel based on the above steps. (2) , W (3) , …, W (K) , respectively construct the slice matrix T (2) , T (3) ,…,T (K) ; S24: Arrange the K slice matrices in order from small to large according to the order of the divided meta-path relationships, and construct a citation network author relationship tensor Y; S25: Based on the known data of the citation network, construct the latent feature matrices G, D, Q of the author links of the multi-layer citation network; Among them, the slice matrix T (1) The size of the slice matrix T is I rows and J columns, where I and J are the number of authors, K is a positive integer, and (1) The element T in (1) ij represents the relationship link value between the i-th author and the j-th author under the first relationship, 1≤i≤I, 1≤j≤J, Y∈R I×J×K , R represents a set of real numbers; among them, the size of the latent feature matrix G is I rows and R columns, the size of the latent feature matrix D is J rows and R columns, and the size of the latent feature matrix Q is K rows and R columns.

4. The method according to claim 3, characterized in that The step S3 comprises: S31: Initialize the process parameters involved in the prediction of author relationship links in the citation network; S32: Construct the target loss function for citation network author relationship link prediction based on the citation network author relationship tensor Y and the latent feature matrices G, D, and Q of the multi-layer citation network author links; S33: using the target loss function to train the latent feature matrix of the author links of each layer of the citation network and iteratively optimize the target loss function; S34: Determine whether the target loss function satisfies the iteration termination condition. If so, terminate the iteration and output the target loss function as the target loss function after iterative optimization; if not, continue the iteration.

5. The method according to claim 4, characterized in that The process parameters include: citation network author relationship tensor Y, latent feature matrices G, D, Q; rank R that determines the dimension of the latent feature space of the latent feature matrices G, D, Q; number of sampling layers N; maximum number of training iterations max Round; training process iteration number control variable round; convergence termination threshold τ.

6. The method according to claim 5, characterized in that Step S32 includes: S321: traverse the multi-layer sampling layer number n∈{1,…,N}, for each layer of the constructed citation network author relationship tensor Y and the process parameters, use the target loss function to train and obtain the latent feature matrix G of each layer of the citation network author connection n ,D n ,Q n ; S322: Iteratively optimize the target loss function using the Euclidean distance as the optimization target.

7. The method according to claim 6, characterized in that Step S322 is to use the SGD algorithm to calculate the hidden feature matrix G of the nth layer. n ,D n ,Q n Training iterations, including: (1) Fixed D n and Q n , train G through formula (1) n : (2) Fixed G n and Q n , train D through formula (2) n : (3) Fixed G n and D n , train Q by formula (3) n : Among them, ← is the assignment operation, G n ,D n ,Q n Represent the hidden feature matrix corresponding to the nth layer, Represents the hidden feature matrix G of the nth layer respectively n ,D n ,Q n The corresponding single element, η is the learning rate, n is the number of sampling layers; y ijk represents the relationship link value of the collaboration between the i-th author and the j-th author on the k-th meta-path relationship in the citation network author relationship tensor Y; represents the final predicted value of the citation network generated data; Where ε is the target loss function, which is expressed by the following formula (4): Among them, the target loss function ε is the instantaneous loss of a single element in the case of known data in the citation network and data generated by the citation network at the nth layer by the following formula: Among them, α and β represent weights, For unknown authorg i In the corresponding meta-path relation q k Link author d j The relationship link prediction value is expressed by the following formula:

8. The method according to claim 7, characterized in that The iteration termination condition of step S34 is that the iteration round control variable round reaches the maximum training iteration round max Round or the difference between the prediction result of the citation network in the previous round and the prediction result of the citation network after the current round is less than the convergence termination threshold τ.

9. The method according to claim 8, characterized in that The step S4 comprises: The K slice matrices into which the citation network author relationship tensor Y is divided are sequentially traversed through each row, a blank position is randomly selected from the known data of the two citation networks, unknown author relationships are predicted and relationship connection prediction values ​​are obtained, and then the relationship link prediction values ​​are mapped through a nonlinear activation function to obtain a final prediction value; Continuously select new unknown author relationships by slice and row order and repeat the above steps until the preset conditions are met, and all the final prediction values ​​and their unknown author relationships are used to generate citation network data; Among them, the relationship connection prediction value is mapped by a nonlinear activation function as follows: in, is the final prediction value; the preset condition is that the amount of data generated by the citation network is the same as the amount of data known in the citation network.

10. A citation network author collaboration link prediction device based on multi-layer prediction sampling, characterized in that: include: A citation network known data acquisition module is used to acquire original citation network known data; The citation network author relationship tensor construction module is used to construct the citation network author relationship tensor and the latent feature matrix of multi-layer citation network author links based on the known data of the original citation network; A citation network training module, for constructing a target loss function for citation network author relationship prediction based on the citation network author relationship tensor, using the target loss function to train the latent feature matrix of each layer of citation network author links and iteratively optimizing the target loss function; The citation network author relationship link prediction module is used to obtain the prediction results of the citation network author relationship link based on the latent feature matrix of each layer of the citation network author link after training.

Images