Multi-expert collaborative network social group characterization method based on modal dynamic fusion

Abstract

The application belongs to the field of network social group analysis, and provides a multi-expert collaborative network social group characterization method based on modal dynamic fusion. First, the multi-modal original data of the network social group is input, and the modal representation adaptive extraction module is used to capture the private feature representation of the multi-modal. Then, through the modal complementary information fusion module, the bidirectional state space model is used to realize the deep fusion of cross-modal information, and a unified group multi-modal fusion representation is generated. Next, the KAN multi-expert network architecture is introduced through the multi-expert collaborative prediction module, different experts are used to adapt to the diversified group characteristic modes such as mainstream, small, temporary, etc., and the dynamic weight is calculated through the gate network to realize the accurate fusion of the expert output. Finally, the long tail boundary perception loss function is combined to optimize the model training, and the multi-dimensional feature characterization result of the group is output. The application can obtain more comprehensive and accurate group feature representation, and improve the characterization performance of the network social group in complex scenes.

Description

Technical Field

[0001] This invention belongs to the field of online social network analysis and relates to a method for characterizing multi-expert collaborative network social groups based on modality dynamic fusion, which can be used in online social network-related analysis, governance and service scenarios. Background Technology

[0002] Online social groups, carrying rich behavioral, relational, and semantic information on online social platforms, have long been important subjects for understanding user interaction patterns, information dissemination laws, and collective decision-making processes. Social network groups not only reflect the interactive relationships between members but also exhibit complex characteristics such as information commonality, topic convergence, and behavioral patterns. These characteristics are significant for tasks such as social influence assessment and user modeling. Unlike traditional static relationship analysis, online social groups often exhibit heterogeneous node types, multi-layered behavioral modalities, and dynamic evolutionary attributes, making it difficult for a single paradigm to fully characterize their complexity. Furthermore, with the explosive growth of social platform data, the interaction relationships between multimodal data (such as behavioral logs, textual semantics, and social relationships) pose new challenges to a deeper understanding of social groups. Currently, characterizing online social groups has become an important research direction in the fields of social computing, network analysis, and graph learning, urgently requiring effective modeling methods to reveal the high-order interaction relationships of multimodal information within groups.

[0003] Current mainstream algorithms for characterizing and representing online social groups can be categorized as follows: Group representation algorithms based on heterogeneous graph neural networks. These methods construct heterogeneous graph models to mine group features based on the diversity of node types (e.g., users, groups, topics) in social platforms. Fu et al.'s 2020 work, "MAGNN: Metapath Aggregated Graph Neural Network for Heterogeneous Graph Embedding," published on WWW, proposed the MAGNN algorithm. This method considers the intermediate node information within the metapath in the heterogeneous graph and learns node and group embeddings containing rich semantics through a multi-channel aggregation mechanism, improving the ability to model complex social relationships. Fan et al.'s 2019 work, "Metapath-guided Heterogeneous Graph Neural Network for Intent Recommendation," published on KDD, proposed MEIREC, which utilizes metapaths to guide the aggregation process of neighbor nodes, effectively capturing complex interaction intentions. In addition, although He et al.'s work "LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation" published at SIGIR in 2020 mainly focuses on recommendation scenarios, the simplified graph convolution idea they proposed has been widely used in group structure modeling. By removing linear transformations and nonlinear activations, it greatly reduces computational complexity and improves the characterization effect based on structural information.

[0004] Social network analysis algorithms based on multimodal fusion. To address the limitations of single-structure information, these methods attempt to fuse text, image, and behavioral data to enhance characterization capabilities. Wei et al.'s 2019 work published in MM, "MMGCN: Multi-modal Graph Convolution Network for Personalized Recommendation of Micro-video," proposed MMGCN. This method constructs independent graph structures for each type of user data (e.g., text, visual) for convolutional learning, and finally fuses multimodal features through a merging operation, achieving refined modeling of user interests. Yu et al.'s 2021 work published in WWW, "Self-Supervised Multi-Channel Hypergraph Convolutional Network for Social Recommendation," proposed SEPT. This method utilizes a hypergraph to model high-order relationships among multiple users and introduces a self-supervised learning mechanism to enhance data representation capabilities under multimodal views, achieving progress in mitigating data sparsity.

[0005] Although current data-driven methods for characterizing online social groups have achieved good performance, there are still two fundamental challenges when dealing with real-world social network data: (1) The multimodal descriptions of social network groups (such as speech text, interaction values, social relationships, and behavioral sequences) have essential differences in data structure, statistical characteristics, and semantics. This heterogeneity leads to semantic barriers between modalities, making it difficult for models to align and associate information from different sources in a unified semantic space. More challenging is that the information carried by different modalities may conflict or be redundant in representing group characteristics. For example, the emotional (positive) sentiment of a group's speech text may contradict the downward trend (numerical) of interaction frequency. Many existing fusion strategies are unable to intelligently identify and reconcile this conflict, either causing information to be diluted or canceled during the fusion process or introducing noise, thus limiting the model's ability to form a comprehensive and consistent understanding of group characteristics. (2) The distribution of characteristics of online social groups does not follow a single rule, but is the result of the combined effect of multiple potential patterns; at the same time, a large number of samples are concentrated in mainstream groups (such as interest communities and strong relationship groups), while rare samples such as niche groups and temporary event groups are scattered at the tail. Existing characterization models, especially single deep neural networks, tend to fit dominant distribution patterns, resulting in insufficient accuracy in characterizing rarer groups at the tail end. Furthermore, their fixed model structures struggle to flexibly adapt to and capture the differentiated characteristics of different group types, thus creating bottlenecks in improving overall accuracy and robustness. Summary of the Invention

[0006] To address the problems existing in the prior art, the purpose of this invention is to solve the suboptimal performance problem of group characterization caused by semantic barriers, information conflicts, and insufficient characterization of rare groups in previous studies by introducing a modality dynamic fusion mechanism and a multi-expert collaborative architecture. This invention proposes a social group characterization method based on modality dynamic fusion and a multi-expert collaborative network, which can accurately capture cross-modal long-distance dependencies, intelligently reconcile information conflicts, and adapt to the differentiated feature patterns of different types of groups through a sparse gating mechanism. Finally, it learns an effective multi-dimensional feature representation of the group for subsequent accurate characterization tasks.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A multi-expert collaborative network social group characterization method based on modality dynamic fusion is proposed. First, an adaptive modality representation extraction module, utilizing channel attention and multi-head self-attention mechanisms, extracts high-quality, highly discriminative private representations from the original data (speech text, interaction values, social relationships, etc.) of each modality in online social networks. Second, a modality complementary information fusion module employs a bidirectional state-space model to perform forward and backward global depth scans on the concatenated multimodal representations, explicitly modeling complex cross-modal interactions and generating a globally aware unified fusion representation. Finally, a multi-expert collaborative prediction module integrates multiple Kolmogorov-Arnold Networks (KANs) with diverse structures and specialized functions, and introduces sparse gating network dynamic routing to achieve specialized processing and collaborative decision-making for group samples with different feature patterns. The entire model is trained end-to-end, and by introducing a long-tail boundary-aware loss function, it effectively addresses the long-tail characteristics of network social group distribution, improving the model's robustness in characterizing niche, temporary, and other scarce group samples. The steps are as follows: Step 1: Adaptive extraction of modal representations to generate highly discriminative private representations for each modality; Step 1.1: Input the raw multimodal data of the online social network, including modal data such as speech text, group interaction values, member social relationships, and behavioral time series. Let the first part be the modal data. The modal data matrix is ,in For sample length, This represents the feature dimension of the modality. Statistical information for each channel is extracted using global average pooling. in, Indicates the first The first modality One channel descriptor; For the first In modal data, the first The sample at the th The values ​​on each feature channel; For the first The total number of feature channels for modal data.

[0008] Step 1.2: Feed the compressed channel descriptors into a bottleneck structure consisting of two fully connected layers to explicitly model the nonlinear correlation between channels and automatically generate channel importance vectors. : in, and This represents the learnable parameters of the two fully connected layers. and This represents the activation function. Channel importance is used as a gating signal, multiplied channel-by-channel with the input data to obtain the recalibrated data matrix. : Step 1.3: Utilize multi-head self-attention mechanism to capture Deep contextual dependencies between internal features. Utilizing three independent learnable linear transformations. ,Will Mapped to queries respectively ,key Sum In three different subspaces: Step 1.4: Calculate the attention score using the scaled dot product. To quantify the relationships between different features within the same modality: In the formula, This is the scaling factor. Multiple independent attention heads are used in parallel, each performing the above attention calculation. The outputs of all heads are concatenated and fused through a linear projection layer to obtain the multi-head attention composite output. In the formula, These are the learnable parameters of the linear projection layer; This represents the concatenation function.

[0009] Step 1.5: Generate the final modal private representation through residual connections and feedforward networks. : In the formula, This represents a feedforward network.

[0010] Step 2: Modal complementary information fusion to construct a globally perceived unified fusion representation; Step 2.1: Concatenate the private representations of all modalities along the feature dimension using a concatenation function to obtain a concatenated representation containing full modal information. : In the formula, This represents the total number of modes.

[0011] Step 2.2: To capture the global dependencies spanning different modalities within the spliced ​​representation, a fusion mechanism based on bidirectional state space scanning is adopted: In the formula, This represents the classic state-space model. This represents the inverse function. and The representation obtained by performing a depth scan of the concatenated representation from both the forward and backward dimensions.

[0012] Step 2.3: Generate the final multimodal fusion representation through a linear transformation layer. : In the formula, The learnable parameters of the linear layer. This bidirectional mechanism ensures that every feature point in the fusion process can fully perceive global contextual information from all modalities and all sequence positions, thereby explicitly modeling complex cross-modal interactions and ultimately generating a globally perceptive, deeply complementary unified fusion representation.

[0013] Step 3: Multi-expert collaborative prediction to output a multi-dimensional feature representation of the population; Step 3.1, given the first Multimodal fusion representation of a sample of online social groups , construct a containing A collection of predictors from specialized KAN experts. Each expert is configured with a different combination of basis functions to specifically learn a particular type of nonlinear mapping pattern for group features; the output of a single expert network is: In the formula, Indicates the first The first expert network generated the first Preliminary characteristics of the sample group are represented. Indicates the first Layer Learnable activation functions for each node. This indicates that the output of the previous function is used as the input of the next function for continuous calculation.

[0014] The activation function takes the form of a linear combination of basis functions and residual activation, and is specifically defined as follows: In the formula, It is a learnable activation function; The activation function is a fixed residual. This is an optional set of basis functions, and the different basis functions constitute the fundamental basis for the expert category division. and This is a learnable scaling factor.

[0015] Step 3.2: Design a dynamic routing network based on sparse gating mechanism. Generate the activation weight vector for each expert network. : In the formula, The first generation generated for expert networks The activation weight of each online social group; The weights are those of the gating network. Before operation Set a maximum weight value and set the rest to zero to ensure that computational efficiency is maintained while ensuring the model's expressive power.

[0016] Step 3.3: Based on the sparse weight vector, obtain the final multi-dimensional feature representation of the population through a convex combination of the outputs of each expert network. : In the formula, For the first The activation weights of an expert network.

[0017] Step 4: End-to-end training optimization, improving model robustness based on long-tail boundary-aware loss function; Step 4.1: Construct the long-tailed boundary-aware objective function It consists of feature matching loss and class marginal adjustment loss, and is mathematically expressed as follows: In the formula, This is the loss balancing hyperparameter; For feature matching loss, This is the category-adjusted loss.

[0018] Step 4.2: Calculate the feature matching loss and use the mean squared error to measure the difference between the predicted group features and the true features, ensuring global characterization accuracy. In the formula, , which represents the true multidimensional feature label of the i-th network social group sample.

[0019] Step 4.3: Calculate the category marginal adjustment loss, compress the decision space of the mainstream group, expand the decision space of the niche group (tail samples), and optimize the decision boundary: In the formula, Number of group types; For the first The number of effective samples in a group; For the sample From the The class was misclassified to the first Resistance of class For the first The first in the class The sample is predicted to belong to the nth sample. The original output of the class, For the first The first in the class The sample is predicted to belong to the nth sample. The original output of the class. Larger The value indicates greater resistance when predicting this category. Dynamically adjusting the decision boundary enhances the model's ability to capture complex population distributions. It means proportional to.

[0020] Step 4.4: Adopt an end-to-end training approach, backpropagate the loss value using the stochastic gradient descent algorithm, and iteratively update all trainable parameters of the model; set the number of training iterations until the loss function converges to obtain the optimal network social group characterization model, ensuring robustness in characterizing the tail-end rare group samples.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: By deeply integrating channel attention, multi-head self-attention, and a bidirectional state-space model, this invention breaks down the semantic barriers of multimodal data, achieving accurate modeling and information conflict reconciliation for cross-modal long-distance dependencies, thus solving the information dilution problem of traditional shallow fusion strategies. Relying on the KAN multi-expert collaborative architecture and sparse gating dynamic scheduling mechanism, it overcomes the limitations of single-model representation, efficiently adapting to multi-modal feature distributions such as mainstream, niche, and temporary event groups, significantly improving the characterization accuracy of rare tail groups. By dynamically optimizing the decision boundary using a long-tail boundary-aware loss function, it balances the learning weights of mainstream and rare groups, enhancing the model's robustness in complex distribution scenarios, and providing more comprehensive, accurate, and reliable technical support for practical scenarios such as social network governance. This invention effectively solves core problems such as multimodal data heterogeneity, group pattern diversity, and long-tail data distribution. Through deep fusion and adaptive adaptation mechanisms, it learns more comprehensive and accurate group feature representations, significantly improving the characterization performance of network social groups in complex scenarios. Attached Figure Description

[0022] Figure 1 This is the basic framework of the present invention. Detailed Implementation

[0023] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0024] like Figure 1 As shown, the specific steps of the multi-expert collaborative network social group characterization method based on modal dynamic fusion of the present invention are as follows: Step 1: Adaptive extraction of modal representations to generate highly discriminative private representations for each modality; Step 1.1: Input the raw multimodal data from the online social network, and extract the channel statistics through global average pooling: Step 1.2: Feed the compressed channel descriptors into a bottleneck structure consisting of two fully connected layers to explicitly model the nonlinear correlation between channels and automatically generate channel importance vectors. Then, the channel importance is used as a gating signal and multiplied channel by channel with the input data to obtain the recalibrated data matrix. .

[0025] Step 1.3: Utilize multi-head self-attention mechanism to capture The deep contextual dependencies between internal features. Mapped to queries respectively ,key Sum In three different subspaces.

[0026] Step 1.4: Calculate the attention score using the scaled dot product. Multiple independent attention heads are used in parallel, each performing the above attention calculation. The outputs of all heads are concatenated and fused through a linear projection layer to obtain the multi-head attention composite output. .

[0027] Step 1.5: Generate the final modal private representation through residual connections and feedforward networks. .

[0028] Step 2: Modal complementary information fusion to construct a globally perceived unified fusion representation; Step 2.1: Concatenate the private representations of all modalities along the feature dimension using a concatenation function to obtain a concatenated representation containing full modal information. .

[0029] Step 2.2: To capture the global dependencies spanning different modalities within the stitched representation, a fusion mechanism based on bidirectional state space scanning is adopted. Forward depth scanning yields... Backward depth scan obtained .

[0030] Step 2.3: Generate the final multimodal fusion representation through a linear transformation layer. .

[0031] Step 3: Multi-expert collaborative prediction to output a multi-dimensional feature representation of the population; Step 3.1, given the first Multimodal fusion representation of a sample of online social groups , construct a containing A collection of predictors from specialized KAN experts. .

[0032] Step 3.2: The activation function is a linear combination of the basis function and the residual activation. .

[0033] Step 3.3: Design a dynamic routing network based on sparse gating mechanism. Generate the activation weight vector for each expert network. .

[0034] Step 3.4: Based on the sparse weight vector, obtain the final multi-dimensional feature representation of the group through a convex combination of the outputs of each expert. .

[0035] Step 4: End-to-end training optimization, improving model robustness based on long-tail boundary-aware loss function.

[0036] Step 4.1: Construct the long-tailed boundary-aware objective function .

[0037] Step 4.2: Calculate the feature matching loss. .

[0038] Step 4.3: Calculate the category marginal adjustment loss .

[0039] Step 4.4: Using an end-to-end training approach, the loss value is backpropagated through the stochastic gradient descent algorithm to iteratively update all trainable parameters of the model; the number of training iterations is set until the loss function converges to obtain the optimal network social group characterization model, ensuring robustness in characterizing the rare tail group samples.

[0040] The algorithm flow of this invention is as follows: Input: Raw data from multimodal online social networks Sample size Number of training iterations Real group characteristics Balance parameters .

[0041] Output: Model parameters after training .

[0042] 1. Randomly initialize model parameters ; 2. When hour( (current iteration number); 3. Multimodal description of each online social group sample ; 4. Adaptive extraction of corresponding representations ; 5. Assemble the representations of each modality. ; 6. Obtaining fused representation through bidirectional state-space scanning ; 7. Calculate the output of each expert. ; 8. Calculate expert weights using a gating network. ; 10. Calculate the feature matching loss and category marginal adjustment loss ; 11. Backpropagation and parameter update using the stochastic gradient descent algorithm; Returns: Post-trained model parameters .

[0043] Based on the solution of this invention, the experimental analysis is as follows: This method uses publicly available group-related datasets collected from online social networks to construct a multimodal group feature dataset, and compares it with current mainstream online social group characterization methods to evaluate the effectiveness of this method.

[0044] (1) Introduction to the dataset The model of this invention was tested on two simulated datasets derived from public data, US-ERD and CRYPTO-MD, to compare their performance. The task was to characterize the multi-dimensional features of online social groups.

[0045] The detailed information of the dataset is shown in Table 1: Table 1. Statistical information of the dataset

[0046] The US-ERD dataset, a simulation set of multimodal features of online social groups, is constructed based on publicly available financial texts and financial data. It contains 1068 simulated group samples: the text modality corresponds to "social group-specific discourse texts", which are converted into 77-dimensional semantic feature vectors through TF-IDF and word embedding; the numerical modality corresponds to "group interaction behavior indicators", from which 13 core quantitative indicators are extracted; and the sample labels correspond to "group feature complexity categories", which are divided into 23 category intervals.

[0047] The CRYPTO-MD dataset, a simulation set of dynamic social group characteristics, is built based on cryptocurrency news and market data and contains 2,684 simulated group samples: the text modality corresponds to "event-related discussion text", which is converted into a 125-dimensional text representation using BERT; the numerical modality corresponds to "group dynamic behavior indicators", which includes 64-dimensional quantitative indicators; and the sample labels correspond to "temporary group feature types", which are divided into 46 category intervals.

[0048] In the experiment, the dataset was divided into a training set and a test set in a 7:3 ratio to simulate the sample to verify the method's ability to characterize the multimodal features of real social groups.

[0049] (2) Comparison of experimental results between this method and other mainstream methods Table 2 shows the comparison of experimental results of this method with other mainstream models in the field of network social group characterization. Among them, WOLSTM and LSR-IGRU are single-modal temporal characterization methods, based on Long Short-Term Memory networks and gated recurrent units respectively, utilizing only single-modal data to mine group temporal features; cD-DNN is a single-modal deep characterization method, based on deep neural networks to extract single-modal group static features; VMD-DNN is a multi-modal fusion method, fusing multi-scale temporal information through variational modality decomposition to model group dynamics; TMVC-EF, MVC-DD, and IMVC are advanced multi-modal characterization methods, employing feature-level fusion, decision-level fusion, and interaction-level fusion strategies respectively to address the heterogeneity problem of multi-modal group data. Comparative experiments on the US-ERD dataset are shown in Table 2. Table 2 shows the population characterization results on the US-ERD dataset.

[0050] The comparative experiments on CRYPTO-MD are shown in Table 3: Table 3 shows the population characterization results on the CRYPTO-MD dataset.

[0051] As can be seen from the results in Tables 2 and 3, our proposed method significantly outperforms other comparative models on all core evaluation metrics across both datasets, demonstrating outstanding group characterization performance. Single-modal methods, limited by a single information source, cannot comprehensively capture the multidimensional features of a group, resulting in generally poor performance. While existing multimodal methods improve characterization by fusing multi-source information, they have inherent limitations in cross-modal deep interaction modeling, adaptation to complex group patterns, and handling long-tail distributions, making it difficult to balance the accuracy of characterizing mainstream and niche groups. Our proposed method effectively addresses the three core challenges of multimodal heterogeneity, group pattern diversity, and long-tailed data distribution through a deep fusion mechanism of a bidirectional state-space model, a KAN multi-expert collaborative architecture, and a long-tailed boundary-aware loss function. It demonstrates significant advantages in both the accuracy and robustness of group feature characterization, fully validating the effectiveness and superiority of our approach.

Claims

1. A method for characterizing multi-expert collaborative network social groups based on modal dynamic fusion, characterized in that, First, through the modality representation adaptive extraction module, private representations are extracted from the original data of each modality in the online social network using channel attention and multi-head self-attention mechanisms. Second, through the modality complementary information fusion module, a bidirectional state space model is used to perform forward and backward global depth scans on the spliced ​​multimodal representations, explicitly modeling complex cross-modal interactions and generating a globally perceived unified fusion representation. Finally, through the multi-expert collaborative prediction module, multiple KANs are integrated and a sparse gating network is introduced for dynamic routing to achieve specialized processing and collaborative decision-making for group samples with different feature patterns. The steps are as follows: Step 1: Adaptive extraction of modal representations to generate highly discriminative private representations for each modality; Step 1.1: Input the raw multimodal data of the online social network, including speech text, group interaction values, member social relationships, and behavioral temporal modal data. Let the first step be... The modal data matrix is ,in For sample length, This is the feature dimension of the modality; Statistical information for each channel is extracted using global average pooling: in, Indicates the first The first modality One channel descriptor; For the first In modal data, the first The sample at the th The values ​​on each channel; For the first Total number of channels for modal data; Step 1.2: Feed the compressed channel descriptors into a bottleneck structure consisting of two fully connected layers to explicitly model the nonlinear correlation between channels and automatically generate channel importance vectors. : in, and This represents the learnable parameters of the two fully connected layers; and The activation function is represented by the channel importance signal, which is then multiplied channel by channel with the input data to obtain the recalibrated data matrix. : Step 1.3: Utilize multi-head self-attention mechanism to capture Deep contextual dependencies between internal features; utilizing three independent learnable linear transformations ,Will Mapped to queries respectively ,key Sum In three different subspaces: Step 1.4: Calculate the attention score using the scaled dot product. To quantify the relationships between different features within the same modality: In the formula, This is the scaling factor; multiple independent attention heads are used in parallel, each performing the above attention calculation, and the outputs of all heads are concatenated and fused through a linear projection layer to obtain the multi-head attention composite output: In the formula, These are the learnable parameters of the linear projection layer; Indicates the concatenation function; Step 1.5: Generate the modal private representation through residual connections and feedforward networks. : In the formula, Indicates a feedforward network; Step 2: Based on the bidirectional state-space model, perform modal complementary information fusion to construct a globally perceived unified fusion representation; Step 2.1: Concatenate the private representations of all modalities along the feature dimension using a concatenation function to obtain a concatenated representation containing full modal information. : In the formula, Indicates the total number of modes; Step 2.2: To capture the global dependencies spanning different modalities within the spliced ​​representation, a fusion mechanism based on bidirectional state space scanning is adopted: In the formula, This represents the classic state-space model; Indicates the inverse function; and The representation obtained by performing a depth scan of this concatenated representation from both the forward and backward dimensions; Step 2.3: Generate the final multimodal fusion representation through a linear transformation layer. : In the formula, These are the learnable parameters for a linear layer; Step 3: Multi-expert collaborative prediction to output a multi-dimensional feature representation of the population; Step 3.1, given the first Multimodal fusion representation of a sample of online social groups , construct a containing A collection of predictors from specialized KAN experts. Each expert is configured with a different combination of basis functions to specifically learn a particular type of nonlinear mapping pattern for group features; the output of a single expert network is: In the formula, Indicates the first The first expert network generated the first Preliminary characteristics of a sample of online social groups are indicated. Indicates the first The first expert network Layer Learnable activation functions for each node. This indicates that the output of the previous function is used as the input of the next function for continuous calculation; The activation function takes the form of a linear combination of basis functions and residual activation, and is specifically defined as follows: In the formula, It is a learnable activation function; The activation function is a fixed residual. This is an optional set of basis functions; and The learnable scaling factor; Step 3.2: Design a dynamic routing network based on sparse gating mechanism. Generate the activation weight vector for each expert network. : In the formula, The first generation generated for expert networks The activation weight of each online social group; The weights of the gating network; Before operation Find the maximum weight value and set the remainder to zero to ensure that computational efficiency is maintained while ensuring the model's expressive power. Step 3.3: Based on the sparse weight vector, obtain the final multi-dimensional feature representation of the population through a convex combination of the outputs of each expert network. : Step 4: End-to-end training optimization, improving model robustness based on long-tail boundary-aware loss function; Step 4.1: Construct the long-tailed boundary-aware objective function It consists of feature matching loss and class marginal adjustment loss, and is mathematically expressed as follows: In the formula, This is the loss balancing hyperparameter; For feature matching loss, Category marginal adjustment loss; Step 4.2: Calculate the feature matching loss and use the mean squared error to measure the difference between the predicted group features and the true features, ensuring global characterization accuracy. In the formula, For the first Real, multi-dimensional feature labels of a sample of online social groups; Step 4.3: Calculate the category marginal adjustment loss, compress the decision space of the mainstream group, expand the decision space of the tail samples, and optimize the decision boundary: In the formula, Number of group types; For the first The number of effective samples in a group; Sample of online social groups From the The class was misclassified to the first Resistance to the class; For the first The first in the class A sample of online social groups is used to predict which group belongs to the [number]th group. The original output of the class; For the first The first in the class A sample of online social groups is used to predict which group belongs to the [number]th group. The original output of the class; Step 4.4: Adopt an end-to-end training approach, backpropagate the loss value using the stochastic gradient descent algorithm, and iteratively update all trainable parameters of the model; set the number of training iterations until the loss function converges to obtain the optimal network social group characterization model, ensuring robustness in characterizing the tail-end rare group samples.

Images