Pseudo-label generation method for low-confidence samples based on social media interaction relations

Abstract

The present invention provides a method for generating pseudo labels for low-confidence samples based on multiple interaction relationships in social media. The method comprises: constructing M different social network account interaction relationship topology graphs using M types of interaction relationships existing in social media; training M graph neural networks using the M different social network account interaction relationship topology graphs to obtain M models; using the M models to obtain predicted labels for sample nodes; calculating the consistency between the M predicted labels for each sample node; if the consistency value corresponding to the sample node is lower than a confidence threshold, the sample node is a low-confidence sample; under each model, for each low-confidence sample, voting to correct the predicted label of the low-confidence sample based on the predicted labels of its neighboring non-low-confidence samples; and aggregating the predicted labels of the confidence samples corrected by the M models to obtain a final predicted label as the pseudo label of the confidence sample.

Description

Technical Field

[0001] The present invention relates to the field of graph neural network technology, and in particular to a method for generating pseudo labels for low-confidence samples based on social media interaction relationships. Background Art

[0002] In recent years, Graph Neural Networks (GNNs) have achieved remarkable success in various graph-based network tasks, such as traffic prediction, recommendation systems, biochemistry and other application fields.

[0003] GNNs typically require large amounts of labeled data for training to achieve satisfactory performance. However, obtaining this labeled data is time-consuming, labor-intensive, and expensive. To address this issue, a semi-supervised learning (SSL) approach has been proposed. This approach involves first training a model on labeled data, then using the trained model to predict labels for unlabeled data, generating pseudo-labels. Finally, the labeled data and the newly generated pseudo-labeled data are combined as new training data, thus making full use of the unlabeled data.

[0004] Therefore, generating reliable pseudo-labels is the key to improving model performance. This is because as training progresses, the confidence of the pseudo-labels will gradually increase. If a large number of unlabeled samples are labeled with incorrect pseudo-labels and used for model training, it will lead to a distribution shift problem between the original labeled dataset and the enhanced dataset, which will seriously affect the performance of the model. At the same time, the gradual increase in the confidence of the pseudo-labels also means that during the model training process, there are a small number of low-confidence samples that are difficult to learn and are submerged in a large amount of data and cannot be fully learned, resulting in the overall learning effect cannot be further improved. Therefore, the inventors believe that realizing the mining of low-confidence samples and optimizing the pseudo-label generation process for low-confidence samples can further improve the performance of the model. Summary of the Invention

[0005] To address the problem of low model performance caused by the difficulty of learning low-confidence samples during model training, the present invention provides a pseudo-label generation method for low-confidence samples based on social media interaction relationships.

[0006] The present invention provides a method for generating pseudo labels for low-confidence samples based on social media interaction relationships, comprising:

[0007] Step 1: Use the M types of interaction relationships existing in social media to construct M different social network account interaction relationship topology graphs;

[0008] Step 2: Use M different social network account interaction relationship topology graphs to train M graph neural networks respectively to obtain M models;

[0009] Step 3: Use M models to obtain the predicted labels of the test set sample nodes;

[0010] Step 4: For each sample node, calculate the consistency between the M predicted labels of the sample node;

[0011] Step 5: If the consistency value corresponding to the sample node is lower than the set confidence threshold, the sample node is a low-confidence sample;

[0012] Step 6: Under each model, for each low-confidence sample, vote to correct the predicted label of the low-confidence sample based on the predicted labels of its neighboring non-low-confidence samples;

[0013] Step 7: For each low-confidence sample, aggregate the predicted labels of the M models after correction of the confidence sample to obtain the final predicted label as the pseudo label of the confidence sample.

[0014] Furthermore, in step 2, the M graph neural networks adopt different model types; and / or, the M graph neural networks adopt different hyperparameters; and / or, use different training sample sets.

[0015] Furthermore, step 4 specifically includes:

[0016] Step 4.1: For each sample node i, obtain the predicted label vector h of M models for the sample node i i ; Wherein, the predicted label vector h i The kth dimension h in i,k Indicates the number of models that judge sample node i as the kth category;

[0017] Step 4.2: Follow the formula Calculate the consistency C between the M predicted labels of the sample node i i .

[0018] Furthermore, in step 7, the predicted labels of the confidence sample corrected by the M models are aggregated, specifically including: voting using the hard labels predicted by each model.

[0019] Beneficial effects of the present invention:

[0020] The present invention utilizes the various relationships between accounts in the social media platform, uses different relationships and training samples to train different GNN models, and determines samples with different classification results from different models as low-confidence samples. The predicted labels of non-low-confidence samples in the neighborhood of the low-confidence samples are voted and corrected. Compared with previous graph neural networks, the method proposed in the present invention effectively utilizes the various interactive relationships between accounts, enables the multiple relationships to supervise and complement each other, and screens out low-confidence samples that are more likely to be misclassified. And by mining the structure of the social network, the output of low-confidence (more likely to be misclassified) samples is corrected, which improves the prediction accuracy of the pseudo-labels of low-confidence samples, thereby improving the accuracy of malicious account detection in the subsequent process of malicious account detection based on the pseudo-label data. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 One of the flow charts of a method for generating pseudo labels for low-confidence samples based on social media interaction relationships provided by an embodiment of the present invention;

[0022] Figure 2 A second flow chart of a method for generating pseudo labels for low-confidence samples based on social media interaction relationships provided by an embodiment of the present invention;

[0023] Figure 3 A schematic diagram of low-confidence sample output correction provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0024] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0025] like Figure 1 As shown, an embodiment of the present invention provides a method for generating pseudo labels for low-confidence samples based on social media interaction relationships, comprising the following steps:

[0026] S101: constructing M different social network account interaction relationship topology graphs using M types of interaction relationships existing in social media;

[0027] Specifically, taking Twitter as an example, we construct M different social network account interaction relationship topologies through M relationships such as RT forwarding relationship, @ mention relationship, follower being followed relationship, following following relationship, reply reply relationship, #Hashtag co-occurrence relationship, and URL co-occurrence relationship. For example, the followers_list contained in user information stores the user's follower IDs, based on which the follower-followed relationship is constructed; the friends_list contained in user information stores the IDs of the account's friends, based on which the friends relationship is constructed. The mention_list contained in user information stores the usernames (screen_names) of other users the user has @ed, based on which @ relationships between accounts are constructed; if two users' tweets contain the same URL, then a URL co-occurrence relationship exists between the two.

[0028] S102: Using M different social network account interaction relationship topology graphs, train M graph neural networks respectively to obtain M models;

[0029] Specifically, the topological diagram of the interaction relationship between different social network accounts They are used as input to M graph neural networks (GNNs) for training. Since the social network account interaction relationship topology constructed by each relationship contains different information, the diversity of the graph neural network can be increased by using different social network account interaction relationship topology maps.

[0030] In addition, in order to further increase the differences between models, one or a combination of the following measures can be used: GNN1 to GNN M Use different types of models (such as GCN, GAT, GraphSAGE, RGCN, etc.); use different hyperparameters (dropout, number of layers, number of hidden layer neurons, learning rate, etc.); use different training sample sets, that is, randomly sample some samples in the training set each time, construct different social network account interaction relationship topology subgraphs, and input different training sample sets for different models.

[0031] S103: Use M models to obtain predicted labels for the sample nodes in the test set;

[0032] Specifically, for the convenience of subsequent description, the j-th model GNN j The predicted label of sample node i is recorded as z i,j , where 0<j<M, 0≤i<N, N is the graph data The number of nodes in the.

[0033] S104: For each sample node, calculate the consistency between the M predicted labels of the sample node;

[0034] Specifically, based on the previous step, statistics are collected on the models with consistent predicted labels for sample node i; for example, if among all models, the predicted labels of model GNN1 and model GNN3 are consistent with the label z of sample node i, i,1 and z i,3 If the label corresponds to category 1, then it is considered that there are two models for category 1. In this way, for each sample node i, the predicted label vector h of M models for the sample node i is obtained. i ; Among them, the predicted label vector h i The kth dimension h in i,k Indicates the number of models that judge sample node i as the kth category.

[0035] Then, according to the formula Calculate the consistency C between the M predicted labels of the sample node i i .

[0036] S105: If the consistency value corresponding to the sample node is lower than the set confidence threshold, the sample node is a low-confidence sample;

[0037] Specifically, based on the above example, if C i <τ, then the sample node i is a low confidence node. τ is a pre-set confidence threshold, 0<τ<1, and the set of low confidence samples is

[0038] In this step, a low-confidence sample set is obtained, which can provide support for fully mining the association relationships between accounts in the social network, thereby realizing automatic detection of malicious accounts in the social network and mining of malicious account groups.

[0039] S106: Under each model, for each low-confidence sample, the predicted label of the low-confidence sample is corrected by voting based on the predicted labels of its neighboring non-low-confidence samples;

[0040] Specifically, for low confidence sample nodes Since the classification accuracy of low-confidence samples is low, the reliability of the predicted label corresponding to the sample node u is poor. Therefore, the predicted label of the low-confidence sample node should be regenerated. Figure 2 As shown, the predicted label voting correction of the non-low confidence samples in the neighborhood of the low confidence sample is adopted. As an implementation method, for the model GNN1, the predicted label of the sample node u is obtained by voting the predicted labels of the non-low confidence samples in the m-order neighborhood (that is, the majority result is taken as the final result). For simplicity, Figure 2 In the example, m is set to 1. Similarly, we can get the model GNN2 to GNN M Output correction result.

[0041] S107: For each low-confidence sample, aggregate the predicted labels of the M models after correction of the confidence sample to obtain the final predicted label as the pseudo label of the confidence sample.

[0042] Specifically, the aggregation method in this embodiment is to use the hard labels predicted by each model to vote.

[0043] The present invention utilizes the various relationships between accounts in the social media platform, uses different relationships and training samples to train different GNN models, and determines samples with different classification results from different models as low-confidence samples. The predicted labels of non-low-confidence samples in the neighborhood of the low-confidence samples are voted and corrected. Compared with previous graph neural networks, the method proposed in the present invention effectively utilizes the various interactive relationships between accounts, enables the multiple relationships to supervise and complement each other, and screens out low-confidence samples that are more likely to be misclassified. And by mining the structure of the social network, the output of low-confidence (more likely to be misclassified) samples is corrected, which improves the prediction accuracy of the pseudo-labels of low-confidence samples, thereby improving the accuracy of malicious account detection in the subsequent process of malicious account detection based on the pseudo-label data.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A pseudo-label generation method for low-confidence samples based on social media interaction relationships, characterized by: include: Step 1: Use the M types of interaction relationships existing in social media to construct M different social network account interaction relationship topology graphs; Step 2: Use M different social network account interaction relationship topology graphs to train M graph neural networks respectively to obtain M models; Step 3: Use M models to obtain the predicted labels of the test set sample nodes; Step 4: For each sample node, calculate the consistency between the M predicted labels of the sample node; Step 5: If the consistency value corresponding to the sample node is lower than the set confidence threshold, the sample node is a low-confidence sample; Step 6: Under each model, for each low-confidence sample, vote to correct the predicted label of the low-confidence sample based on the predicted labels of its neighboring non-low-confidence samples; Step 7: For each low-confidence sample, aggregate the predicted labels of the M models after correction of the confidence sample to obtain the final predicted label as the pseudo label of the confidence sample.

2. The method for generating pseudo labels for low-confidence samples based on social media interaction relationships according to claim 1, characterized in that: In step 2, the M graph neural networks adopt different model types; and / or, the M graph neural networks adopt different hyperparameters; and / or, use different training sample sets.

3. The method for generating pseudo labels for low-confidence samples based on social media interaction relationships according to claim 1, characterized in that: Step 4 specifically includes: Step 4.1: For each sample node i, obtain the predicted label vector h of M models for the sample node i i ; Wherein, the predicted label vector h i The kth dimension h in i,k Indicates the number of models that judge sample node i as the kth category; Step 4.2: Follow the formula Calculate the consistency C between the M predicted labels of the sample node i i .

4. The method for generating pseudo labels for low-confidence samples based on social media interaction relationships according to claim 1, characterized in that: In step 7, the predicted labels of the confidence samples corrected by the M models are aggregated, specifically including: voting using the hard labels predicted by each model.

Images