A method for graph network data encoding based on mixed-order anonymous random walks
Through the graph network data encoding method based on hybrid-order anonymous random walk, the problem of insufficient data mining capabilities of graph network data in the existing technology is solved, and the hierarchical representation and information capture capabilities of graph structure are improved, which is improved. Graph the classification accuracy of network data.
Patent Information
- Application Number
- CN202110406833.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-04-15
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2041-04-15
AI Technical Summary
The prior art is difficult to effectively mine the information and connections contained in graph network data, especially lacking the ability to hierarchically represent the graph structure and accurately capture graph network data information.
Using a graph network data encoding method based on hybrid-order anonymous random walk, the graph network data set is constructed, and the mixed-order anonymous random walk structure is captured, and the encoding under different anonymous random walks is calculated using natural language processing technology, and the encoding under different anonymous random walks is calculated through The softmax function and gradient descent method optimize weights to realize mixed encoding of graph network data.
The hierarchical representation of the graph structure is realized and the ability to capture graph network data information is improved, and the classification accuracy of graph network data is improved.
Smart Images

Figure CN113239665B_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of graph network data analysis and mining, and in particular to a method for graph network data encoding based on mixed-order anonymous random walks. Background Art
[0002] In recent years, the methods in the field of graph network data analysis and mining mainly include neural network encoding methods and kernel function encoding methods. Among them, neural network encoding methods, including but not limited to graph neural networks (GNNs), graph convolutional neural networks (GCNs, PSCNs, etc.), aim to (convolute) encode large-scale graph network data. The disadvantage is that a large amount of data is required, and the final encoding result is usually obtained through iterative calculations, so that the refined structure of the graph network data is coupled and lacks accurate representation of the refined structure. The kernel function encoding method obtains the encoding of the graph data network by directly calculating on the graph network data. The kernel function is usually divided into two categories: based on walks and based on finite subgraph structures. The disadvantage of the method based on finite subgraph structure is that the calculation of the subgraph structure (finite subgraph) needs to consider the isomorphism problem of the subgraph, which is an NP-complete problem and is unrealistic to calculate in finite polynomial time; the method based on walks is often limited to a fixed walk length (i.e., the same order), so the walk structure obtained is also of fixed size, which lacks a hierarchical representation of the graph structure.
[0003] With the advancement of data mining technology, graph network data, as a structured data representation form, has shown powerful information representation capabilities. How to fully mine the information and connections contained in graph data networks has become a hot issue in the field of data mining today. Graph coding technology maps graph network data to vector space and represents graph network data with vectors of finite dimensions. This normalizes graph network data in an irregular form, which is more conducive to data mining tasks such as classification and clustering. Summary of the invention
[0004] The purpose of the present invention is to provide a method for encoding graph network data based on mixed-order anonymous random walks, which is used to solve the problem of sufficient mining of information and connections contained in graph network data in the prior art, obtain a hierarchical representation of the graph structure and improve the ability to capture graph network data information.
[0005] Given a series of graph network data G = (G1,...,G N ), graph encoding is done through the encoding function G→R 1×d Calculate the encoding matrix X N×d , where the i-th column represents the graph G i The d-dimensional encoding vector in the encoding space.
[0006] In the first aspect, a method for encoding graph network data based on mixed-order anonymous random walk is provided, and a graph network data set G = (G1, ..., G N ), the method comprising:
[0007] Traversing the random walk length set L = (l1,...,l T ), capturing mixed-order anonymous random walk structures of graph network data;
[0008] According to the positions where different points in the random walk first appear in the random walk, the random walk is converted into an anonymous random walk, so as to mine anonymous random walk structures of different orders in graph network data.
[0009] Using natural language processing techniques, we calculate different encodings of graph network data under anonymous random walks of different orders. We nonlinearly combine different encodings under anonymous random walks of different orders to obtain a mixed encoding of graph network data, where each weight is represented by a softmax function with parameters.
[0010] Based on the graph network data classification task, the gradient descent method is used to calculate the weights assigned in the mixed encoding under different graph network data to achieve the optimal classification result.
[0011] In one embodiment, the mixed-order anonymous random walk structure of the captured graph network data is to start from each point and perform a random walk to obtain random walk data of different orders, wherein the order represents the path length from the path point in the random walk to the initial point.
[0012] In one embodiment, in a single graph G, a random walk w is defined as a finite sequence of points (v0, v1, v2, ..., v l ), where v0 is the initial point, v i+1 represents the point selected from its neighboring points after the i-th point is determined; random walk w=(v0,v1,...,v l ), the Position function f is defined as f(v i )=|(v0,...,v i′ )|, where i′ is a i′ =v i The smallest number; for a random walk w=(v0,v1,...,v l ), the corresponding anonymous random walk is defined as a=(f(v0),f(v1),...,f(v l )), where f is the Position function.
[0013] In one embodiment, for a graph network dataset G=(G1,...,GN ) is used to mine anonymous random walks with a step length of l for all graph network data in the dataset. Let the set of anonymous random walks with a step length of l be Traverse the random walk step set L = (l1,...,l T ), get all anonymous random walks of different step lengths
[0014] In one embodiment, the graph network G i ∈G=(G1,G2,...,G N ) is encoded under an anonymous random walk of length l; in the graph network dataset G i ∈G=(G1,G2,...,G N ), i=1,2,...,N, assuming that the set of all anonymous random walks with a step length of l mined is Apply the document encoding method PV-DBOW in natural language processing to learn each graph G i ∈G=(G1,G2,...,G N ) is encoded in an anonymous random walk with a step size of l as a vector The corresponding anonymous random walk matrix in It has 1×d dimensions, where d is the number of pixels per graph G. i The encoding dimension of ; anonymous random walk matrix With |A l |×d dimension, |A l | is the number of all sampled anonymous random walks of length l, d is the encoding dimension of each anonymous random walk; use the matrix M l Represents the encoded representation of all graph network data under anonymous random walk with step length l, where each line is a graph G i Encoded representation of anonymous random walk with step size l Right now:
[0015]
[0016] In one implementation, the PV-DBOW method is used to convert the graph network dataset G = (G1, ..., G N ) as a document collection, and each graph network data G i As a document, G=(G1,...,G N Each anonymous random walk structure with a step length l mined in G is considered to be contained in G i words in;
[0017] For the mined graph G i An anonymous random walk of length l in Optimize the following objective function, that is, make the random walk Appears in Figure G i The highest probability:
[0018]
[0019] Probability Defined as a vector using the softmax function With G i The dot product of With G i They are With G i The encoding vector of :
[0020]
[0021] Where |A l | represents the number of anonymous random walks with a step length of l; by optimizing the objective function, the matrix M is obtained l ;
[0022] Similarly, traverse all the step lengths L = (l1,...,l T ), we get that for all step sizes L = (l1,...,l T ) is the encoding representation matrix of all graph network data under
[0023] In one embodiment, Represents the mixed encoding matrix of the graph network dataset, where α i With parameters θ=(θ1,...,θ T )’s softmax function definition Represents the encoding matrix The weights in the mixed coding matrix; then
[0024]
[0025] The i-th behavior of M
[0026]
[0027] Representation graph G i Encoding of mixed-order anonymous random walks.
[0028] In one embodiment, the i-th row of M represents the hybrid coding applied to the graph classification task, and the loss function is defined. The optimal parameter θ of the hybrid coding matrix is calculated using the gradient descent method. T ) solution;
[0029] Assume that the classifier is represented by a function g(x), whose input is the encoding of each graph; the classifier includes SVM and Logistic Regression; if the classifier is for graph G i If the classification is correct, then let g(G i )=-1; otherwise let g(G i )=1; then the loss function is defined as
[0030]
[0031] Use gradient descent to update the parameters θ i , i = 1, ..., N, until convergence is reached; finally, the parameters of the hybrid coding matrix θ = (θ1, ..., θ T ) and weights α1,...,α T , to achieve the optimal nonlinear composition of hybrid coding.
[0032] In a second aspect, the present invention provides a graph network data encoding device based on mixed-order anonymous random walk, comprising:
[0033] The capture unit is used to traverse the random walk length set L = (l1,...,l T ), capturing mixed-order anonymous random walk structures of graph network data;
[0034] A conversion unit is used to convert the random walk into an anonymous random walk according to the positions where different points in the random walk first appear in the random walk, so as to realize the mining of anonymous random walk structures of different orders of graph network data;
[0035] The encoding unit is used to calculate different encodings of the graph network data under anonymous random walks of different orders by using the technical means of natural language processing; the different encodings under anonymous random walks of different orders are nonlinearly combined to obtain a mixed encoding of the graph network data;
[0036] The optimization unit is used to calculate the weights allocated in the mixed encoding under different graph network data based on the graph network data classification task, using the gradient descent method to obtain the optimal classification result.
[0037] In one embodiment, the mixed-order anonymous random walk structure of the crawling unit crawling graph network data is obtained by selecting a step length l∈L=(l1,...,l T ), for G i ∈G=(G1,G2,...,G N ), starting from each point on i=1,2,...,N, a random walk with a step length of l is performed;
[0038] The conversion unit converts all random walks into anonymous random walks. Let the set of anonymous random walks with a step length of l be Complete the mining of all anonymous random walks with a step length of l on the graph network dataset, and then traverse the random walk step length set L = (l1,...,l T ), get all anonymous random walks of different step lengths
[0039] The encoding unit uses the PV-DBOW algorithm of natural language processing to encode the mined random walks and graph data networks with a step length of l, and obtains the encoding representation matrix M of all graph network data under the anonymous random walk with a step length of l. l Similarly, we can obtain the encoding matrix of all graph network data under different random walk step length sets
[0040] Using the parameter θ=(θ1,...,θ T )’s softmax function represents the weight α i Hybrid encoding matrix for graph network dataset
[0041] The optimization unit constructs a loss function based on the graph classification task and uses the gradient descent method to calculate the parameter θ=(θ1,...,θ T ) and weight α i .
[0042] The present invention pre-sets the walking step length set L of the mixed-order random walk = (l1, ..., l T ), give the definition of random walk Position function and anonymous random walk; then mine anonymous random walks by selecting step size l∈L=(l1,...,l T ), for G i ∈G=(G1,G2,...,G N ), starting from each point on i=1,2,...,N, we mine random walks with a step length of l, and transform all random walks into anonymous random walks. Let the set of anonymous random walks with a step length of l mined be In this way, the mining of all anonymous random walks with a step length of l on the graph network dataset is completed. According to this method, the random walk step length set L = (l1,...,l T ), get all anonymous random walks of different step lengths Then, the PV-DBOW algorithm of natural language processing is used to encode the mined random walks and graph data networks with a step length of l, so that the encoding representation matrix M of all graph network data under the anonymous random walk with a step length of l is obtained. lAccording to this method, the encoding representation matrix of all graph network data under different random walk step length sets is obtained. Using the parameter θ=(θ1,...,θ T )’s softmax function represents the weight α i Hybrid encoding matrix for graph network dataset Based on the graph classification task, a loss function is constructed, and the gradient descent method is used to calculate the parameter θ=(θ1,...,θ T ) and weight α i The present invention can be used to obtain a hierarchical representation of a graph structure and improve the ability to capture graph network data information. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] The above and / or additional aspects and advantages of the present disclosure will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:
[0044] Figure 1 is a flow chart of the method and apparatus according to the present disclosure. DETAILED DESCRIPTION
[0045] The English term for order in the present invention is order; order is sometimes also represented by jump, and the English term is hop.
[0046] The PV-DBOW algorithm for natural language processing in the present invention is called Paragraph Vector-Distributed Bag-of-Words in English.
[0047] The optimization calculation idea adopted by the present invention is to use word2vec and doc2vec in the field of natural language processing (NLP) to determine the relationship between the number of anonymous random walks and the number of random walk sampling estimates in the field of graph data mining. Word2vec is used to encode word vectors, while doc2vec is used to encode sentences and paragraphs.
[0048] In word2vec, if two words are close in position in a sentence, then after encoding, their positions in the encoding space are also close. Given a vocabulary set V = {w1,w2,...,w |V|}, target vocabulary w t and a series of neighboring words w around it in a sentence t-c ,...,w t ,...,w t+c , word2vec uses Skipgram to optimize the probability of neighboring words appearing around the target word:
[0049]
[0050] The conditional probability Pr(w t-c ,...,w t ,...,w t+c |w t ) uses the independence assumption to calculate:
[0051]
[0052] Furthermore, the conditional probability Pr(w j |w t )Use the softmax function to represent
[0053]
[0054] where w t , w j and w i Respectively represent w t , w j and w i The encoding vector of . In summary, (1) is expressed as
[0055]
[0056] In doc2vec, given a text set D = {D1, D2, ..., D N}, where all the words are a set V = {w1,w2,...,w |V|}, the target text is D t , which contains the vocabulary Doc2vec uses PV-DBOW to optimize the objective by maximizing Contained in text D t Probability Let's learn text D t The encoding, conditional probability is defined as follows:
[0057]
[0058] Where D t , and w i Respectively represent w t , and w i The encoding vector of .
[0059] The relationship between the number of anonymous random walks and the number of random walk sampling estimates proposed by Shervashidze et al. in their paper “Efficient graphlet kernels for large graph comparison” published at the 12th International Conference on Artificial Intelligence and Statistics is as follows:
[0060] For a graph G∈G, when it is large enough, the computational complexity required to sample ξ random walks of step length l is O(ξl). The relationship between the number of samples ξ and the number of anonymous walks μ in O(ξl) is determined by the artificially set ε and δ:
[0061]
[0062] Where ε>0, 0≤δ≤1.
[0063] Embodiments of the present disclosure are described in detail below, examples of which are shown in the accompanying drawings, and the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present disclosure, and should not be construed as limiting the present disclosure.
[0064] (1) In step S101, a set of mixed-order random walk step lengths L = (l1, ..., l T ). Before a detailed description is given, the present disclosure predefines a random walk, a Position function, and an anonymous random walk.
[0065] Definition of random walk: In graph G, a random walk w is defined as a finite sequence of points (v0,v1,v2,...,v l ), where v0 is the initial point, v i+1 Indicates v i A point is selected from its neighboring points after the point is determined.
[0066] Define the Position function: In the random walk w=(v0,v1,...,v l ), the Position function f is defined as f(v i )=|(v0,...,v i′ )|, where i′ is a i′ =v i The smallest number of .
[0067] Define anonymous random walk: For a random walk w = (v0, v1, ..., vl ), the corresponding anonymous random walk is defined as a=(f(v0),f(v1),...,f(v l )), where f is the Position function.
[0068] The advantage of random walks is that they can fully mine graph information, including points and edges. However, because the mined information is too detailed, especially in large graph data networks, the amount of information is too large, the distribution of random walk categories is too sparse, and the computational requirements are too high, which is not conducive to fast encoding of graph data networks. Although anonymous random walks give up some information, they retain the structural information of the graph to a large extent, greatly reducing the amount of computation. Existing theories have proved that the graph data network information mined by anonymous walks can reconstruct the original graph.
[0069] (2) In step S102, for the graph network dataset G = (G1, ..., G N ), for graph G i ∈G=(G1,G2,...,G N ) to perform anonymous random walk mining, in Figure G i Mining random walks with number ξ The length of each random walk is l, so its corresponding anonymous random walk is A Gi ={a1,a2,...,a μ}.
[0070] For G=(G1,...,G N ) are mined by anonymous random walks with a step length of l. Let the set of anonymous random walks with a step length of l be
[0071] According to this method, the random walk step set L = (l1,...,l T ), get all anonymous random walks of different step lengths
[0072] (3) In step S103, the graph data network G is now calculated i ∈G=(G1,G2,...,G N ) under the encoding of an anonymous random walk of length l.
[0073] Graph Network G i ∈G=(G1,G2,...,G N ) is encoded under an anonymous random walk of length l. In the graph network dataset G i ∈G=(G1,G2,...,G N), i=1,2,...,N, assuming that the set of all anonymous random walks with a step length of l mined is This paper applies a document encoding method PV-DBOW (Paragraph Vector-Distributed Bag-of-Words) in natural language processing to learn each graph G i ∈G=(G1,G2,...,G N ) is encoded in an anonymous random walk with a step size of l as a vector The corresponding anonymous random walk matrix in It has 1×d dimensions, where d is the number of pixels per graph G. i The encoding dimension of ; anonymous random walk matrix With |A l |×d dimension, |A l | is the number of all sampled anonymous random walks of length l, and d is the encoding dimension of each anonymous random walk. l Represents the encoded representation of all graph network data under anonymous random walk with step length l, where each line is a graph G i The encoding representation of the anonymous random walk with a step length of l Right now:
[0074]
[0075] PV-DBOW transforms the graph network dataset G = (G1,...,G N ) as a document collection, and each graph network data G i As a document, G=(G1,...,G N Each anonymous random walk structure with a step length l mined in G is considered to be contained in G i The words in .
[0076] Assume that all mined graphs G i The anonymous random walk of length l in is The present disclosure optimizes the following objective function, that is, the random walk Appears in Figure G i The highest probability of occurrence:
[0077]
[0078] Probability Defined as a vector using the softmax function With G i The dot product of With G i They are With Gi The encoding vector of :
[0079]
[0080] Where |A l | represents the number of anonymous random walks with a step length of l. By optimizing the objective function, we get the matrix M l .
[0081] Similarly, using the above method, traverse all step lengths L = (l1,...,l T ), we get that for all step sizes L = (l1,...,l T ) is the encoding representation matrix of all graph network data under
[0082] (4) In step S104, Represents the mixed encoding matrix of the graph network dataset, where α i With parameters θ=(θ1,...,θ T )’s softmax function definition Represents the encoding matrix The weights in the hybrid coding matrix.
[0083]
[0084] The i-th behavior of M
[0085]
[0086] Representation graph G i Encoding of mixed-order anonymous random walks.
[0087] (5) In step S105, the hybrid coding obtained in step S104 is applied to the graph classification task, a loss function is defined, and the optimal parameters of the hybrid coding matrix θ = (θ1, ..., θ T )’s solution.
[0088] Assume that the classifier is g(x), and the input is the encoding of each graph. i If the classification is correct, then let g(G i )=-1; otherwise let g(G i )=1. Then the loss function is defined as
[0089]
[0090] (1) Use gradient descent to update the parameter θ i, i = 1, ..., N, until convergence is reached. Finally, the parameter θ = (θ1, ..., θ T ) and weight α i .
Claims
1. A method for encoding graph network data based on mixed-order anonymous random walks, based on the acquired graph network data, constructs a graph network dataset G = (G 1 ,...,G N ), It is characterized in that The method comprises: Traverse the random walk step set L = (l 1 ,...,l T ), capturing mixed-order anonymous random walk structures of graph network data; According to the positions where different points in the random walk first appear in the random walk, the random walk is converted into an anonymous random walk, so as to mine anonymous random walk structures of different orders in graph network data. Using natural language processing technology, we calculate different encodings of graph network data under anonymous random walks of different orders; we nonlinearly combine different encodings under anonymous random walks of different orders to obtain hybrid encodings of graph network data; Based on the graph network data classification task, the gradient descent method is used to calculate the weights allocated in the hybrid encoding under different graph network data to achieve the optimal classification result; The natural language processing technique is used to calculate the different encodings of the graph network data under different orders of anonymous random walks. Specifically, the PV-DBOW method is used to convert the graph network data set G = (G 1 ,...,G N ) as a document collection, and each graph network data G i As a document, G=(G 1 ,...,G N ) is considered as the anonymous random walk structure of each step length l mined in G i words in; For the mined graph G i An anonymous random walk with a step size of l in Optimize the following objective function, that is, make the random walk Appears in Figure G i The highest probability: Probability Defined as a vector using the softmax function With G i The dot product of With G i They are With G i The encoding vector of : Where |A l | represents the number of anonymous random walks with a step length of l; by optimizing the objective function, we get the matrix M l ; Similarly, traverse all the step lengths L = (l 1 ,...,l T ), we get that in all step lengths L = (l 1 ,...,l T ) is the encoding representation matrix of all graph network data under 2. The method according to claim 1, Features: The mixed-order anonymous random walk structure of the captured graph network data, that is, starting from each point, performing a random walk to obtain random walk data of different orders, wherein the order represents the path length from the path point in the random walk to the initial point.
3. The method according to claim 1, Features: In a single graph G, a random walk w is defined as a finite sequence of points (v 0 ,v 1 ,v 2 ,...,v l ), where v 0 is the initial point, v i+1 represents the point selected from its neighboring points after the i-th point is determined; random walk w = (v 0 ,v 1 ,...,v l ), the Position function f is defined as f(v i )=|(v 0 ,...,v i′ )|, where i′ is a i′ =v i The smallest number; for random walk w=(v 0 ,v 1 ,...,v l ), the corresponding anonymous random walk is defined as a=(f(v 0 ),f(v 1 ),...,f(v l )), where f is the Position function.
4. The method according to claim 1, Features: For the graph network dataset G=(G 1 ,...,G N ) is used to mine anonymous random walks with a step length of l for all graph network data in the dataset. Let the set of anonymous random walks with a step length of l be Traverse the random walk step set L = (l 1 ,...,l T ), get all anonymous random walks of different step lengths 5. The method according to claim 1, Features: Graph Network G i ∈G=(G 1 ,G 2 ,...,G N ) is encoded under an anonymous random walk with a step size of l; in the graph network dataset G i ∈G=(G 1 ,G 2 ,...,G N ), i=1,2,...,N, assuming that the set of all anonymous random walks with a step length of l mined is Apply the document encoding method PV-DBOW in natural language processing to learn each graph G i ∈G=(G 1 ,G 2 ,...,G N ) is encoded in an anonymous random walk with a step size of l as a vector The corresponding anonymous random walk matrix in It has 1×d dimensions, where d is the number of pixels per graph G. i The encoding dimension of ; anonymous random walk matrix With |A l |×d dimension, |A l | is the number of anonymous random walks with a step length of l obtained by all sampling; the matrix M l Represents the encoded representation of all graph network data under anonymous random walk with step length l, where each line is a graph G i Encoded representation of anonymous random walk with step size l Right now:
6. The method according to claim 1, Features: use Represents the mixed encoding matrix of the graph network dataset, where α t With the parameter θ=(θ 1 ,...,θ T )’s softmax function definition Represents the encoding matrix The weights in the mixed coding matrix; then The i-th behavior of M Representation graph G i Encoding of mixed-order anonymous random walks.
7. The method according to claim 6, Features: The i-th row of M represents the hybrid coding and is applied to the graph classification task. The loss function is defined and the gradient descent method is used to find the optimal parameter θ = (θ 1 ,...,θ T ) solution; Assume that the classifier is represented by a function g(x), whose input is the encoding of each graph; the classifier includes SVM and LogisticRegression; if the classifier is for graph G i If the classification is correct, then let g(G i )=-1; otherwise let g(G i )=1; then the loss function is defined as Use gradient descent to update the parameters θ k , k = 1, ..., T, until convergence is reached; finally, the parameter θ of the hybrid coding matrix that achieves the optimal classification result under the classification task is obtained. 1 ,...,θ T ) and weight α 1 ,...,α T , to achieve the optimal nonlinear composition of hybrid coding.
8. A graph network data encoding device based on mixed-order anonymous random walks for executing the method of claim 1, It is characterized in that include: The capture unit is used to traverse the random walk step set L = (l 1 ,...,l T ), capturing mixed-order anonymous random walk structures of graph network data; A conversion unit is used to convert the random walk into an anonymous random walk according to the positions where different points in the random walk first appear in the random walk, so as to realize the mining of anonymous random walk structures of different orders of graph network data; The encoding unit is used to use natural language processing technology to calculate different encodings of graph network data under different orders of anonymous random walks; The different codes under different orders of anonymous random walks are nonlinearly combined to obtain the hybrid code of graph network data; The optimization unit is used to calculate the weights allocated in the mixed encoding under different graph network data based on the graph network data classification task, using the gradient descent method to obtain the optimal classification result.
9. The device according to claim 8, It is characterized in that The mixed-order anonymous random walk structure of the crawling unit crawling graph network data is achieved by selecting the step size l∈L=(l 1 ,...,l T ), for G i ∈G=(G 1 ,G 2 ,...,G N ), starting from each point on i=1,2,...,N, a random walk with a step length of l is performed; The conversion unit converts all random walks into anonymous random walks. Let the set of anonymous random walks with a step length of l be Complete the mining of all anonymous random walks with a step length of l on the graph network dataset, and then traverse the random walk step length set L = (l 1 ,...,l T ), get all anonymous random walks of different step lengths The encoding unit uses the PV-DBOW algorithm of natural language processing to encode the mined random walks and graph data networks with a step length of l, and obtains the encoding representation matrix M of all graph network data under the anonymous random walk with a step length of l. l Similarly, we can obtain the encoding matrix of all graph network data under all random walk step length sets Using the parameter θ=(θ 1 ,...,θ T )’s softmax function represents the weight α t Hybrid encoding matrix for graph network dataset The optimization unit constructs a loss function based on the graph classification task and uses the gradient descent method to calculate the parameter θ=(θ 1 ,...,θ T ) and weight α t .
