A high-dimensional rumor text causality discovery method based on a graph selection attention mechanism

Abstract

The application discloses a high-dimensional rumor text causality discovery method based on a graph selection attention mechanism, and belongs to the field of natural language processing. In view of the problem that existing rumor detection ignores the causality correlation between event entities, rumor text data is taken as a research object, a HSIC Lasso feature selection method is adopted to restore an initial graph skeleton in view of the high-dimensional variable problem of rumor text, the graph search process of a rumor causality model is optimized, and a graph selection attention mechanism is introduced into a rumor causality discovery process in view of the problem that semantic information between rumor causality variables is ignored. Experiment proves that the rumor causality discovery precision in a high-dimensional variable scene can be further improved, and the method has more outstanding performance on rumor data containing semantic information. The rumor causality relationship discovered on rumor text can realize timely judgment of rumors through causality knowledge.

Description

TECHNICAL FIELD

[0001] The present application relates to a high-dimensional rumor text causal discovery method based on a graph selection attention mechanism, and belongs to the field of natural language processing. By adopting the HSIC Lasso feature selection method to alleviate the accuracy problem of high-dimensional causal discovery, the graph selection attention mechanism is introduced to alleviate the problem that the existing causal discovery method does not pay attention to semantic information, which is helpful to timely distinguish rumors through causal knowledge. BACKGROUND

[0002] The spread of rumors can cause many harms to society, such as causing huge losses to product producers, causing public panic, and endangering the national image. Therefore, it is crucial to timely distinguish rumors and reduce the spread of rumors. However, network rumors involve many aspects such as society, economy, politics, medicine, food, etc., and are usually difficult to distinguish. At the same time, with the growth of information and data, the study of causal relationships is becoming more and more extensive in the process of massive data analysis.

[0003] In recent years, causal discovery has received more and more attention in the field of artificial intelligence and machine learning, and has played an important role in the fields of neuroscience, economics and finance, psychology, genomics, etc. People often infer the causality between things based on existing knowledge, such as staying up late may cause headaches, and using prior knowledge to discover the causal relationship between events to detect rumors. From the perspective of causal relationship, the analysis of rumors has high research value and practical significance. Existing rumor detection mostly focuses on the spread pattern, and a small part of research work focuses on the authenticity of rumor content, but all ignore the causal relationship between event entities, which requires the help of causal discovery method to discover the causal relationship in rumor text, and promotes the discrimination of rumors from the perspective of causality.

[0004] The rumor causal discovery problem mainly has the following difficulties:

[0005] (1) Internet rumors spread rapidly and involve a wide range of fields, which leads to a large number of entities in rumor texts, and the causal relationship between entities is complex. The number of directed acyclic graphs on multiple causal variables reaches an exponential level, and traditional causal discovery methods are not accurate enough in the high-dimensional variable scenario of rumor text.

[0006] (2) Unlike structured attribute variables (height, weight, etc.), rumor causal variables contain a lot of semantic information, but traditional causal discovery methods do not fully utilize semantic information. SUMMARY

[0007] The present application aims at the problem of the traditional causal discovery method in the high-dimensional variable scene of rumor text processing, which is not accurate and does not process semantic information, and provides a high-dimensional rumor text causal discovery method based on graph selection attention mechanism.

[0008] To achieve the above-mentioned purpose, the present application comprises the following steps:

[0009] a. Process data to obtain a word list of causal variable words, and perform vectorization representation on each text;

[0010] b. The processed causal variables in the text sequence are input into the HSIC Lasso feature selection layer to obtain an initial causal graph;

[0011] c. The node vector set of the initial graph after feature selection is input into the graph attention layer, and the graph attention layer performs representation learning on the initial graph to obtain semantic representation of the initial graph;

[0012] d. The causal variables of the rumor and the semantic representation of the initial graph are input into the rumor causal generation network to obtain the final causal graph.

[0013] In step a, the specific steps of vectorizing each text can be: first, using a Chinese word segmentation tool to segment the text content, and then filtering the stop words and some meaningless words to obtain the word as the word list of the causal variable words, and performing one-hot vectorization on each text according to the word list of the causal variable words.

[0014] In step b, the processed causal variables in the text sequence are input into the HSIC Lasso feature selection layer, and based on the HSIC Lasso feature selection method, the non-redundant features with strong statistical dependence on the output value are found through the input features.

[0015] In step c, the input of the graph attention layer is the node vector set in the initial graph after feature selection, and the output is a new node feature set; the graph attention layer is a single-layer forward neural network, and LeakyReLU is used as the activation function; after obtaining the attention weight, the attention weight is used to calculate the linear combination of the corresponding features, and the final features are calculated by using the sigmoid function.

[0016] In step d, the rumor causal generation network receives the graph representation of the attention layer output, learns the rumor variable node representation using the graph selection attention mechanism, learns the rumor variable generation model through back propagation, trains and optimizes the model structure and network weight using the score-based method, and finds the optimal rumor causal graph using greedy search.

[0017] Compared with the prior art, the present application has the following advantages and effects:

[0018] (1) As can be seen from the example analysis, the rumor causal discovery method for text rumor data proposed in the present application can pay more attention to the causal relationship in the rumor, which helps to use causal knowledge to timely distinguish rumors.

[0019] (2) As can be seen from the example analysis, the high-dimensional rumor causal graph search method based on HSIC Lasso feature selection proposed in the present application can further improve the rumor causal discovery accuracy in the high-dimensional causal variable scenario by introducing the HSIC Lasso feature selection method to restore the initial graph skeleton and optimize the graph search process of the rumor causal model.

[0020] (3) As can be seen from the example analysis, the rumor causal variable semantic representation method based on the graph selection attention mechanism proposed in the present application can better focus on the semantic information between rumor causal variables by introducing the graph selection attention mechanism in the rumor causal discovery process, and has more outstanding performance on rumor data containing semantic information. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 It is a general structure diagram of the present application. The left half is the process of generating an initial graph using a text sequence, and the right half is the part of generating a causal graph using causal variables and an initial graph.

[0022] Figure 2 It is a graph attention layer module structure of the present application. The left half of the figure shows the main calculation process of the graph selection attention mechanism, and the right half of the figure shows the process of obtaining rumor causal variable representation using the multi-head graph selection attention mechanism. DETAILED DESCRIPTION

[0023] The embodiment of the present application includes the following steps:

[0024] Step 1: The processed causal variables in the text sequence are generated into an initial graph by the HSIC Lasso feature selection method;

[0025] Step 2: The graph attention layer learns the representation of the initial graph;

[0026] Step 3: The rumor causal generation network generates the final causal graph structure.

[0027] Further, the specific content of the step 1 is:

[0028] An alternative implementation of HSIC Lasso using feature-based nonlinear Lasso is used. Given input features {u1, u2,..., u k}, assuming that the output {y1, y2,..., y n} is known, the output here represents the result of causal discovery, and the output of the feature algorithm represents a feature subset, i.e., a subset of rumor causal variables in the rumor causal discovery task. The objective function in the HSIC Lasso algorithm introduces the Frobenius norm and adds regularization, as follows:

[0029]

[0030] s.t.α1,...,α d ≥0

[0031] By minimizing the objective function in the HSIC Lasso algorithm, the purpose of approximating the linear combination of the input kernel matrix to the output and matrix is achieved, and the input feature combination obtained in this process is considered as a feature that has strong statistical dependence on the output value, which is taken as the final result of the feature selection process. Here, K and L can be different, and a non-negative constraint is added, the first term of the objective function represents the output kernel matrix obtained by linear combination of the feature-based input kernel matrix

[0032] The first term in the above objective function formula can be expanded as follows:

[0033]

[0034]

[0035] Where HSIC represents a kernel-based independence measure, also known as the Hilbert-Schmidt independence criterion, the value of HSIC is limited to be non-negative, when using a Gaussian kernel similar to a generalized reproducing kernel, the value of HSIC is 0 when two random variables are independent of each other, if the kth feature has a high dependence on the output, HSIC will take a large value, and therefore α k will also get a larger value. When u k is independent of the output y, the value of HSIC(u k , y) will tend to 0, and α k will be basically eliminated in the L1 regularization process, which means that HSIC Lasso will select features with strong dependence.

[0036] When u k and u l have strong dependence, the value of HSIC(u k , u l ) will be large, at this time one of a k and a l will be set to 0, which means that HSIC Lasso will eliminate redundant features in the process of feature selection, and it tends to find the optimal solution with less redundancy in features.

[0037] The specific steps of step 2 are as follows:

[0038] The input of the graph attention layer is the node vector set of the initial graph after feature selection Use h to represent the node instead of the initial x, the initial graph node after feature selection, N is the number of variables at this time, that is, the number of initial graph nodes obtained after HSIC Lasso feature selection, F represents the number of features of each node, and the output is a new node feature set In order to obtain an output feature with sufficient expression, a learnable linear transformation is performed on the input feature, and a weight matrix W ∈ R F’×F is initialized for each node. For each node, use the self-graph selection attention mechanism to calculate the attention coefficient. The formula of the attention coefficient is as follows:

[0039]

[0040] This attention coefficient represents the importance of the features of node j to node i. In the general graph selection attention mechanism, each node is involved in the attention calculation of other nodes, which ignores the information of the graph structure. By masked attention, the graph structure information is also put into the graph selection attention mechanism, that is, for i node, only the attention coefficient e ij of the neighbor node j and i is calculated. In order to make the attention coefficients between different nodes more easily comparable, the attention coefficients of all neighbor nodes j selected by the i node are normalized to obtain the attention weight a ij , the specific calculation formula is as follows:

[0041]

[0042] In the model, the attention layer is a single-layer feedforward neural network, which uses LeakyReLU as the activation function, so the above attention weight can be expanded as follows, where w represents the weight vector, T represents the transpose, and || represents splicing.

[0043]

[0044] After obtaining the attention weights, a linear combination of the corresponding features is calculated using the attention weights, and the final features are calculated using the sigmoid function.

[0045]

[0046] To make the self-attention learning process more stable, the above graph selection attention mechanism is extended to a multi-head graph selection attention mechanism. The following feature representation is obtained using the k-head graph selection attention mechanism, and the specific calculation formula is shown below:

[0047]

[0048] in, The graph selection attention mechanism a represents the k-th head. k The normalized attention weights, W, are calculated. k It is the linear transformation matrix corresponding to the input. The final output h' obtained by using the k-head graph selection attention mechanism contains KF' features.

[0049] Since performing multi-head attention operations in the last layer of a neural network is meaningless, the final output is obtained by averaging the results from the last layer. The specific calculation formula is shown below:

[0050]

[0051] The node representation of the initial graph is obtained through the graph selection attention mechanism described above. Then, it is used in the subsequent rumor causal generation network to generate the final rumor causal graph structure.

[0052] The specific steps of step 3 are as follows:

[0053] The input to the rumor causal generation network is the representation of N node variables learned by the graph attention layer. For any N variables... The definition of a causal generative neural network (G,f,ε) is as follows: This represents a directed acyclic graph (DAG) that represents each variable... and its parent variable set Connect them. For i variables i∈[1,N], the causal mechanism It has one hidden layer and n k A regressive neural network with hidden units:

[0054]

[0055] n h ∈N represents the number of hidden units. are parameters of the neural network, and σ is a continuous activation function. Each variable E i is independent of cause X i , all noise variables are independent of each other and follow the same distribution ε.

[0056] The probability distribution generated by the generative model is a Markov distribution about , and is acyclic, and the noise variables E i are independent of each other. In addition, the generative model also supports intervention, that is, fixing the variable X i into a constant, which also shows that the rumor causal generation network can ignore all causal effects on and set its value to X i The intervention distribution

[0057] Figure 1 The overall structure diagram of the present application is shown. The left half is the process of generating an initial graph using a text sequence, and the right half is the part of generating a causal graph using causal variables and an initial graph.

[0058] Figure 2 The structure of the graph attention layer module of the present application is shown. The left half of the figure shows the main calculation process of the graph selection attention mechanism, and the right half of the figure shows the process of obtaining rumor causal variable representation using a multi-head graph selection attention mechanism.

[0059] The present application is suitable for rumor text data, and the following experiments are carried out on three representative data sets, namely Dream4 data, Sachs flow cytometry data and pneumonia epidemic data DXY-Data, as shown in Table 1. Dream data is multi-factor artificial data, which is generated by GeneNetWeaver 2.0. The data set contains 5 data, and the target graph is a causal graph with 100 nodes. The experiment uses one of them labeled 4. Sachs flow cytometry data is real data. The pneumonia epidemic data DXY-Data is the text data used by the present application for non-structured text causal discovery research. The causal discovery of DXY-Data data belongs to the problem of high-dimensional data, and compared with structured attribute variables, it belongs to unstructured text data, and the semantic information between variables needs to be considered.

[0060] Table 1 Details of the data set

[0061] Data Data type Sample size Number of variables Sachs Real data 7466 11 Dream4 Artificial data 100 100 DXY-Data Text data 284 954

[0062] The task of the application is to find the causal relationship in the rumor data, and the purpose of rumor data processing in rumor causal discovery is to obtain the vectorization representation of the text. In the experiment, the text content is first segmented by using the jieba segmentation tool, and after segmentation, the stop words and some meaningless words are filtered once, and the obtained words are used as the word table of the causal variable word. According to the causal variable word table, each text is one-hot vectorized.

[0063] The application uses three causal discovery methods for comparative experiments, which are a method based on conditional constraints, a method based on a scoring function and a method based on a functional causal model. In the method based on conditional constraints, the PC algorithm is selected for the experiment. In the method based on the scoring function, the GES method, the CAM method and the CGNN method are selected for the experiment. In addition, in the method based on the functional causal model, the LiNGAM method, the ANM method, the CDS method, the RECI method and the IGCI method are selected for comparative experiments.

[0064] The penalty parameter of the GES method in the experiment is set to 3, the number of hidden layer units of the rumor causal generation network is set to 20, the threshold value in the CAM experiment is set to 0.001, the structure of the generation network is a neural network with one hidden layer, the activation function is ReLU, the bandwidth γ of the Gaussian kernel in the MMD score is in the range of {0.005, 0.05, 0.25, 0.5, 1, 5, 50}, nb run is set to 32, the distribution ε of the noise variable is set to N(0, 1), n epoch is set to 1000, and the learning rate of the Adam training is set to 0.01.

[0065] The high-dimensional rumor text causal discovery method based on graph selection attention mechanism, the technical effects of the application are demonstrated from the following three aspects: rumor causal discovery model performance evaluation, rumor causal discovery ablation experiment and introduction of knowledge graph to evaluate the rumor causal discovery effect.

[0066] 1. Rumor causal discovery model performance evaluation:

[0067] The experimental results of the rumor causal discovery model performance of the application are shown in Tables 2, 3 and 4, and the best experimental results in the experiment are shown in bold in the table. The indicators in the experiment are the evaluation of the predicted graph. Due to the diversity of the graph structure, the indicators AP, SID and SHD are used to evaluate the predicted graph from different angles.

[0068] The experimental results of the Sachs flow cytometer data are shown in Table 2, and HACGNN is a causal generation method based on graph selection attention mechanism proposed by the application. As can be seen from the results in Table 2, the method of the application is better than other causal discovery methods in various indicators. The AP value in the experimental results is generally low because there will be some errors in the evaluation of the graph structure by the AP index, so it is only used for relative comparison between different methods. The difference between the HACGNN method proposed by the application and the CGNN method is not large, because the Sachs data is a low-order variable sparse graph search problem, there is no graph search difficulty and there is no semantic information between variables, so the effect of HACGNN is not much different from CGNN, and both are through back propagation to learn the differentiable generation model of the data.

[0069] Table 2 Experimental results of different causal discovery methods on Sachs data

[0070] Method AP SID SHD PC 0.127 101 32 GES 0.464 64 22 CAM 0.148 91 30 LiNGAM 0.133 110 33 ANM 0.131 98 30 RECI 0.272 74 24 CDS 0.188 88 26 IGCI 0.159 84 28 CGNN 0.432 60 21 HACGNN (the method of the present application) 0.471 58 20

[0071] The experimental results of the dream4 data are shown in Table 3, and the values of each indicator on the dream4 data are obviously not as good as those on the sachs data, the value of SHD has exceeded 200, and the value of SID has even reached 8000, which fully illustrates that the evaluation of the graph structure in the causal discovery task is relative, and the values of the indicators on different experimental environments, different graph structures or different types of data may be at different levels.

[0072] Table 3 Experimental results of different causal discovery methods on dream4 data

[0073] Method AP SID SHD PC 0.133 8283 222 GES 0.250 8210 216 CAM 0.313 8012 212 LiNGAM 0.137 8264 238 ANM 0.116 8202 220 RECI 0.145 8117 218 CDS 0.181 8114 216 IGCI 0.151 8109 218 CGNN 0.265 8103 216 HACGNN (the method of the present application) 0.571 7810 182

[0074] As shown in Table 3, the method HACGNN proposed in the application also shows the best effect on the dream4 data, and in terms of the promotion amplitude, the promotion of the method of the application on the Dream4 data is more obvious than that on the Sachs data. Since the Dream4 data contains 100 variables, which is 10 times the order of the variables of the Sachs data, and the result graph is also a directed graph containing 100 nodes, the causal graph produced by the Dream4 is obviously more complex than the Sachs data result graph. It shows that the method of the application is effective for high-order variable scenarios, and the CAM method performs similarly to other methods on the Sachs data, but performs better than other traditional causal discovery methods on the Dream4 data, and is only inferior to the HACGNN method of the application. This is because the CAM method realizes feature selection between variables through non-regular maximum likelihood estimation, and also simplifies the structure search and estimation problem of the causal model. In the high-dimensional scene, the effect will be more significant. This is consistent with the idea of the HACGNN method using Lasso feature selection algorithm to restore the initial graph skeleton and then optimize the graph structure search. Therefore, in the dream4 data experiment, the CAM method and the HACGNN method of the application perform better than other traditional causal discovery methods, which also shows that the CAM method and the HACGNN method of the application are effective in high-dimensional scenarios.

[0075] Compared with the results on the Sachs data, the performance of the CGNN method on the Dream4 data is significantly reduced, and the CAM performs better than the CGNN. This is because the CGNN only uses a generative neural network to learn the generative model of the data, which may be effective when applied to general data. However, in the high-order variable scenario, it obviously loses its advantage. In contrast, the CAM and HACGNN methods use feature selection to simplify the graph structure search problem, so they perform outstandingly in high-order scenarios.

[0076] Table 4 Experimental results of different causal discovery methods on DXY-Data data

[0077] Method AP SID SHD PC 0.083 1503 197 GES 0.087 1472 182 CAM 0.203 1156 143 LiNGAM 0.043 1582 203 ANM 0.092 1463 189 RECI 0.146 1470 185 CDS 0.116 1473 187 IGCI 0.091 1470 189 CGNN 0.108 1351 162 HACGNN (the method of the present application) 0.463 874 106

[0078] The experimental results of the DXY-Data data are shown in Table 4. From the data in the table, it can be seen that the HACGNN method proposed in the application is more significant than the other two data sets. Because the DXY-Data data contains 954 variables of text data, it not only involves the causal discovery problem of high-order variables, but also considers the semantic information of text data. Other traditional causal discovery methods cannot well handle these two problems.

[0079] As shown in the experimental results of Table 4, the HACGNN has been obviously superior to all other comparative methods, under the text data of DXY-Data, the CAM method is still superior to other traditional causal discovery methods, and is slightly superior to the CGNN method, but the effect does not show the obvious advantage as before in the Dream4 data, because the DXY-Data text data is still a high-order variable problem, the CAM method performs well in the high-order scene by virtue of the simplified graph structure search, and can still be slightly superior to other traditional methods on the DXY-Data data. But the DXY-Data data is a text data, and there is semantic information association between the node variables, the CAM method does not consider the processing of the variable semantics, and the performance on the text data is obviously inferior to the HACGNN method proposed in the application.

[0080] 2. Rumor causal discovery ablation experiment:

[0081] In order to show the effect of the causal model of the application in the high-dimensional data scene and the effect of processing semantic information, two data sets are selected for ablation experiment. Table 5 is the ablation experiment result on the Dream4 data, and Table 6 is the ablation experiment result on the DXY-Data data. The Dream4 data does not contain semantic information high-dimensional data, and the DXY-Data data contains semantic information high-dimensional data.

[0082] In the ablation experiment results of Tables 5 and 6, CGNN represents learning the generation model of data only by back propagation, CGNN+Lasso represents restoring the initial graph skeleton by using Lasso feature selection algorithm first, and then learning the generation model of data by back propagation, and CGNN+Attention represents learning the semantic information between texts by using attention mechanism to obtain the representation of text variables, and then learning the generation model of data by back propagation. The hierarchical graph selection attention mechanism is used, which is slightly different from the graph attention layer in the HACGNN. The HACGNN first obtains the initial graph structure by using Lasso and then learns the attention information. In the case where the initial graph structure is not obtained, the CGNN+Attention uses the hierarchical graph selection attention mechanism to obtain the variable representation, which also learns the text semantic information.

[0083] Table 5 is the ablation experiment result on the Dream4 data

[0084] Method AP SID SHD CGNN 0.265 8103 216 CGNN+Lasso 0.413 8042 207 CGNN+Attention 0.301 8163 218 HACGNN (the method of the present application) 0.571 7810 182

[0085] Since the Dream4 data is artificially synthesized high-dimensional data, and the variables themselves do not carry any semantic information, the causal method capable of processing the high-dimensional data scene will have more obvious advantages on the Dream4 data. As shown in the results in Table 5, the causal discovery method HACGNN based on the graph selection attention mechanism proposed in the present application performs best, and the CGNN+Lasso result is only second to the HACGNN method, which shows that the HACGNN method and the CGNN+Lasso method using Lasso feature selection to optimize the graph search process have good effects in the high-dimensional data scene. The CGNN+Attention method has only a slight improvement compared with the CGNN, and the effect is not as significant as the CGNN+Lasso and the HACGNN, which shows that for the Dream4 data not containing any semantic information, the advantage of the graph selection attention mechanism is not obvious.

[0086] Table 6 shows the ablation experiment results on the DXY-Data data, and the HACGNN method still performs best, but compared with the Dream4 data results in Table 5, the CGNN+Lasso method does not perform outstandingly as in the Dream4 data, and is not superior to the CGNN+Attention, because the causal discovery task on the DXY-Data data is not only a causal discovery task in a high-dimensional data scene, but also a text data causal discovery task containing semantic information, and the HACGNN method uses feature selection to optimize the graph search problem and uses the graph selection attention mechanism to learn the context semantic information, and has a significant effect in real text data similar to the DXY-Data.

[0087] Method AP SID SHD CGNN 0.108 1351 162 CGNN+Lasso 0.237 1203 146 CGNN+Attention 0.287 1123 137 HACGNN (the method of the present application) 0.463 874 106

[0088] Table 6 shows the ablation experiment results on the DXY-Data data, and the HACGNN method still performs best, but compared with the Dream4 data results in Table 5, the CGNN+Lasso method does not perform outstandingly as in the Dream4 data, and is not superior to the CGNN+Attention, because the causal discovery task on the DXY-Data data is not only a causal discovery task in a high-dimensional data scene, but also a text data causal discovery task containing semantic information, and the HACGNN method uses feature selection to optimize the graph search problem and uses the graph selection attention mechanism to learn the context semantic information, and has a significant effect in real text data similar to the DXY-Data.

[0089] 3. Introducing a knowledge graph to evaluate the rumor causal discovery effect

[0090] The present application selects the existing ownthink graph with a larger scale and a wider knowledge coverage range and a general concept knowledge graph (ConceptNet), and comprehensively obtains 140492140 triples in the experiment. For the DXY-Data data, 894 variables among the 954 variables in the data can retrieve related knowledge in the graph. The result causal pairs obtained by the rumor causal discovery method in the present application are retrieved in the knowledge graph, and the results are shown in Table 7.

[0091] Table 7 shows the embodiment of the causal pairs obtained by different causal discovery models in the knowledge base causal relationship

[0092]

[0093] From the data in Table 7, it can be seen that most rumor causal variables can be retrieved in the knowledge graph, which shows that the selected knowledge graph has a wide coverage of words in general text data and is more suitable for knowledge retrieval. Secondly, although most variables can be retrieved in the knowledge graph, the retrieval rate of the causal pair is still low, which shows that most of the causal relationships obtained by the traditional causal discovery method are not correct from the perspective of the knowledge graph. In contrast, the causal relationship results obtained by the rumor causal discovery method HACGNN of the present application can be basically verified in the knowledge graph, which shows the effectiveness of the rumor causal discovery method proposed in the present application from the perspective of the knowledge graph.

[0094] The above experiments and results show that the present application has the following advantages and effects compared with the prior art:

[0095] (1) From the example analysis, it can be seen that the causal discovery method for text rumor data proposed in the present application can pay more attention to the causal relationship in the rumor, which is helpful for using causal knowledge to timely identify rumors.

[0096] (2) From the example analysis, it can be seen that the high-dimensional rumor causal graph search method based on HSIC Lasso feature selection proposed in the present application can further improve the rumor causal discovery accuracy in the high-dimensional causal variable scenario by introducing the HSIC Lasso feature selection method to restore the initial graph skeleton and optimize the graph search process of the rumor causal model.

[0097] (3) From the example analysis, it can be seen that the rumor causal variable semantic representation method based on graph selection attention mechanism proposed in the present application can better focus on the semantic information between rumor causal variables by introducing the graph selection attention mechanism in the rumor causal discovery process, and has more outstanding performance on rumor data containing semantic information.

Claims

1. A high-dimensional rumor text causal discovery method based on graph selection attention mechanism, characterized in that... Includes the following steps: a. Process the data to obtain a vocabulary of causal variable words, and vectorize each text. b. Input the processed causal variables from the text sequence into the HSIC Lasso feature selection layer to obtain the initial causal graph; c. Pass the set of node vectors of the initial graph after feature selection into the graph attention layer. The graph attention layer performs representation learning on the initial graph to obtain the semantic representation of the initial graph. The graph attention layer takes as input a set of node vectors from the initial graph after feature selection, and outputs a new set of node features. To obtain sufficiently expressive output features, a learnable linear transformation is performed on the input features. A weight matrix is ​​initialized for each node, and an attention coefficient is calculated for each node using a self-graph selection attention mechanism. Graph structure information is also incorporated into the graph selection attention mechanism through masked attention. i For a node, only its neighboring nodes are counted. j and i Attention coefficient e ij To make it easier to compare the attention coefficients between different nodes, i All neighboring nodes selected by the node j The attention coefficients are normalized to obtain the attention weights. α ij The attention layer in the diagram is a single-layer feedforward neural network that uses LeakyReLU as the activation function. After obtaining the attention weights, the attention weights are used to calculate a linear combination of the corresponding features, and the sigmoid function is used to calculate the final features. The input to the graph attention layer is the set of node vectors in the initial graph after feature selection, and the output is a new set of node features. The graph attention layer is a single-layer feedforward neural network that uses LeakyReLU as the activation function. After obtaining the attention weights, the attention weights are used to calculate a linear combination of the corresponding features, and the sigmoid function is used to calculate the final features. d. The causal variables of the rumor and the semantic representation of the initial graph are fed into the rumor causal generation network to obtain the final causal graph. The rumor causal generation network receives the graph representation output by the attention layer, learns the node representation of the rumor variable using the graph selection attention mechanism, learns the generation model of the rumor variable through backpropagation, trains and optimizes the model structure and network weights using a score-based method, and finds the optimal rumor causal graph using greedy search.

2. The high-dimensional rumor text causal discovery method based on graph selection attention mechanism as described in claim 1, characterized in that... In step a, the specific steps for vectorizing each text are as follows: first, the text content is segmented using a Chinese word segmentation tool; after segmentation, stop words and some meaningless words are filtered out, and the resulting words are used as a vocabulary of causal variable words; then, each text is one-hot vectorized based on the causal variable vocabulary.

3. The high-dimensional rumor text causal discovery method based on graph selection attention mechanism as described in claim 1, characterized in that... In step b, the processed causal variables in the text sequence are fed into the HSIC Lasso feature selection layer. Based on the HSIC Lasso feature selection method, non-redundant features that have strong statistical dependence on the output value are found through the input features.

Images