Social network depression radical co-occurrence individual detection method based on mixed time sequence fine-grained emotion and aggressive outburst

Abstract

The application provides a social network depression radical co-occurrence individual detection method based on mixed time sequence fine-grained emotion and aggressive outbreak, focuses on online social network text, excavates psychological characteristics of language expression of depression users from the perspective of pure text, fully considers language characteristics, can more effectively identify potential depression patients, selects a new perspective of depression radical co-occurrence to carry out research, realizes accurate identification of target users by fusing fine-grained emotion analysis technology with time sequence characteristics and an aggressive outbreak detection model, and solves the early identification problem of depression radical groups. The method provided by the application not only can promote accurate identification of depression-companied behaviors, but also can provide key data support for relevant departments, so as to reduce the risk of harming others and society possibly caused by the specific group.

Description

Technical Field

[0001] This invention relates to the field of individual detection technology in social networks, specifically to a method for detecting co-occurrence of depressive and radical individuals in social networks based on mixed temporal fine-grained emotions and aggressive outbursts. Background Technology

[0002] Depression is the most common and serious mental health problem and a leading cause of physical disability worldwide. In recent years, the age of onset of depression has been trending younger, accompanied by an increase in aggressive outbursts, raising concerns about early identification and intervention. However, most existing research focuses on identifying patients with early-stage depression or those exhibiting self-aggressive behavior, lacking identification of depressed individuals with aggressive outbursts. Aggressive outbursts can have serious consequences, including but not limited to physical harm, psychological trauma, social disruption, economic loss, and social impact, having a profound and undeniable negative impact on the health, stability, and development of the attacked individuals, organizations, and society. Therefore, it is necessary and urgent to identify individuals with potential aggressive outburst traits among depressed patients as early as possible through social networks.

[0003] Currently, existing methods for identifying depressed patients through social networks include: using a multi-task learning scheme to train a deep neural network classification model FusionNet (FN), which identifies depressed patients by fusing ten statistical features extracted from user text, social behavior, and posted images with text word features. This method integrates posted images and various behaviors; however, the meaning of images becomes uncertain during network propagation, making it interference information for the model (Yiding W, Zhenyi W, Chenghao L, et al. Online social network individual depression detection using a multitask heterogenous modality fusion approach[J]. Information Sciences, 2022, 609: 727-749.); using a depression detection framework based on multivariate time series features of user depressive symptoms, extracting the variation features of user depressive symptoms in the form of multivariate time series is a relatively effective method for identifying depressed patients. However, this method only has good identification results for users with obvious depressive characteristics and cannot effectively identify depressed patients who have depressive psychology but have not posted obvious text about their illness on social networks. Summary of the Invention

[0004] To address the neglect of identifying depressed users with potential aggressive outbursts in existing research, this invention defines depressed users with aggressive outburst potential as depressive-aggressive co-occurring individuals. It proposes a social network depressive-aggressive co-occurring individual detection method, TFE-AOD (Mixed Temporal Fine-grained Emotions and Aggression Outburst Detection Method), which focuses on online social network text, mining the psychological characteristics of depressed users' language expressions from a pure textual perspective. By fully considering linguistic characteristics, it can more effectively identify potential depressed patients. Furthermore, it adopts a novel perspective on depressive-aggressive co-occurrence, achieving accurate identification of target users by integrating fine-grained emotion analysis technology incorporating temporal features with an aggression outburst detection model, thus solving the problem of early identification of depressive-aggressive groups.

[0005] The technical solution of this invention is as follows:

[0006] The method for detecting co-occurrence of depressive and radical individuals in social networks based on mixed temporal fine-grained emotions and aggressive outbursts includes the following steps:

[0007] Step 1: Extract text content from social networks that reveals user attributes; take a portion of the text content and manually set feature labels, including multivariate temporal emotion features, auxiliary features, and aggression features; obtain a first training dataset composed of user attributes, text content, and manually set feature labels; the user attributes include user ID, label values ​​for whether the user is a depressed user, and label values ​​for whether the user is an aggressive depressed user.

[0008] Step 2: Construct a fine-grained sentiment analysis model and train the model using user attributes, text content, and manually set multivariate temporal sentiment feature labels from the first training dataset.

[0009] An offensive feature analysis model is constructed, and the offensive feature analysis model is trained using user attributes, text content, and manually set offensive feature labels from the first training dataset.

[0010] The text content containing user attributes extracted from the social network in step 1 is input into the trained fine-grained sentiment analysis model, and the output is the probability value of the multi-dimensional temporal sentiment features corresponding to the text content.

[0011] The text content containing user attributes extracted from the social network in step 1 is input into the trained aggressive feature analysis model, and the aggressive feature labels corresponding to the text content are output.

[0012] Step 3: Construct a multivariate temporal depression identification model;

[0013] Using the probability values ​​of the multi-emotional features corresponding to the text content obtained in step 2 and the user attributes corresponding to the text content, calculate the emotional similarity between users. Based on the emotional similarity between users, establish a graph model input G = (V, E), where V is the set of user nodes and E is the set of edges. For each user node, there are edge connections between it and the top N user nodes sorted by emotional similarity from largest to smallest.

[0014] Input the graph model G, the probability values ​​of the multivariate temporal emotion features corresponding to the text content obtained in step 2, and the auxiliary features corresponding to the text content into the multivariate temporal depression recognition model for training; obtain the trained multivariate temporal depression recognition model; the model output is the feature value of whether the user is depressed.

[0015] Step 4: Construct an aggressive individual detection model incorporating aggressive features; the aggressive individual detection model incorporating aggressive features consists of an ERNIE network + an LSTM network + a fully connected layer;

[0016] Based on the user attributes in step 1, select the text content of depressed users and the corresponding aggressive feature labels as model inputs to train the radical individual detection model, and obtain the trained radical individual detection model; the model output is the feature value of whether the user is radical.

[0017] Step 5: Input the user text content obtained from the social network into the model trained in Step 3 and Step 4. If the model output is 1 for both, it means that the user is a depressive-radical co-occurring individual.

[0018] Furthermore, in step 1, the multi-temporal emotional features are nine emotional features in total: anger, disgust, fear, happiness, sadness, surprise, shame, confusion, and irony.

[0019] Furthermore, in step 1, the auxiliary feature is nighttime activity level.

[0020] Furthermore, in step 1, the aggressive characteristics are four types: alcoholism, drug dependence, hidden hatred, and exposure of violence.

[0021] Furthermore, in step 2, the fine-grained sentiment analysis model consists of a Roberta network and a multi-head attention classification network; the text content in the first training dataset is processed by word segmentation and encoding, combined with user attributes and manually set multi-dimensional temporal sentiment feature labels, and then input into the fine-grained sentiment analysis model for training, finally obtaining the trained fine-grained sentiment analysis model.

[0022] Furthermore, in step 2, the aggressive feature analysis model consists of a Roberta network and a multi-head attention classification network; the text content in the first training dataset is processed by word segmentation and encoding, combined with user attributes and manually set aggressive feature labels, and then input into the aggressive feature analysis model for training, finally obtaining the trained aggressive feature analysis model.

[0023] Furthermore, in step 2, the NSP prediction task module described later is removed from the Roberta network.

[0024] Furthermore, in step 3, the multivariate temporal depression recognition model is obtained by an MTGNN network plus two fully connected layers.

[0025] Beneficial effects

[0026] The proposed method for detecting co-occurrence of depression and radicalism in social networks, based on hybrid temporal fine-grained emotion and aggression outbreaks, mainly achieves accurate identification of target users by fusing fine-grained emotion analysis techniques incorporating temporal features with an aggression outbreak detection model.

[0027] First, this invention introduces the irony characteristic of Chinese, refining emotions into nine categories. It performs fine-grained sentiment analysis on individual users' posts over a timeline, incorporating a multi-head attention mechanism to assign different weights to each emotion to better capture subtle individual emotional changes. This consideration of long-term emotional trends in depressed users effectively avoids the problem of generalizing from specific moments' emotions to represent the overall user experience, while also enabling more accurate capture of emotional changes, thus contributing to a better characterization of depressed users.

[0028] Secondly, this invention proposes using nighttime activity levels to characterize the higher urge to confide in depressed individuals compared to non-depressed individuals during the night. It also integrates time series data and fine-grained emotion data to capture the temporal dependence of depressed individuals, thereby identifying those with depression who exhibit long-term cumulative effects of depressive symptom representation. At this stage, the invention borrows ideas from the Spatiotemporal Graph Neural Network (MTGNN) for identifying depressed users. This approach not only considers the effects of long-term time series data but also preserves the spatial dependencies between individual users and key features, enhancing the characteristics of depressed individuals from both temporal and spatial perspectives.

[0029] Finally, through in-depth analysis of individual users' statements and behaviors, four typical characteristics of aggressive outbursts (alcoholism, drug dependence, hidden hatred, and violence exposure) were summarized and refined. These four characteristics were then automatically analyzed and calculated to identify signs of radical behavior, and the final depressive-radical co-occurrence individuals were obtained based on the classifier.

[0030] The method proposed in this invention can not only facilitate the accurate identification of comorbid behaviors of depression, but also provide key data support for relevant departments, thereby reducing the risk of this specific group potentially harming others and society.

[0031] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0032] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0033] Figure 1 : A schematic diagram of the principle of this invention;

[0034] Figure 2 Method: TFE-AOD training process. Detailed Implementation

[0035] The embodiments of the present invention are described in detail below. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0036] The proposed method for detecting co-occurring depressive and aggressive individuals in social networks, based on hybrid temporal fine-grained emotion and aggression bursts, achieves accurate identification of target users by fusing fine-grained emotion analysis techniques incorporating temporal features with an aggression burst detection model. It comprises three main parts: fine-grained emotion analysis incorporating irony, multivariate temporal depression identification incorporating nighttime activity, and aggressive individual detection incorporating aggression features. The model ultimately outputs co-occurring depressive and aggressive individuals; explanations of each stage will follow in subsequent steps. Table 1 summarizes commonly used symbols in this embodiment for clarity.

[0037] Table 1. Symbol Explanation

[0038]

[0039]

[0040] like Figure 1 As shown, the specific steps of this embodiment are as follows:

[0041] Step 1: Extract text content from social networks that reveals user attributes; take a portion of this text content and manually set feature labels, including multivariate temporal emotion features, auxiliary features, and aggression features; obtain a first training dataset composed of user attributes, text content, and manually set feature labels; the user attributes include user ID, label values ​​indicating whether the user is a depressed user, and label values ​​indicating whether the user is an aggressive depressed user. The multivariate temporal emotion features are nine emotion features: anger, disgust, fear, happiness, sadness, surprise, shame, confusion, and irony. The auxiliary feature is nighttime activity level. The aggression features are four aggression features: alcoholism, drug dependence, hidden hatred, and exposure to violence.

[0042] This invention has two key entity types, namely user sets U = u1, u2, u3, ..., u M and content set C = c1, c2, c3, ..., c P For each user u, e (k) Defined as text content c p The emotions expressed Let a represent the k-th emotion expressed in the text posted by user i at time t. (l) Defined as text content c p The aggressive characteristics it exhibits The l-th aggressive characteristic manifested in the text posted by user i. To further represent the user's emotional changes over time, Z is defined. i,t The multivariate temporal sentiment characteristics of user i at time point t are expressed as follows: Where K = 9. External auxiliary features of the data, defined as S. i,t ∈R N×D To exclude the emotional features Z of user i i,t In addition to the above, auxiliary features include nocturnal activity and aggressive characteristics such as alcoholism, drug dependence, hidden hatred, and violent exposure, where N represents a multivariate variable of dimension N, and B is the feature dimension.

[0043] Step 2: Construct a fine-grained sentiment analysis model and train the model using user attributes, text content, and manually set multivariate temporal sentiment feature labels from the first training dataset.

[0044] The fine-grained sentiment analysis model consists of a Roberta network and a multi-head attention classification network. The text content in the first training dataset is processed by word segmentation and encoding, combined with user attributes and manually set multi-dimensional temporal sentiment feature labels, and then input into the fine-grained sentiment analysis model for training, finally obtaining the trained fine-grained sentiment analysis model.

[0045] An offensive feature analysis model is constructed, and the offensive feature analysis model is trained using user attributes, text content, and manually set offensive feature labels from the first training dataset.

[0046] The offensive feature analysis model consists of a Roberta network and a multi-head attention classification network. The text content in the first training dataset is processed by word segmentation and encoding, combined with user attributes and manually set offensive feature labels, and then input into the offensive feature analysis model for training, finally obtaining the trained offensive feature analysis model.

[0047] Furthermore, in order to improve model performance, the context prediction (NSP) task module in the Roberta network was removed, and the multi-head attention classification network was further combined with Roberta.

[0048] The text content containing user attributes extracted from the social network in step 1 is input into the trained fine-grained sentiment analysis model, and the output is the probability value of the multi-dimensional temporal sentiment features corresponding to the text content.

[0049] The text content containing user attributes extracted from the social network in step 1 is input into the trained aggressive feature analysis model, which outputs the aggressive feature labels corresponding to the text content.

[0050] In the initial in-depth analysis of the texts of depressed users, it was found that their statements contained a variety of typical emotions, including frustration, helplessness, fear, and anger. Furthermore, considering the frequent use of irony and satire in Chinese by depressed users, introducing irony as a textual expression of emotion can better detect and identify depressed users. Combining the above factors with Ekman's basic emotion theory, this invention extracts and summarizes nine fine-grained emotions: anger, disgust, fear, happiness, sadness, surprise, shame, confusion, and irony. Compared to simple emotion classification, fine-grained emotion detection can better identify and differentiate the negative emotions of depressed patients, such as sadness, despair, and pain, thereby improving the accuracy of depression detection.

[0051] The text content C, from which user attributes are known, extracted from the social network in step 1, is converted into computer-recognizable Token(C) bytes using TokenEmbedding and PositionEmbedding. The TokenEmbedding is Roberta's own publicly available training vector. The PositionEmbedding is a calculation function, derived using the following sine and cosine formulas. These two formulas are used interchangeably during encoding to encode the user's sentiment text, where pos represents the word position and d... modelis the dimension of the embedded vector.

[0052] Token(C) = TokenEmbedding(C) i )+PositionEmbedding(C i )

[0053]

[0054] Input the obtained Token (C) into Roberta's Transformer. (l) In the structure and multi-head attention network, the text content C is obtained to correspond to different emotional features. The result value S of the attack feature i,t ∈R N×D .

[0055] The Transformer (l) The architecture consists of self-attention, a residual network, a normalized layer, and a feedforward neural network. The residual connection network Concat has multiple attention heads. H and a weight parameter W i V LayerNorm merges the residual join results into a single layer, resulting in MHA(X). (l-1) The intermediate quantity is obtained by merging. The fully connected layer FFN has its W1 weights and b1 bias parameters. Output results The text data is then processed by LayerNorm, and the process is repeated in this loop.

[0056] X = Token(C)

[0057] X (l) =Transformer (l) (X (l-1) for l = 1, ..., L

[0058] head i =Attention(X) (l-1) W i Q ,X (l-1) W i K ,X (l-1) W i V )

[0059] MHA(X (l-1) = Concat(head1,...,head) H W O

[0060]

[0061]

[0062]

[0063] The multi-head attention network consists of three parts: an attention layer, a concatenation layer, and a fully connected layer. The input to this part is the Transformer. (l) The output of the structure is the desired result. X (l) For the Transformer (l) The structure output is processed by the FC_head function to obtain the multi-head attention result, which is then concatenated by the MHA_output and input into the linear layer. When outputting the FC_output, W is introduced. fc Weighting indicators and b fc Bias, these two data points are model parameters in the FC (Fully Connected) layer; after passing through the FC layer, the result is classified, where... Belongs to category e i The probability of emotions.

[0064] FC_head i =Attention(X) (L) W i Q ,X (L) W i K ,X (L) W i V )

[0065] MHA_output=Concat(head1,...,head9)W O

[0066] FC_output = MHA_outputW fc +b fc

[0067]

[0068] Step 3: Construct a multivariate temporal depression identification model that incorporates nighttime activity levels. The multivariate temporal depression identification model that incorporates nighttime activity levels is obtained by an MTGNN network plus two fully connected layers.

[0069] Using the probability values ​​of the multi-emotional features corresponding to the text content obtained in step 2, and the user attributes corresponding to the text content, calculate the emotional similarity between users. The cosine similarity calculation formula can be used. Based on the emotional similarity between users, establish a graph model input G = (V, E), where V is the set of user nodes and E is the set of edges. For each user node, there are edges connecting it to the top N user nodes sorted by emotional similarity from largest to smallest.

[0070] Input the graph model G, the probability values ​​of the multivariate temporal emotion features corresponding to the text content obtained in step 2, and the auxiliary features corresponding to the text content into the multivariate temporal depression recognition model for training; obtain the trained multivariate temporal depression recognition model; the model output is the feature value of whether the user is depressed.

[0071] The MTGNN network consists of graph learning layers, graph convolutions, and temporal convolutions. The graph convolution and temporal convolution modules work alternately. The temporal convolution module filters the input by sliding a one-dimensional window along the time axis and node axis, while the graph convolution module filters the input at each step.

[0072] The graph learning layer first receives multi-dimensional temporal emotion features representing user i at time node t. The emotional similarity δ(u) between nodes is obtained. i ,u j Finally, the adjacency matrix A is obtained by processing the ReLU and tanh activation functions.

[0073] The graph convolution module consists of two hybrid skip propagation layers, with the input being the adjacency matrix A obtained from the graph learning layer. The outputs of the previous graph's convolutional module and temporal convolutional module serve as intermediate values ​​during model training. The working mechanism of the hybrid skip propagation layer is to first propagate information horizontally, and then select information vertically. The information propagation process is defined as follows:

[0074]

[0075] Where q is the propagation depth, and H in This represents the input hidden state of the output of the previous layer. β is a hyperparameter used to preserve a certain proportion of the original node states during propagation. And H... (0) =H in , as well as Considering the issues of node information loss and useless noise, information selection needs to be introduced to filter out important information at each step. The information selection process is defined as follows:

[0076]

[0077] Information propagation recursively propagates node information and the given graph structure. Where H... out W represents the output hiding state of the current layer. (q) This is the parameter matrix that acts as the feature selector.

[0078] The temporal convolution module consists of two dilated initial layers, with the adjacency matrix A obtained from the graph learning layer as input. The outputs of the previous convolutional module and the temporal convolutional module serve as intermediate values ​​during model training. The dilated initial layer includes a dilated convolutional inception structure, a tanh activation function, and a sigmoid activation function. The dilated convolution can increase the receptive field (RF) without increasing parameters. The inception structure captures information at different scales by using convolutional kernels of different sizes. The calculation method for the receptive field differs from that of regular convolution, and is as follows:

[0079]

[0080] Among them, RF n This represents the receptive field size of the output feature map of the current layer, RF. n-1 This represents the receptive field size of the feature map in the previous layer, where s is the stride. This represents the product of all step sizes from the first layer to the (n-1)th layer, where d is the dilation rate of the dilated convolution in the current layer, and ck is the kernel size of the current layer. For the first layer, RF0 is typically set to 1.

[0081] Step 4: Construct an aggressive individual detection model incorporating aggressive features; the aggressive individual detection model incorporating aggressive features consists of an ERNIE network + an LSTM network + a fully connected layer;

[0082] Based on the user attributes in step 1, select the text content of depressed users and the corresponding aggressive feature labels as model inputs to train the radical individual detection model, and obtain the trained radical individual detection model; the model output is the feature value of whether the user is radical.

[0083] The ERNIE network takes text content C as input and outputs S. i,t ∈R N×D , excluding user i's emotional characteristics Z i,t In addition to the characteristics (alcoholism, drug dependence, hidden hatred, and aggressive characteristics such as violence), where N represents a multivariate variable of dimension N, and B is the feature dimension; the output is the label value of whether the user is an aggressive user.

[0084] The ERNIE network incorporates knowledge graph content and uses a graph attention network (GAT) to process the content in the knowledge graph and the content in text C. The processing result is then fed into the same Transformer described in Roberta. (l) The block yields a corresponding vector representation. This vector representation is then input into a classification network to obtain the aggression feature result S. i,t ∈R N×D .

[0085] The knowledge graph contains entity and relation embeddings KG = (entity1, r 1,2 ,entity2),...,(entity i ,r i,j entity j The graph attention network GAT is an existing functional open module; the classification network is a multi-head attention network, which also uses the multi-head attention network described in step 2.

[0086] The output S of the ERNIE network i,t ∈R N×D Aggressive traits such as alcoholism, drug dependence, hidden hatred, and overt violence are input into a Long Short-Term Memory (LSTM) neural network, a fully connected layer, and an activation function to obtain a label value indicating whether the user is aggressive. The execution process of the LSTM neural network and the fully connected layer is shown in the following formula. E final The values ​​are the initial values ​​and the output of the ERNIE network. This classification process is represented by the Classify function, and after passing through an LSTM layer, W·Hlstm T +b fc The data represents the weights passed through a fully connected layer, where W is the weight parameter of the fully connected layer, and b... fc This represents the bias parameter of the fully connected layer. This data is processed by the Sigmoid function to obtain the final label value result {U non-rad U rad}

[0087] Hlstm0 = Classify(E final )

[0088] Hlstm t =LSTM(Logits) t Hlstm t-1 )

[0089] {U non-rad U rad} = Sigmoid(W·Hlstm) T +b fc )

[0090] Step 5: Input the user text content obtained from the social network into the model trained in Step 3 and Step 4. If the model output is 1 for both, it means that the user is a depressive-radical co-occurring individual.

[0091] Here is a specific example:

[0092] This study utilized two datasets: the publicly available Weibo dataset for depression (SWDD) and a self-constructed Douban dataset (DBDD), both in Chinese. SWDD is a large, publicly released Weibo dataset for depression containing complete user tweet histories. It collected user profiles and complete historical tweet records under the Weibo hashtag "#depression" up to June 4, 2021, including 3,711 users with depression and 19,526 users without depression. DBDD is the Douban depression dataset constructed for this study, collecting user and user broadcast data from "depression" and "happiness" related groups on Douban. This dataset collected data from 727 users with more than 100 posts up to November 20, 2023, including personal information and broadcast information from 618 users with depression and 109 normal users.

[0093] Each data point has 13 label items, including 9 emotions (anger, disgust, fear, happiness, sadness, surprise, shame, confusion, and irony) and 4 aggressive traits (alcoholism, drug dependence, hidden hatred, and exposure to violence). The amount of data used is shown in Table 2.

[0094] Table 2 shows the data volume.

[0095]

[0096] Training a fine-grained sentiment analysis model that incorporates irony:

[0097] The pre-trained model chinese-roberta-wwm-ext was loaded and trained for 20 epochs on all texts in the dataset. A warm-up step of 500 and a weight decay coefficient of 0.01 were set to prevent overfitting during training. To ensure numerical stability, a loss function combining binary cross-entropy loss and logit transformation was used.

[0098] Training a multivariate temporal depression identification model incorporating nighttime activity levels:

[0099] The constructed model was trained using the SWDD and DBDD datasets, combined with the attached... Figure 2Analysis shows the decline of the model's loss function. Initially, the loss decreases rapidly, indicating that the model quickly improves its predictive ability in the early stages of learning. On the SWDD dataset, the rate of loss decline begins to slow after approximately 25 epochs and gradually stabilizes during subsequent training. On the DBDD dataset, due to its relatively small data volume, the model only stabilizes after approximately 250 epochs. These results suggest that after sufficient training rounds, the model converges to a lower loss value, indicating that the model's fit to the data has become relatively stable. The performance results for the depression user identification stage are shown in Table 3.

[0100] Table 3. Indicators for identifying depressed users

[0101]

[0102] Training a radical individual detection model that incorporates aggressive features:

[0103] Based on the obtained set of depressed users, radical depressed users were identified, and the results are shown in Table 4. The results demonstrate that the TFE-AOD model performs superiorly in detecting co-occurring individuals with radical depressed users, and can predict and classify them more accurately, proving the reliability of our proposed method. Furthermore, the high recall and precision of the TFE-AOD method indicate its high sensitivity and accuracy in identifying relevant instances, which is crucial in practical applications.

[0104] Table 4. Identification indicators of radical depression

[0105]

[0106] Method-specific ablation analysis:

[0107] The TFE-AOD method enhances the accuracy of identifying individuals with co-occurring depression and radicalism by applying graph structures to the text content set and considering external features (nighttime activity). To evaluate the effectiveness of these two components, this study performed ablation analysis by removing the corresponding components. Table 5 shows the results of the ablation analysis, indicating that the introduction of graph structures has a slight impact on model performance. When graph structures are removed, the model's accuracy, precision, recall, and F1 score all decrease. Furthermore, when nighttime activity is not considered, all performance metrics decrease, showing a more significant performance decline, with accuracy dropping to 0.9075 on the SWDD dataset and 0.7917 on the DBDD dataset.

[0108] Table 6 Ablation Analysis Results

[0109]

[0110] Identification of individuals with co-occurrence of depression and radicalism:

[0111] The data to be tested is fed into a trained model, which outputs predicted labels for individual users. Users with a label of 1 are the target users, i.e., individuals exhibiting both depression and aggression. Furthermore, individuals with a label of 1 indicating depression can be identified during the model's intermediate processing. In this scenario, the model effectively identifies subtle signals characterizing depression and aggressive outbursts, and effectively distinguishes individuals exhibiting both depression and aggression from ordinary users.

[0112] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A method for detecting co-occurrence of depressive and radical individuals in social networks based on mixed temporal fine-grained emotion and aggressive outbursts, characterized in that: Includes the following steps: Step 1: Extract text content from social networks that reveals user attributes; take a portion of the text content and manually set feature labels, including multivariate temporal emotion features, auxiliary features, and aggression features; obtain a first training dataset composed of user attributes, text content, and manually set feature labels; the user attributes include user ID, label values ​​for whether the user is a depressed user, and label values ​​for whether the user is an aggressive depressed user. Step 2: Construct a fine-grained sentiment analysis model and train the model using user attributes, text content, and manually set multivariate temporal sentiment feature labels from the first training dataset. An offensive feature analysis model is constructed, and the offensive feature analysis model is trained using user attributes, text content, and manually set offensive feature labels from the first training dataset. The text content containing user attributes extracted from the social network in step 1 is input into the trained fine-grained sentiment analysis model, and the output is the probability value of the multi-dimensional temporal sentiment features corresponding to the text content. The text content containing user attributes extracted from the social network in step 1 is input into the trained aggressive feature analysis model, and the aggressive feature labels corresponding to the text content are output. Step 3: Construct a multivariate temporal depression identification model; Using the probability values ​​of the multi-emotional features corresponding to the text content obtained in step 2 and the user attributes corresponding to the text content, calculate the emotional similarity between users. Based on the emotional similarity between users, establish a graph model input G = (V, E), where V is the set of user nodes and E is the set of edges. For each user node, there are edge connections between it and the top N user nodes sorted by emotional similarity from largest to smallest. Input the graph model G, the probability values ​​of the multivariate temporal emotion features corresponding to the text content obtained in step 2, and the auxiliary features corresponding to the text content into the multivariate temporal depression recognition model for training; obtain the trained multivariate temporal depression recognition model; the model output is the feature value of whether the user is depressed. Step 4: Construct an aggressive individual detection model incorporating aggressive features; the aggressive individual detection model incorporating aggressive features consists of an ERNIE network + an LSTM network + a fully connected layer; Based on the user attributes in step 1, select the text content of depressed users and the corresponding aggressive feature labels as model inputs to train the radical individual detection model, and obtain the trained radical individual detection model; the model output is the feature value of whether the user is radical. Step 5: Input the user text content obtained from the social network into the model trained in Step 3 and Step 4. If the model output is 1 for both, it means that the user is a depressive-radical co-occurring individual.

2. The method for detecting co-occurrence of depressive and radical individuals in social networks based on mixed temporal fine-grained emotion and aggressive outbursts, as described in claim 1, is characterized in that: In step 1, the multi-temporal emotional features are nine emotional features: anger, disgust, fear, happiness, sadness, surprise, shame, confusion, and irony.

3. The method for detecting co-occurrence of depressive and radical individuals in social networks based on mixed temporal fine-grained emotion and aggressive outbursts, as described in claim 1, is characterized in that: In step 1, the auxiliary feature is nighttime activity.

4. The method for detecting co-occurrence of depressive and radical individuals in social networks based on mixed temporal fine-grained emotion and aggressive outbursts, as described in claim 1, is characterized in that: In step 1, the aggressive characteristics are four types: alcoholism, drug dependence, hidden hatred, and exposure of violence.

5. The method for detecting co-occurrence of depressive and radical individuals in social networks based on mixed temporal fine-grained emotion and aggressive outbursts, as described in claim 1, is characterized in that: In step 2, the fine-grained sentiment analysis model consists of a Roberta network and a multi-head attention classification network. The text content in the first training dataset is processed by word segmentation and encoding, and then combined with user attributes and manually set multivariate temporal sentiment feature labels, and input into the fine-grained sentiment analysis model for training, finally obtaining the trained fine-grained sentiment analysis model.

6. The method for detecting co-occurrence of depressive and radical individuals in social networks based on mixed temporal fine-grained emotion and aggressive outbursts, as described in claim 1, is characterized in that: In step 2, the offensive feature analysis model consists of a Roberta network and a multi-head attention classification network. The text content in the first training dataset is processed by word segmentation and encoding, combined with user attributes and manually set offensive feature labels, and then input into the offensive feature analysis model for training, finally obtaining the trained offensive feature analysis model.

7. The method for detecting co-occurrence of depressive and radical individuals in social networks based on mixed temporal fine-grained emotion and aggressive outbursts, as described in claim 5 or 6, is characterized in that: In step 2, the NSP prediction task module was removed from the Roberta network.

8. The method for detecting co-occurrence of depressive and radical individuals in social networks based on mixed temporal fine-grained emotion and aggressive outbursts, as described in claim 1, is characterized in that: In step 3, the multivariate temporal depression recognition model is obtained by an MTGNN network plus two fully connected layers.

Images