CN114742564A - A false reviewer group detection method incorporating complex relationships - Google Patents

Abstract

The invention belongs to the field of artificial intelligence, and provides a false comment group detection method fusing complex relationships, which is used for false comment group detection on an online trading platform. The method comprises three stages of node representation updating, model training and false comment group detection. The method applies the trained model to a real data set, can identify the false reviewers, and can well distinguish the false reviewer group from the normal reviewers. The method is based on the complex relation characteristics of the nodes, makes full use of valuable relation information among reviewers, integrates the embedding process and the clustering detection process to obtain a false reviewer group detection model taking a target as a guide, and can overcome the problems of poor universality, low detection effect and the like of the conventional group detection method.

Description

A group detection method of fake reviewers integrating complex relationships

technical field

The invention relates to the field of artificial intelligence, in particular to a false commenter group detection method integrating complex relationships.

Background technique

The rapid popularity of online review systems has made reviews an important basis for people to buy products. More and more people will check reviews on the platform before purchasing products, and will also make comments on products after purchasing them. These reviews can provide customers with useful information and first-hand product experience, so the quality of online reviews is particularly important. Fake reviews that do not conform to the facts of the product will affect the reputation of the product and confuse buyers.

Most of the existing fake review detection technologies are realized by big data and artificial intelligence methods. relationship between users. In the past, researchers mainly focused on detecting individual fake reviewers. However, groups of fake reviews tend to cause more harm to online review systems, and it is difficult to find fake reviewers in groups: fake reviews in groups may be normal individual reviews. , the previous detection technology of fake reviews alone is difficult to play a role. In addition, it is difficult to establish relationships between fake reviewers, and this complex relationship enables the model to capture the connections between reviewers within the group, thereby assisting the detection of fake reviewer groups.

The current fake comment group detection methods can be classified into the following categories:

Detection method based on clustering algorithm. Detection algorithms based on clustering algorithms usually use algorithms such as graph neural networks to learn node embedding representations, then cluster nodes through clustering algorithms, and finally detect fake comment groups through detection methods. Common clustering algorithms such as partition-based clustering algorithm KMeans, density-based clustering algorithm DBSCAN.

(1) The KMeans clustering algorithm mainly divides all points in the sample space into K groups, and the similarity is usually measured by Euclidean distance. The main process of the algorithm is as follows: randomly place K centroids, and each cluster There is a centroid in . The distance from each point to the centroid is calculated, and each data point is assigned to its nearest centroid, forming a cluster. In the iterative process, the position of the centroid K is recalculated.

(2) The DBSCAN clustering algorithm first determines the type of each point, and each data point in the data set may be a core point or a boundary point. A data point is a core point if there are at least M points in its neighborhood within the specified radius R, if there are less than M data points in the neighborhood of a data point, or it can be reached from a core point, That is, it is within the R distance from the core point, indicating that it is a boundary point. Core points that are neighbors will be connected and placed in the same cluster, and border points will be assigned to each cluster.

Graph-based detection methods. Starting from the subgraph, use the attributes of the node or subgraph to judge the suspiciousness of the group, so as to realize the whole detection process. Some methods aggregate relationships from differences in graph topology, timing, and ratings, using joint probabilities to detect groups of fake reviewers. This method ignores the structural characteristics of the nodes themselves, and does not consider the complex relationships between nodes. There are also methods that propose several main characteristics of a population, such as synchronicity, moderation, and dispersion, and detect population anomalies by computing certain metrics. This method lacks universality in practical applications. For different networks or data sets, specific indicators need to be proposed to better complete the task of detecting fake comment groups. If it is promoted, the detection accuracy will drop significantly. In addition, this method only considers the characteristics within the group, and still lacks the consideration of the complex relationship between reviewers.

SUMMARY OF THE INVENTION

In the existing false comment group detection methods, the embedding process and the subsequent clustering and detection process are separated, and the training process lacks target guidance. If the result of representation learning is not suitable for detection, the obtained false comment group detection results will be. very poor. In addition, the complex relationships in the review network are ignored, and the valuable relationship information among reviewers within the group cannot be utilized.

Aiming at the problems existing in the prior art, the present invention proposes a method for detecting groups of false commenters that integrates complex relationships, which is target-oriented, based on the complex relationship features of nodes, uses the features to learn the complex relationship representation of nodes, and uses automatic encoder to reconstruct the topological information of the graph, this method is used for the detection of fake comment groups on online sales platforms; in order to integrate the embedding process and clustering with the detection process, this method adopts a self-supervised training model, using clustering and detection. The results guide the optimization of the model.

In order to achieve the above purpose, the technical solutions adopted in the present invention are as follows: a method for detecting groups of false commenters integrating complex relationships, using a graph neural network based on an attention mechanism to characterize and update the comment nodes in the comment network; design graph reconstruction Loss and self-supervised distribution loss are used for model training, and the optimal model is obtained and applied to false reviewer group detection to identify false reviewer groups in the review network; the specific steps are as follows:

The first step is to update the node representation to obtain the reconstruction graph; the model extracts the adjacency matrix and attribute matrix of the review network, and obtains the complex relationship matrix according to the adjacency matrix. After obtaining the complex relation matrix, the graph attention encoder integrates the complex relation with the message passing process, effectively encodes the high-order structural information and node attribute information of the network, and then updates the node representation. The graph neural network based on the attention mechanism is used as the encoder; the initial feature of the node is used as the initial embedding of the node, and the complex relationship of the node is integrated into the graph neural network based on the attention mechanism, so that the node representation can express both high-order structural features and attribute features;

1.1) Calculate the similarity of nodes; in order to simplify the calculation and reduce the model parameters, the nodes are limited to the first-order neighbor nodes of the central node, and the calculation formula is as follows:

c _ij =a(Wh _i , Wh _j )#(1)

In the formula, c _ij represents the importance of node j to node i, W represents the weight matrix; hi and h _j represent the eigenvectors of node i and node _j respectively; a represents the function of calculating the similarity of nodes;

1.2) Calculate the complex relationship matrix; the review network has complex structural relationships, and the complex relationships between its nodes contain valuable information. The complex relationship matrix of a node is obtained by considering the node's higher-order neighbor nodes:

M=(B+B ² +...+B ^t )/t# (2)

In the formula, B is the transformation matrix. When there is an edge between node i and node j, B _ij = 1/d _i , where d _i is the degree of the node; when there is no edge between node i and node j, B _ij = 0; the matrix M represents a complex relationship matrix, and M _ij is the complex relationship between node i and node j at order t;

1.3) Fusion of complex relationships; with a single-layer feedforward neural network as the calculation method, the complex relationship matrix M is fused with the graph neural network based on the attention mechanism. When calculating the similarity between nodes, not only the similarity between node representations, but also the impact of complex relationships between nodes on the similarity must be considered; LeakyReLU is selected as the activation function to increase the nonlinear factors of the model and enhance the feature expression ability of the model. After fusing complex relations, the importance expression of node j to node i is rewritten as:

1.4) Update node representation; the softmax function normalizes the importance of neighbor nodes, so that the importance of first-order neighbor nodes to the central node is distributed between [0, 1], and aggregates the features of neighbor nodes to update node representation ;

In formula (4), α _ij represents the normalized attention coefficient; N _i represents the first-order neighbor set of node i;

为节点i的相邻节点j在第l层上的表示，

代表节点i在第l+1上的表示；节点最终表征经多层聚合获得；In formula (5),

is the representation of the adjacent node j of node i on the lth layer,

represents the representation of node i on the l+1th; the final representation of the node is obtained by multi-layer aggregation;

The second step is model training; the model first uses the autoencoder to reconstruct the topological information of the graph, and calculates the loss by calculating the difference between the adjacency matrix of the original graph and the reconstructed graph. This loss is the first part of the loss. The second part of the loss is obtained by the self-supervised training method. The model uses the DBSCAN clustering algorithm to determine the core points in the review network, and calculates the distance between all nodes and the core points, and uses the KL divergence as the loss of this part. The final loss function is composed of the above two loss functions together and is used to jointly train the model. After the loss is calculated, the model parameters are updated using the gradient descent algorithm, and the training is complete.

Design the graph reconstruction loss function and the self-supervised distribution loss function, update the parameters of the graph neural network model based on the attention mechanism to complete the training. The specific steps are:

2.1) Calculate the reconstruction loss function of the graph; according to the topology information of the reconstructed graph of the encoder, calculate the difference between the adjacency matrices to obtain the reconstruction loss of the reconstructed graph and the original graph; the formula is:

为邻接矩阵；H为更新后的节点表征矩阵；σ为激活函数；In the formula,

is the adjacency matrix; H is the updated node representation matrix; σ is the activation function;

During training, cross entropy is used as the loss function:

代表重构的邻接矩阵中的相应元素。此部分训练需要最小化重构损失，重构损失定义如下：where y represents the value of an element in the adjacency matrix,

represents the corresponding element in the reconstructed adjacency matrix. This part of the training needs to minimize the reconstruction loss, which is defined as follows:

2.2) Calculate the self-supervised distribution loss function; one of the challenges of false comment detection methods is that there is no label to guide the training of the model; this model adopts a self-supervised training method, and uses pseudo-labels to optimize node embedding representation; clustering algorithms are used to cluster nodes. This model uses the K-Means algorithm for clustering:

In the formula, μ _i is the mean of all nodes in Si, and _k is the number of sets to be clustered.

After obtaining all the fake comment groups, the DBSCAN clustering algorithm is used to determine the core points in the comment network, and the distance distribution between each node and the core points is calculated;

During the training process, it is necessary to continuously learn the distribution of data to distinguish normal nodes from abnormal nodes. p _iu represents the pseudo-marker calculated by the model, and _qiu represents the distance distribution between the features of all nodes and the core points detected by DBSCAN. The definition of q _iu is as follows:

In the formula, u _u represents the representation of the core point detected by DBSCAN; Z _i represents the representation of the current processing node; _uk represents the representation of the core point of the kth category. The formula calculates the distance between the representation of the node and the representation of the core point. If the distance between the node and the core point is close enough, the node can be considered to belong to the group, and it is considered to be a normal node. Assuming that a certain node is far away from the core point, it can be considered that this node is an outlier, that is, it is considered to be a corresponding false comment group. Node labels can be obtained by the following formula:

S _i =argmax·q _iu #(11)

Use KL divergence as a loss function to measure the difference between the distance distribution between nodes and core points and their pseudo-labels;

KL divergence mainly measures the difference between the probability distribution Q and the reference probability distribution P. Different from the label obtained by equation (11), the target distribution p _iu is considered as the real label, which is calculated by Q in the training process, and p _iu depends on the P distribution and is updated according to the stage, which is regarded as the stage. Internal self-supervised labels. The main role of the target distribution is to supervise the learning of the model and guide the update of the distribution Q. The formula for calculating P is as follows:

In the formula, q _ik represents the distance distribution between the features of all nodes and the core point of the k-th category. The loss function for self-supervised optimized embedding is as follows:

2.3) Calculate the joint loss function; the expression of the joint loss function is:

L=·L _r +βL _c #(14)

In the formula, L _r is the graph reconstruction loss function, L _c is the self-supervised distribution loss function, and β is the weight between the two loss functions;

2.4) Model training, setting the initial parameters of the graph neural network model based on the attention mechanism, based on the joint loss function, iterative training process, to obtain the best parameters of the graph neural network model based on the attention mechanism;

The third step is the detection of false comment groups; the graph neural network model based on the attention mechanism obtained in the second step is used to detect the real comment network and save the detection results.

The graph reconstruction loss function adopts the cross entropy loss function; the clustering algorithm for clustering the nodes adopts the KMeans clustering algorithm.

The specific method of model training in 2.4) is as follows:

Set the initial parameters of the graph neural network model based on the attention mechanism, including the number of aggregation layers of the graph neural network model based on the attention mechanism, the node embedding dimension, the number of clusters and the number of training iterations of the KMeans clustering algorithm;

During the model training process, the parameters are continuously adjusted, and the optimal parameters are determined according to the decline of the joint loss function during the training process or the final detection result of the model;

Specifically: input the review network and the adjacency matrix of the network into the model, run and train the model, record the detection performance of the model after this training, repeat the training multiple times under the same set of hyperparameters, and take the average value of the detection accuracy as the final Result detection accuracy; after completing the model training under a set of parameters, adjust the parameters in the model according to the control variable method, adjust a certain parameter of the model in the direction of increasing the average accuracy, and keep other parameters unchanged; repeat the adjustment of parameters, Retain a set of parameter settings that maximizes the average discriminant accuracy of the model, and the model is trained.

Beneficial effects of the present invention: the method can not only identify false reviewers, but also distinguish the group of false reviewers from normal reviewers. Based on the complex relationship characteristics of nodes, this method makes full use of the valuable relationship information between reviewers, integrates the embedding process with the clustering detection process, and obtains a target-oriented false reviewer group detection model. The group detection method has problems such as poor universality and low detection effect.

Description of drawings

Fig. 1 is the basic frame diagram of the present invention;

Fig. 2 is the working flow chart of the present invention;

Fig. 3 is the recall rate change chart in the training process of an embodiment of the present invention;

FIG. 4 is a change diagram of a loss function in a training process according to an embodiment of the present invention;

FIG. 5 is a visualization diagram of a population detection result according to an embodiment of the present invention.

Detailed ways

The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

A false reviewer group detection method integrating complex relationships, including three stages: node representation update; model training; false review group detection.

In the first step, the node representation is updated. In this stage, the graph neural network based on attention mechanism is used as the encoder, the initial node features are used to represent the initial embedding of nodes, and the complex relationship of nodes is integrated into the graph neural network based on attention mechanism, so that the node representation has the ability to express high-order structural features and attributes. characteristic ability.

1.1) Calculate node similarity. In order to simplify the calculation and reduce the model parameters, the nodes are limited to the one-hop neighbors of the central node. The calculation formula is as follows:

c _ij =a(Wh _i , Wh _j )#(1)

In the formula, c _ij represents the importance of node j to node i, W represents the weight matrix; hi and h _j represent the eigenvectors of node i and node _j respectively; a represents the function of calculating the similarity of nodes;

1.2) Calculate the complex relationship matrix. The review network has complex structural relationships, and the complex relationships between its nodes contain valuable information. The complex relationship matrix of a node can be obtained by considering its higher-order neighbors:

M=(B+B ² +...+B ^t )/t#(2)

In the formula, B is the transformation matrix. If there is an edge between node i and node j, B _ij = 1/d _i , where d _i is the degree of the node. When there is no edge between node i and node j, B _ij = 0. The matrix M represents a complex relationship matrix, and M _ij is the complex relationship between node i and node j at order t.

1.3) Integrate complex relationships. The single-layer feedforward neural network is selected as the calculation method, so that the complex relationship matrix M is integrated with the graph attention network. The similarity between node representations should also consider the impact of complex relationships between nodes on the similarity. Finally, LeakyReLU is selected as the activation function to increase the nonlinear factor of the model, so that the feature expression ability of the model is enhanced. After fusing complex relations, the importance expression of node j to node i is rewritten as:

1.4) Update node representation. In order to make the importance of neighbor nodes to the central node distributed between [0, 1], the softmax function is used to normalize the importance of neighbor nodes, and the features of neighbor nodes are aggregated to update the node representation.

为节点i的相邻节点j在第l层上的表示，

代表节点i在第l+1上的表示。节点最终的表征经多层聚合获得。In equation (4), α _ij represents the normalized attention coefficient, and Ni represents the first-order neighbor set of node _i ; in equation (5),

is the representation of the adjacent node j of node i on the lth layer,

represents the representation of node i on the l+1th. The final representation of the node is obtained through multi-layer aggregation.

The second step is model training. First, the loss function is designed. After using the designed loss function to calculate the loss, the model parameters are updated to complete the training. The model first uses the decoder to reconstruct the original network to calculate the adjacency matrix difference loss between the original network and the reconstructed network. Since the nodes in the false commenter group detection task have no labels, the self-supervised training method is used to optimize the embedding. First, the DBSCAN clustering algorithm is used to generate the core points in the comment network, and the KL divergence is used to measure the distance between the core points and other nodes. , and then compute the difference between the pseudo-labels and the learned embedding distribution. After the loss calculation is completed, the gradient descent algorithm is used to update the model parameters, completing the training.

2.1) Computational graph reconstruction loss. The inner product is used to reconstruct the original graph, and the reconstruction formula is:

为重构图的邻接矩阵，为使得重构出来的邻接矩阵

与输入的邻接矩阵尽可能的相似。训练过程中，采用交叉熵作为损失函数：In the formula, H is the embedding vector of the learned node,

is the adjacency matrix of the reconstructed graph, so that the reconstructed adjacency matrix

Be as similar as possible to the input adjacency matrix. During training, cross entropy is used as the loss function:

代表重构的邻接矩阵中的相应元素。此部分训练需要最小化重构损失，重构损失定义如下：where y represents the value of an element in the adjacency matrix,

represents the corresponding element in the reconstructed adjacency matrix. This part of the training needs to minimize the reconstruction loss, which is defined as follows:

2.2) Calculate the distribution loss. One of the challenges of fake review detection methods is that there is no label to guide the training of the model. This model adopts a self-supervised way, using pseudo-labels to optimize node embedding representation. Since the nodes in the graph are independent, in the training process, all nodes are clustered first. This model uses the K-Means algorithm for clustering:

In the formula, μ _i is the mean of all nodes in Si, and _k is the number of sets to be clustered. After getting all the fake comment groups, the DBSCAN algorithm is used to detect abnormal groups. The DBSCAN algorithm first distinguishes the core points and boundary points in the graph, uses the detected core points as the core points in the training model, and calculates the distance between the representation of other nodes and the representation of the core point. During the training process, it is necessary to continuously learn the distribution of data to distinguish normal nodes from abnormal nodes. p _iu represents the pseudo-marker calculated by the model, and _qiu represents the distance distribution between the features of all nodes and the core points detected by DBSCAN. The definition of q _iu is as follows:

In the formula, u _u represents the characterization of the core point detected by DBSCAN. The formula calculates the distance between the representation of the node and the representation of the core point. If the distance between the node and the core point is close enough, the node can be considered to belong to the group, and it is considered to be a normal node. Assuming that a certain node is far away from the core point, it can be considered that this node is an outlier, that is, it is considered to be a corresponding false comment group. Node labels can be obtained by the following formula:

S _i =argmax·q _iu #(11)

This model uses KL divergence to measure the difference between the pseudo-label and the learned distribution. KL divergence mainly measures the difference between the probability distribution Q and the reference probability distribution P. Different from the label obtained by equation (11), the target distribution p _iu is considered as the real label, which is calculated by Q in the training process, and p _iu depends on the P distribution and is updated according to the stage, which is regarded as the stage. Internal self-supervised labels. The main role of the target distribution is to supervise the learning of the model and guide the update of the distribution Q. The formula for calculating P is as follows:

The loss function for self-supervised optimized embedding is as follows:

2.3) Calculate the joint loss function. The overall loss function of the model consists of a graph reconstruction loss function and a self-supervised distribution loss function. The final loss function expression is:

L=·L _r +βL _c #(14)

In the formula, L _r is the reconstruction loss, L _c is the distribution loss, and β is used to control the weight between the two losses.

2.4) Model training. The training of the model is carried out according to the following steps: setting the initial hyperparameters, including the number of aggregation layers of the graph attention network, the node embedding dimension, the number of clusters of the KMeans clustering algorithm, and the number of training iterations.

During the model training process, the hyperparameters need to be manually adjusted to achieve the best detection effect of the model. Generally speaking, hyperparameters need to be determined according to the decline of the loss function during the training process or the final detection result of the model. After the hyperparameters are set, input the information such as the comment network and the adjacency matrix of the network into the model, run the model, wait for the model training to complete, record the detection performance of the model after this training, and repeat the above process multiple times under the same set of hyperparameters. Take the average of detection accuracy as the final result detection accuracy. After the model training under a set of hyperparameters is completed, the hyperparameters in the model are adjusted according to the control variable method, and a certain hyperparameter of the model is adjusted in the direction of increasing the average accuracy, and other parameters are kept unchanged. Repeat the hyperparameter adjustment process, keep a set of hyperparameter settings that maximizes the average discriminant accuracy of the model, and complete the model training.

The third step is to detect fake comment groups. Use the model trained in the previous step and the hyperparameters to detect the real review network and save the detection results of the model on the review network.

Table 1 The running process of this algorithm

In conjunction with the scheme of the present invention, carry out experimental analysis as follows:

The present invention verifies the false comment group detection effect on the Amazon data set processed by the researchers. The basic situation of the data set is shown in Table 2. The relationship type U-P-U in the table means that two users have reviewed at least one of the same product. U-S-U means that two reviewers reviewed the same rating within a week. U-V-U represents two reviewers with similar reviews. The experiments are carried out on four datasets, corresponding to the above three relations, and a dataset consisting of three relations. The four datasets are Amazon_p, Amazon_s, Amazon_v and Amazon datasets.

Table 2 The basic situation of the fake comment dataset used in the experiment

The experimental analysis process of a false reviewer group detection method integrating complex relationships can be divided into two parts: the method is compared with the existing false review group detection methods, and the recall rate is used as an evaluation index to verify the superiority of the present invention ; Visualize the training process and detection results, so as to more intuitively analyze the rationality of the model design and the validity of the detection effect.

(1) Comparison experiment of test results

The comparison experiment selects several false comment group detection methods that have been proposed by researchers to compare with this method, in which Graph-Strainer uses a graph-based method to discover target items, and on this basis detects false commenter groups, through 2 jumpers The graph solves the crowd detection problem. ColluEage uses Markov random fields to detect colluding fake reviewers and fake review activity. DeFrauder utilizes product review graphs, combined with behavioral signals to detect candidate fraud groups, maps candidate fraud groups into an embedding space, assigns scores to each group, and finally identifies groups of fake reviewers based on their scores. In addition to the above comparison methods, in order to verify the effectiveness of the modules in this method, two additional decoupling detection methods are added to the experiment: GCN+KMeans+DBSCAN and GAT+KMeans+DBSCAN. The first method uses GCN to embed the initial data set, and the second method uses GAT to embed the initial data set. After obtaining the embedding, both methods use KMeans clustering method and DBSCAN method to detect the embedding results.

The experimental results of this method and the comparative method are shown in Table 3. Through the longitudinal comparison of the experimental results, it can be seen that the performance of this method is obviously better than that of other methods, and the detection effect is greatly improved. The results of GAT+KMeans+DBSCAN are better than GCN+KMeans+DBSCAN, proving the effectiveness of using GAT as a graph encoder. Compared with GCN, GAT can aggregate neighbor features according to the similarity between the central node and neighbor nodes, so that the representation results of normal nodes will not aggregate the information of a large number of false nodes. Through the horizontal comparison of the experimental results, it can be seen that the method has achieved the best results when considering three different relationships and considering all the relationships, which shows that the KMeans clustering algorithm is integrated in the deep learning model, and in the During the training process, the core points are continuously updated iteratively to obtain more accurate detection results.

Table 2 Test results

(2) Visualization experiment of training process and detection results

The purpose of the visualization experiment is to show the rationality of the design of this method by analyzing the loss in the training process and the change of the recall rate of the result, and to intuitively show the validity of the detection result by visualizing the detection result.

Figure 3 shows the change of recall rate during the training process. The overall situation of the graph shows that the recall rate of the detection results is continuously improving as the training progresses, and the rationality of the model design has been confirmed.

The change of the loss function during training is shown in Figure 4. Combined with Figure 3 and Figure 4 for analysis, with the continuous reduction of the loss function, the recall rate is also continuously improved, which shows that the method is constantly updating the representation learning results, and the obtained representation learning results can also be applied to fake reviewers. Crowd detection. Conversely, the loss function we designed can feed the loss well to the model and supervise the learning of the model, solving the problem that the representation learning result may not be suitable for the detection method.

Figure 5 shows the clustering results of the model on the Amazon data set. It can be seen that the present invention has a good effect on the problem of false reviewer group detection. The black entities represent fake comment groups, which are mainly concentrated in the lower left, and the gray entities represent normal comment nodes, which are mainly concentrated in the upper right.

The above-mentioned embodiments only represent the embodiments of the present invention, but should not be construed as a limitation on the scope of the patent of the present invention. Several modifications and improvements can also be made, which all belong to the protection scope of the present invention.

Claims (3)

1. A false reviewer group detection method fused with complex relationships, characterized in that the false reviewer group detection method fused with complex relationships uses a graph neural network based on an attention mechanism to characterize and update the review nodes in the review network; Design graph reconstruction loss and self-supervised distribution loss for model training, and after obtaining the optimal model, apply it to false reviewer group detection and identify false reviewer groups in the review network; the specific steps are as follows: The first step is to update the node representation to obtain the reconstructed graph; the graph neural network based on the attention mechanism is used as the encoder; the initial feature of the node is used as the initial embedding of the node, and the complex relationship of the node is integrated into the graph neural network based on the attention mechanism, Make node representation express high-order structural features and attribute features at the same time; 1.1) Calculate the node similarity; the node is limited to the first-order neighbor nodes of the central node, and the calculation formula is as follows: c _ij =a(Wh _i , Wh _j )#(1) In the formula, c _ij represents the importance of node j to node i, W represents the weight matrix; hi and h _j represent the eigenvectors of node i and node _j respectively; a represents the function of calculating the similarity of nodes; 1.2) Calculate the complex relationship matrix; obtain the complex relationship matrix of the node by considering the high-order neighbor nodes of the node: M=(B+B ² +...+B ^t )/t# (2) In the formula, B is the transformation matrix. When there is an edge between node i and node j, B _ij = 1/d _i , where d _i is the degree of the node; when there is no edge between node i and node j, B _ij = 0; the matrix M represents a complex relationship matrix, and M _ij is the complex relationship between node i and node j at order t; 1.3) Fusion of complex relationships; using a single-layer feedforward neural network as the calculation method, the complex relationship matrix M and the graph neural network based on the attention mechanism are fused, specifically multiplying the complex relationship matrix and the node similarity; LeakyReLU is selected as the activation function, which fuses complex relationships, the expression for the importance of node j to node i is rewritten as:

1.4) Update node representation; the softmax function normalizes the importance of neighbor nodes, so that the importance of first-order neighbor nodes to the central node is distributed between [0, 1], and aggregates the features of neighbor nodes to update node representation ;

In formula (4), α _ij represents the normalized attention coefficient; N _i represents the first-order neighbor set of node i; In formula (5),

is the representation of the adjacent node j of node i on the lth layer,

represents the representation of node i on the l+1th; the final representation of the node is obtained by multi-layer aggregation; The second step is model training; design the graph reconstruction loss function and the self-supervised distribution loss function, and update the parameters of the graph neural network model based on the attention mechanism to complete the training. The specific steps are: 2.1) Calculate the reconstruction loss function of the graph; according to the topology information of the reconstructed graph of the encoder, calculate the difference between the adjacency matrices to obtain the reconstruction loss of the reconstructed graph and the original graph; the formula is:

In the formula,

is the adjacency matrix; H is the updated node representation matrix; σ is the activation function; 2.2) Calculate the self-supervised distribution loss function; adopt the self-supervised training method, use pseudo-labels to optimize the node embedding representation; use the clustering algorithm to cluster the nodes, and use the DBSCAN clustering algorithm to determine the core points in the comment network, and calculate each node. Distance distribution between nodes and core points; use KL divergence as a loss function to measure the difference between the distance distribution between nodes and core points and their pseudo-labels; 2.3) Calculate the joint loss function; the expression of the joint loss function is: L=·L _r +βL _c # (7) In the formula, L _r is the graph reconstruction loss function, L _c is the self-supervised distribution loss function, and β is the weight between the two loss functions; 2.4) Model training, setting the initial parameters of the graph neural network model based on the attention mechanism, based on the joint loss function, iterative training process, to obtain the best parameters of the graph neural network model based on the attention mechanism; The third step is the detection of false comment groups; the graph neural network model based on the attention mechanism obtained in the second step is used to detect the real comment network and save the detection results. 2. The false commenter group detection method fused with complex relationships according to claim 1, wherein the graph reconstruction loss function adopts a cross-entropy loss function; the clustering algorithm for clustering the nodes adopts KMeans clustering algorithm. 3. the false commenter group detection method of fusion complex relationship according to claim 2, is characterized in that, the concrete practice of model training in described 2.4) is as follows: Set the initial parameters of the graph neural network model based on the attention mechanism, including the number of aggregation layers, the node embedding dimension, the number of clusters and the number of training iterations of the KMeans clustering algorithm of the graph neural network model based on the attention mechanism; During the model training process, the parameters are continuously adjusted, and the optimal parameters are determined according to the decline of the joint loss function during the training process or the final detection result of the model; Specifically: input the review network and the adjacency matrix of the network into the model, run and train the model, record the detection performance of the model after this training, repeat the training multiple times under the same set of hyperparameters, and take the average value of the detection accuracy as the final Result detection accuracy; after completing the model training under a set of parameters, adjust the parameters in the model according to the control variable method, adjust a certain parameter of the model in the direction of increasing the average accuracy, and keep other parameters unchanged; repeat the adjustment of parameters, Retain a set of parameter settings that maximizes the average discriminant accuracy of the model, and the model is trained.

Images