An archive data automatic classification method and system based on machine learning

Abstract

The application discloses a kind of based on machine learning's archive data automatic classification method and system, including the following steps: step one: acquisition archive ontology data and behavior trajectory data;Step two: obtain text, image and structure vector;Step three: behavior trajectory data is arranged in time sequence, input to improved TS2Vec model, and introduce multi-view behavior projection mechanism, output behavior residual vector;Step four: generate archive representation vector;Step five: construct semantic graph structure, adopt graph neural network to carry out node embedding propagation, generate preliminary classification result;Step six: construct misclassified pulse path, propagate disturbance signal along misclassified pulse path, generate disturbance signal intensity matrix;Step seven: apply mirror image to train against loss function, obtain trained graph neural network;Step eight: output final classification result.The application fuses improved TS2Vec model and disturbance signal propagation, realizes archive data accurate intelligent classification.

Description

Technical Field

[0001] This invention relates to the field of archival information management technology, and in particular to an automatic classification method and system for archival data based on machine learning. Background Technology

[0002] With the increasing demand for digitized and intelligent management of archival information, automatic classification technology for multi-source archival data has become an important support for the construction of intelligent archival systems. Existing automatic archival classification methods mainly rely on keyword matching of text content or deep learning models based on single-modal features for label prediction, but these methods generally suffer from the following problems in practical applications:

[0003] Archival ontology data typically includes multimodal content such as text, images, and structured tables. Existing methods lack the ability to fuse and process multimodal information, failing to fully explore the complementary semantic relationships between different modalities. This results in low accuracy in archival representation and affects classification performance. Behavioral trajectory data, as an important dimension reflecting users' archival operation habits and usage preferences, is often ignored or only subjected to simple statistics, lacking in-depth modeling of temporal evolution patterns and semantic change features. For misclassified samples in the preliminary classification results, existing methods mostly adopt static loss function optimization strategies, lacking dynamic perception and intervention mechanisms for confusion relationships between samples. This leads to easy misidentification in situations where the boundaries between categories are blurred and the similarity of sample features is high, affecting the overall classification robustness.

[0004] Therefore, how to provide an automatic classification method and system for archival data based on machine learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an automatic classification method and system for archival data based on machine learning. This invention integrates an improved TS2Vec model with perturbation signal propagation, extracts archival behavior afterimage vectors by introducing a multi-view behavior projection mechanism, constructs a semantic graph structure by combining archival ontological features, utilizes graph neural networks for node embedding learning, and introduces misclassified pulse paths and mirror adversarial training strategies to optimize the classification boundary. This enables effective identification and correction of frequently confused categories, significantly improving the accuracy and robustness of automatic archival classification. It has significant advantages such as strong modeling ability, controllable error, and excellent generalization performance.

[0006] An automatic classification method for archival data based on machine learning according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Collect the archive data and behavioral trajectory data;

[0008] Step 2: Perform text semantic extraction, image embedding generation, and structured information parsing on the archive ontology data to generate text vectors, image vectors, and structure vectors;

[0009] Step 3: Arrange the behavior trajectory data in chronological order, input it into the improved TS2Vec model, and introduce a multi-view behavior projection mechanism to output behavior afterimage vectors;

[0010] Step 4: Concatenate and fuse the text vector, image vector, structure vector, and behavior afterimage vector to generate an archive representation vector;

[0011] Step 5: Construct a semantic graph structure based on the file representation vector. Each node in the semantic graph structure corresponds to a file representation vector. Use a graph neural network to perform node embedding propagation and generate preliminary classification results.

[0012] Step 6: Identify misclassified samples in the preliminary classification results, construct misclassified pulse paths, propagate perturbation signals along the misclassified pulse paths in the semantic graph structure, and generate a perturbation signal intensity matrix;

[0013] Step 7: Select frequently confused category pairs based on the perturbation signal strength matrix, construct mirror sample pairs, apply the mirror adversarial loss function to the mirror sample pairs for training, and obtain the trained graph neural network;

[0014] Step 8: Input the file representation vector of each file to be classified into the trained graph neural network to generate classification labels and classification confidence scores, and output the final classification results.

[0015] Optionally, the archive ontology data specifically includes text content, image information, and structured table data; the behavior trajectory data specifically includes the user's access, modification, approval, and borrowing behavior types, as well as the operator's identity, the time of the behavior, and terminal environment information.

[0016] Optionally, step two specifically includes:

[0017] The text content in the archive ontology data is normalized by character encoding and segmented. The segmentation results are input into a BERT language model containing a multi-layer Transformer encoder to extract context-related embedding vectors and perform average pooling to generate text vectors.

[0018] The image information in the archive ontology data is normalized by pixel matrix, and the processed image is input into a convolutional neural network containing convolutional layers, batch normalization layers and activation functions. The intermediate layer feature maps are extracted, and the image vector is obtained by global average pooling.

[0019] The structured table data in the archive ontology data is subjected to cell boundary detection and row and column structure recognition. The table content is parsed into a sequence of key-value pairs. Each key-value pair is converted into word vectors and then concatenated with position encoding. The result is input into a multilayer perceptron structure for nonlinear transformation to generate a structure vector.

[0020] Optionally, the improved TS2Vec model specifically includes a time-step encoding module, a hierarchical temporal modeling module, a dynamic viewpoint interleaving module, and a temporal compression output module;

[0021] The time step encoding module receives behavior trajectory data arranged in chronological order, embeds the behavior type, operator identity, behavior occurrence time and terminal environment information in each behavior record, outputs four equal-length vectors and concatenates them according to the field order to generate a time step vector sequence.

[0022] The hierarchical temporal modeling module receives a time step vector sequence and sequentially inputs three one-dimensional convolutional channels. The convolutional kernel lengths are 3, 5, and 7, respectively, and the sliding stride is 1. The module extracts local change patterns and concatenates the outputs of the three convolutional channels in the channel dimension to output a sequence of behavioral feature vectors.

[0023] The dynamic view interleaving module receives a sequence of behavioral feature vectors and introduces a multi-view behavioral projection mechanism to construct three perturbation views: semantic, temporal, and density.

[0024] The three perspective tensors are subjected to a channel-wise weighted summation operation with the sequence of behavioral feature vectors, and the weighted result is non-linearly activated using the ReLU activation function to generate a fused behavioral representation tensor.

[0025] The temporal compression output module receives the fused behavior representation tensor, performs average pooling processing, and inputs it into a two-layer fully connected network. The first layer of the fully connected network performs a dimension compression operation, compressing the dimension of the input feature vector from a high-dimensional space to a preset low-dimensional space. The second layer of the fully connected network uses a hyperbolic tangent activation function to perform a nonlinear transformation, generating a behavior afterimage vector.

[0026] Optionally, the multi-view behavior projection mechanism specifically includes a semantic rearrangement perspective construction method, a temporal perturbation projection method, and a density aggregation projection method;

[0027] The semantic rearrangement perspective construction method calculates the difference between the feature vector of each time step in the embedding space and the overall average vector of the behavior feature vector sequence. The difference value is obtained by subtracting the values ​​of the corresponding positions of each dimension of the two vectors, taking the absolute value, and summing them. The time steps are then reordered according to the difference value from smallest to largest. The rearranged behavior feature vector sequence constitutes the semantic rearrangement perspective tensor.

[0028] The time perturbation projection method statistically analyzes the time difference between every two adjacent time steps, calculates the average time interval of the entire behavioral feature vector sequence, and sets twice the average time interval as the time segmentation threshold. When the time difference between adjacent time steps is less than the time segmentation threshold, they are determined to be in the same time period. When the time difference between adjacent time steps is greater than the time segmentation threshold, they are divided into a new time period.

[0029] The behavioral feature vectors of each time period are summed and divided by the number of behavioral feature vectors in that time period to obtain the average vector of the time period. The average vectors of the time periods are arranged in chronological order to form a sequence of average vectors of the time periods, which serves as the time perturbation perspective tensor.

[0030] The density aggregation projection method counts the number of times each time step belongs to the behavior type in the entire behavior feature vector sequence to obtain the frequency of each behavior type; the proportion of each behavior type in the total frequency at each time step is used as a weighting factor, and multiplied element-wise by the behavior feature vector of the corresponding time step to obtain a weighted vector; all weighted vectors are summed and divided by the total number of time steps to obtain the density-aggregated behavior feature vector, and copied into a tensor of the same length as the behavior feature vector sequence to form a frequency-view tensor.

[0031] Optionally, step five specifically includes:

[0032] A semantic graph structure is constructed based on all file representation vectors. Each file representation vector is used as a graph node. The cosine similarity between any two file representation vectors is calculated. If the cosine similarity exceeds a set similarity threshold, an edge connection is established between the corresponding nodes, and the cosine similarity is used as the edge weight value to form an adjacency matrix.

[0033] The semantic graph structure is input into a graph neural network model, which includes multiple graph convolutional layers. Each graph convolutional layer receives the node embedding representation and adjacency matrix from the previous layer and updates the node embedding representation through neighborhood feature aggregation operations.

[0034] The neighborhood feature aggregation operation performs a weighted summation of the target node embedding representation and the representations of its directly connected neighbor nodes, and then concatenates them to obtain an intermediate embedding representation. The intermediate embedding representation is then input into the ReLU activation function for nonlinear mapping to generate an updated node embedding representation.

[0035] The node embedding representation output from the last graph convolutional layer is input into the fully connected layer and then into the Softmax classifier to obtain the preliminary classification label and confidence score for each node, which serves as the preliminary classification result.

[0036] Optionally, step six specifically includes:

[0037] The node samples with confidence scores lower than the preset confidence threshold in the preliminary classification results are compared with the actual labels, and the nodes whose actual labels do not match the predicted labels are selected as misclassified samples.

[0038] For each misclassified sample node, a misclassification pulse path is constructed based on the true category center node and the predicted category center node in the semantic graph structure. The misclassification pulse path starts from the misclassified sample node and sequentially connects the predicted category center node, all intermediate propagation nodes on the shortest connection path between the predicted category center node and the true category center node, and the true category center node, as a perturbation signal propagation path. The intermediate propagation nodes are selected from the adjacency matrix using the Dijkstra algorithm.

[0039] In the semantic graph structure, a perturbation signal is injected into each misclassified sample node. The initial strength of the perturbation signal is set to 1. The perturbation signal propagates hop by hop along the misclassification pulse path. During each hop propagation process, the current perturbation signal strength is multiplied by the edge weight value between the previous hop node and the current node as the new perturbation signal strength, until the end of the misclassification pulse path is reached.

[0040] Statistical analysis is performed on the propagation records of disturbance signals in all misclassified pulse paths. Based on the correspondence between the final propagated true class label and the initial predicted class label, a disturbance signal intensity matrix is ​​constructed. Each element in the disturbance signal intensity matrix represents the sum of the cumulative received disturbance signal intensities between class pairs.

[0041] Optionally, step seven specifically includes:

[0042] Based on the disturbance signal strength matrix, category pairs with cumulative disturbance signal strength greater than a preset strength threshold are selected and determined to be frequently confused category pairs;

[0043] From the training samples, select archive samples with the true labels of two different categories in the frequently confused category pair, obtain the archive representation vector of each archive sample, and take the archive representation vector of each archive sample in one category (category A) as the benchmark, and find the archive representation vector with the highest cosine similarity in the other category (category B) as the matching object to form a mirror sample pair; each mirror sample pair consists of a category A sample and a category B sample with the most similar features.

[0044] For each pair of mirror samples, the mirror adversarial loss value is calculated. The mirror adversarial loss function consists of two parts: label cross-entropy loss and contrast preservation loss. The label cross-entropy loss is obtained by calculating the cross-entropy value between the predicted probability distribution of each file sample and the true label, and averaging the cross-entropy values ​​of the two file samples.

[0045] The contrast preservation loss is obtained by calculating the square of the difference between the values ​​at each dimension of the probability distribution vector of the two archival samples and summing them.

[0046] The cross-entropy loss value is added to the contrast-preservation loss value to obtain the mirror adversarial loss value;

[0047] The mirror adversarial loss value is input into the graph neural network training process, the backpropagation algorithm is executed, and the graph neural network parameters are trained according to the gradient direction corresponding to the mirror adversarial loss value to obtain the trained graph neural network.

[0048] Optionally, step eight specifically includes:

[0049] The file representation vector of each file to be classified is input into the trained graph neural network. The node embedding representation output by the graph neural network is mapped to the classification space through a fully connected layer. The normalized probability of each category is calculated by the Softmax classifier, which serves as the confidence probability distribution of the file belonging to each category.

[0050] Based on the index corresponding to the category with the highest confidence probability distribution, determine the classification label of the file, use the highest probability value as the confidence of the classification result, and output the final classification result.

[0051] An automatic classification system for archival data based on machine learning according to an embodiment of the present invention includes the following modules:

[0052] The data acquisition module is used to collect data on the archive itself and behavioral trajectory data;

[0053] The ontology feature extraction module is used to extract text vectors, image vectors, and structural vectors from the archival ontology data;

[0054] The behavior modeling module is used to encode the behavior trajectory data into a multi-dimensional time series in chronological order, input it into the improved TS2Vec model, and introduce a multi-view behavior projection mechanism to output behavior afterimage vectors.

[0055] The vector fusion module is used to concatenate and fuse text vectors, image vectors, structural vectors, and behavioral afterimage vectors to generate archive representation vectors;

[0056] A semantic graph construction module is used to construct a semantic graph structure based on the file representation vector and form an adjacency matrix;

[0057] The graph neural network module is used to perform node embedding propagation and generate preliminary classification results;

[0058] The misclassification identification and perturbation propagation module is used to identify misclassified samples in the preliminary classification results, construct misclassification pulse paths, propagate perturbation signals along the misclassification pulse paths in the semantic graph structure, and generate a perturbation signal intensity matrix.

[0059] The mirror adversarial training module is used to select frequently confused category pairs based on the perturbation signal strength matrix, construct mirror sample pairs, apply a mirror adversarial loss function to the mirror sample pairs for training, and obtain the trained graph neural network.

[0060] The classification reasoning module is used to input the file representation vector of each file to be classified into the trained graph neural network, generate classification labels and classification confidence scores, and output the final classification result.

[0061] The beneficial effects of this invention are:

[0062] This invention addresses the challenges of diverse modalities in archival ontology data, complex temporal sequences in behavioral trajectory data, and susceptibility to classification errors by combining multimodal feature extraction with graph neural network training. It proposes an ontology feature extraction module that integrates text semantic extraction, image embedding generation, and structured information parsing. This module, combined with an improved TS2Vec model and a multi-view behavior projection mechanism, extracts temporally dynamic behavioral afterimage vectors. A vector fusion module performs deep concatenation of multi-source features to generate a complete archival representation vector. A high-dimensional semantic graph structure is constructed using a semantic graph construction module and input into the graph neural network module for node embedding and propagation, achieving preliminary classification. For misclassified samples in the preliminary classification results, a misclassification identification and perturbation propagation module is designed. This module constructs misclassification pulse paths and propagates perturbation signals, statistically generating a perturbation signal intensity matrix. Furthermore, a mirror adversarial training module identifies frequently confused category pairs, constructs mirror sample pairs, applies a mirror adversarial loss function, and optimizes the graph neural network parameters according to the gradient direction. Finally, the classification inference module achieves the joint output of classification labels and classification confidence, effectively improving the accuracy, robustness, and ability to distinguish easily confused categories in automatic archival data classification. Attached Figure Description

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0064] Figure 1 This is an overall flowchart of an automatic classification method for archival data based on machine learning proposed in this invention;

[0065] Figure 2 This is a schematic diagram of the structure of an automatic classification system for archival data based on machine learning proposed in this invention. Detailed Implementation

[0066] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0067] refer to Figure 1 An automatic classification method for archival data based on machine learning includes the following steps:

[0068] Step 1: Collect the archive data and behavioral trajectory data;

[0069] Step 2: Perform text semantic extraction, image embedding generation, and structured information parsing on the archive ontology data to generate text vectors, image vectors, and structure vectors;

[0070] Step 3: Arrange the behavior trajectory data in chronological order, input it into the improved TS2Vec model, and introduce a multi-view behavior projection mechanism to output behavior afterimage vectors;

[0071] Step 4: Concatenate and fuse the text vector, image vector, structure vector, and behavior afterimage vector to generate an archive representation vector;

[0072] Step 5: Construct a semantic graph structure based on the file representation vector. Each node in the semantic graph structure corresponds to a file representation vector. Use a graph neural network to perform node embedding propagation and generate preliminary classification results.

[0073] Step 6: Identify misclassified samples in the preliminary classification results, construct misclassified pulse paths, propagate perturbation signals along the misclassified pulse paths in the semantic graph structure, and generate a perturbation signal intensity matrix;

[0074] Step 7: Select frequently confused category pairs based on the perturbation signal strength matrix, construct mirror sample pairs, apply the mirror adversarial loss function to the mirror sample pairs for training, and obtain the trained graph neural network;

[0075] Step 8: Input the file representation vector of each file to be classified into the trained graph neural network to generate classification labels and classification confidence scores, and output the final classification results.

[0076] In this embodiment, the archive ontology data specifically includes text content, image information, and structured table data; the behavior trajectory data specifically includes the user's access to, modification, approval, and borrowing behavior types, as well as the operator's identity, the time of the behavior, and terminal environment information.

[0077] In this embodiment, step two specifically includes:

[0078] The text content in the archive ontology data is normalized by character encoding and segmented. The segmentation results are input into a BERT language model containing a multi-layer Transformer encoder to extract context-related embedding vectors and perform average pooling to generate text vectors.

[0079] The image information in the archive ontology data is normalized by pixel matrix, and the processed image is input into a convolutional neural network containing convolutional layers, batch normalization layers and activation functions. The intermediate layer feature maps are extracted, and the image vector is obtained by global average pooling.

[0080] The structured table data in the archive ontology data is subjected to cell boundary detection and row and column structure recognition. The table content is parsed into a key-value pair sequence. Each key-value pair is converted into word vectors and concatenated with position encoding. The result is then input into a multilayer perceptron structure for nonlinear transformation to generate a structure vector.

[0081] In the text content processing process, this invention uses character encoding standardization to unify the character encoding format in the document, avoiding garbled text or word segmentation errors caused by inconsistent encoding. Specifically, encodings such as GBK, ISO-8859-1, and UTF-16 are uniformly converted to UTF-8 encoding, and invisible control characters and illegal characters are removed. The word segmentation results are fed into the BERT language model, which contains multiple layers of Transformer encoders. This language model is composed of multiple stacked TransformerBlocks. Each Block includes a self-attention mechanism layer, a feedforward neural network layer, and a residual connection structure, which can perform deep semantic modeling of each word in the context and extract word vectors with context relevance.

[0082] In image processing, pixel matrix normalization normalizes the pixel values ​​of the original image to the range of 0 to 1, which helps improve the convergence speed and numerical stability of convolutional neural networks. A convolutional neural network contains multiple consecutive convolutional layers, batch normalization layers, and a non-linear ReLU activation function to extract local and global features of the image. The intermediate layer feature maps are compressed into fixed-length image vectors through global average pooling.

[0083] In structured table processing, cell boundary detection determines the row and column boundaries of each cell in the table through HTML tag parsing, and performs structural semantic mapping of the table content in conjunction with the table header. After parsing the fields and corresponding values ​​in the table into key-value pairs, word vector encoding and positional encoding are concatenated to form an input vector. This vector is then fed into a multilayer perceptron, which consists of multiple linear transformation layers and ReLU activation functions, for nonlinear mapping transformation to obtain a structure vector that can capture the relationships between fields and the spatial features of the table structure.

[0084] In this embodiment, the improved TS2Vec model specifically includes a time step encoding module, a hierarchical temporal modeling module, a dynamic viewpoint interleaving module, and a temporal compression output module;

[0085] The time step encoding module receives behavior trajectory data arranged in chronological order, embeds the behavior type, operator identity, behavior occurrence time and terminal environment information in each behavior record, outputs four equal-length vectors and concatenates them according to the field order to generate a time step vector sequence.

[0086] The hierarchical temporal modeling module receives a time step vector sequence and sequentially inputs three one-dimensional convolutional channels. The convolutional kernel lengths are 3, 5, and 7, respectively, and the sliding stride is 1. The module extracts local change patterns and concatenates the outputs of the three convolutional channels in the channel dimension to output a sequence of behavioral feature vectors.

[0087] The dynamic view interleaving module receives a sequence of behavioral feature vectors and introduces a multi-view behavioral projection mechanism to construct three perturbation views: semantic, temporal, and density.

[0088] The three perspective tensors are subjected to a channel-wise weighted summation operation with the sequence of behavioral feature vectors, and the weighted result is non-linearly activated using the ReLU activation function to generate a fused behavioral representation tensor.

[0089] The temporal compression output module receives the fused behavior representation tensor, performs average pooling processing, and inputs it into a two-layer fully connected network. The first layer of the fully connected network performs a dimension compression operation, compressing the dimension of the input feature vector from a high-dimensional space to a preset low-dimensional space. The second layer of the fully connected network uses a hyperbolic tangent activation function to perform a nonlinear transformation, generating a behavior afterimage vector.

[0090] In this embodiment, the multi-view behavior projection mechanism specifically includes a semantic rearrangement perspective construction method, a temporal perturbation projection method, and a density aggregation projection method;

[0091] The semantic rearrangement perspective construction method calculates the difference between the feature vector of each time step in the embedding space and the overall average vector of the behavior feature vector sequence. The difference value is obtained by subtracting the values ​​of the corresponding positions of each dimension of the two vectors, taking the absolute value, and summing them. The time steps are then reordered according to the difference value from smallest to largest. The rearranged behavior feature vector sequence constitutes the semantic rearrangement perspective tensor.

[0092] The time perturbation projection method statistically analyzes the time difference between every two adjacent time steps, calculates the average time interval of the entire behavioral feature vector sequence, and sets twice the average time interval as the time segmentation threshold. When the time difference between adjacent time steps is less than the time segmentation threshold, they are determined to be in the same time period. When the time difference between adjacent time steps is greater than the time segmentation threshold, they are divided into a new time period.

[0093] The behavioral feature vectors of each time period are summed and divided by the number of behavioral feature vectors in that time period to obtain the average vector of the time period. The average vectors of the time periods are arranged in chronological order to form a sequence of average vectors of the time periods, which serves as the time perturbation perspective tensor.

[0094] The density aggregation projection method counts the number of times each time step belongs to the behavior type in the entire behavior feature vector sequence to obtain the frequency of each behavior type; the proportion of each behavior type in the total frequency of each time step is used as a weighting factor, and multiplied element-wise by the behavior feature vector of the corresponding time step to obtain a weighted vector. All weighted vectors are summed and divided by the total number of time steps to obtain the density-aggregated behavior feature vector, which is then copied into a tensor of the same length as the behavior feature vector sequence to form a frequency-view tensor.

[0095] To enhance the temporal representation capability and robustness of the TS2Vec model in archival behavior trajectory modeling, this invention proposes a structurally improved TS2Vec model, specifically including a time step encoding module, a hierarchical temporal modeling module, a dynamic viewpoint interleaving module, and a temporal compression output module. This model introduces a multi-view behavior perturbation mechanism, reconstructing and perturbing the behavior feature sequence from three dimensions: semantics, time, and density. This enhances the model's ability to perceive potential structural changes in the behavior sequence and strengthens the robustness of the behavior representation to heterogeneity and local changes.

[0096] The time-step encoding module constructs a time-step vector sequence containing multi-source heterogeneous information by independently embedding and concatenating information such as behavior type, operator identity, behavior occurrence time, and terminal environment, providing a sufficient input foundation for modeling. The hierarchical temporal modeling module uses three one-dimensional convolutional channels with different kernel sizes to extract local change patterns, effectively improving the model's ability to capture local features while preserving dynamic information of different granularities in the behavior sequence.

[0097] The core improvement lies in the multi-view behavior projection mechanism introduced in the dynamic perspective interleaving module. This mechanism constructs three structural perturbation methods: semantic rearrangement perspective, temporal perturbation perspective, and density aggregation perspective. The semantic rearrangement perspective rearranges behavior by calculating the difference between each time step and the average behavior vector, guiding the model to focus on segments in the behavior sequence that deviate from the global average pattern, thereby enhancing the modeling ability for atypical behaviors. The temporal perturbation perspective generates time segment representations based on a dynamic segmentation strategy of behavior time difference, enabling the model to model non-uniform temporal behavior patterns. The density aggregation perspective introduces category-level semantic density constraints through global frequency weighted fusion of behavior types, which helps to highlight the importance of high-frequency behaviors and enhance feature stability.

[0098] The three perspective tensors are weighted and fused with the original behavioral feature sequence and then processed nonlinearly by the ReLU activation function to generate a fused behavioral representation tensor. The pooling and fully connected transformations are completed by the temporal compression output module to output the behavioral afterimage vector, providing a structurally stable and highly discriminative behavioral representation for semantic graph construction and classification.

[0099] Through the above structural improvements, the improved TS2Vec model effectively overcomes the insensitivity of the original model to abnormal temporal distributions and imbalances in behavioral categories, enhances the model's expressive and generalization capabilities in automatic classification scenarios of archival data, significantly improves classification accuracy, and strengthens the model's robustness and adaptability when facing complex behavioral trajectory data.

[0100] In this embodiment, step five specifically includes:

[0101] A semantic graph structure is constructed based on all file representation vectors. Each file representation vector is used as a graph node. The cosine similarity between any two file representation vectors is calculated. If the cosine similarity exceeds a set similarity threshold, an edge connection is established between the corresponding nodes, and the cosine similarity is used as the edge weight value to form an adjacency matrix.

[0102] The semantic graph structure is input into a graph neural network model, which includes multiple graph convolutional layers. Each graph convolutional layer receives the node embedding representation and adjacency matrix from the previous layer and updates the node embedding representation through neighborhood feature aggregation operations.

[0103] The neighborhood feature aggregation operation performs a weighted summation of the target node embedding representation and the representations of its directly connected neighbor nodes, and then concatenates them to obtain an intermediate embedding representation. The intermediate embedding representation is then input into the ReLU activation function for nonlinear mapping to generate an updated node embedding representation.

[0104] The node embedding representation output from the last graph convolutional layer is input into the fully connected layer and then into the Softmax classifier to obtain the preliminary classification label and confidence score for each node, which serves as the preliminary classification result.

[0105] In this invention, after the graph neural network undergoes multiple layers of graph convolution operations, it outputs the final node embedding representation of each graph node (i.e., the archive data representation vector), which is represented as a set of high-dimensional vectors to comprehensively reflect the semantic features of the archive in the global semantic graph. In order to transform these vector-form embedding information into specific classification labels and their corresponding confidence values, the final node embedding representation is input into a fully connected layer neural network.

[0106] Specifically, the fully connected layer maps the input final node embedding representation into a vector of length K, where K is the total number of predefined file categories. Each element in this vector represents the unnormalized probability score of the node belonging to the corresponding category. This vector is then input into the Softmax classifier for normalization. The Softmax function exponentially calculates the unnormalized probability score for each category and normalizes it by summing the exponents of all categories, resulting in a K-dimensional probability distribution vector. Each dimension represents the confidence level of the file belonging to the corresponding category, with values ​​ranging from 0 to 1, and the sum of the confidence levels of all categories is 1.

[0107] The category corresponding to the maximum probability value is selected based on the probability distribution vector and used as the initial classification label for that node. The maximum probability value is recorded as the classification confidence index.

[0108] In this embodiment, step six specifically includes:

[0109] The node samples with confidence scores lower than the preset confidence threshold in the preliminary classification results are compared with the actual labels, and the nodes whose actual labels do not match the predicted labels are selected as misclassified samples.

[0110] For each misclassified sample node, a misclassification pulse path is constructed based on the true category center node and the predicted category center node in the semantic graph structure. The misclassification pulse path starts from the misclassified sample node and sequentially connects the predicted category center node, all intermediate propagation nodes on the shortest connection path between the predicted category center node and the true category center node, and the true category center node, as a perturbation signal propagation path. The intermediate propagation nodes are selected from the adjacency matrix using the Dijkstra algorithm.

[0111] Specifically, in the construction of the misclassified pulse path, each path starts with a misclassified sample node, which is the node in the preliminary classification result output by the graph neural network where the true class label and the predicted class label are inconsistent. The second node of the path is the center node of the predicted class to which the misclassified sample belongs. This center node is the node with the highest average cosine similarity among all nodes predicted to belong to this class, representing the feature centroid of this class in the semantic graph. Starting from the center node of the predicted class, the shortest path to the true class center node is searched using Dijkstra's shortest path algorithm through the edge connections already constructed in the semantic graph structure. All intermediate nodes on this path, except for the starting and ending points, are selected and arranged in order in the path as intermediate propagation nodes of the misclassified pulse path. The entire misclassified pulse path includes, in sequence: the misclassified sample node, the predicted class center node, several intermediate propagation nodes, and the true class center node, forming a path channel for the propagation of the perturbation signal. This guides the perturbation to focus in the class boundary region and records the interference relationship between classes, while also reflecting the propagation context of the model's erroneous decision in the graph space.

[0112] In the semantic graph structure, a perturbation signal is injected into each misclassified sample node. The initial strength of the perturbation signal is set to 1. The perturbation signal propagates hop by hop along the misclassification pulse path. During each hop propagation process, the current perturbation signal strength is multiplied by the edge weight value between the previous hop node and the current node as the new perturbation signal strength, until the end of the misclassification pulse path is reached.

[0113] Statistical analysis is performed on the propagation records of disturbance signals in all misclassified pulse paths. Based on the correspondence between the final propagated true class label and the initial predicted class label, a disturbance signal intensity matrix is ​​constructed. Each element in the disturbance signal intensity matrix represents the sum of the cumulative received disturbance signal intensities between class pairs.

[0114] In this embodiment, step seven specifically includes:

[0115] Based on the disturbance signal strength matrix, category pairs with cumulative disturbance signal strength greater than a preset strength threshold are selected and determined to be frequently confused category pairs;

[0116] From the training samples, select archive samples with the true labels of two different categories in the frequently confused category pair, obtain the archive representation vector of each archive sample, and take the archive representation vector of each archive sample in one category (category A) as the benchmark, and find the archive representation vector with the highest cosine similarity in the other category (category B) as the matching object to form a mirror sample pair; each mirror sample pair consists of a category A sample and a category B sample with the most similar features.

[0117] For each pair of mirror samples, the mirror adversarial loss value is calculated. The mirror adversarial loss function consists of two parts: label cross-entropy loss and contrast preservation loss. The label cross-entropy loss is obtained by calculating the cross-entropy value between the predicted probability distribution of each file sample and the true label, and averaging the cross-entropy values ​​of the two file samples.

[0118] The contrast preservation loss is obtained by calculating the square of the difference between the values ​​at each dimension of the probability distribution vector of the two archival samples and summing them.

[0119] The cross-entropy loss value is added to the contrast-preservation loss value to obtain the mirror adversarial loss value, which represents a comprehensive error measure of the mirror sample pair in terms of label prediction accuracy and prediction consistency.

[0120] The mirror adversarial loss value is input into the graph neural network training process, the backpropagation algorithm is executed, and the graph neural network parameters are trained according to the gradient direction corresponding to the mirror adversarial loss value to obtain the trained graph neural network.

[0121] The parameters of the graph neural network include the weight matrix and bias term in the graph convolutional layer, and the weight matrix and bias term in the fully connected layer. Taking the graph convolutional layer as an example, the parameter update process is as follows: First, calculate the partial derivative of the mirror adversarial loss value with respect to the output of the current layer. Then, based on the derivative of the ReLU activation function of the current layer, backpropagate to the weight matrix of that layer, and finally obtain the gradient of the loss function with respect to the weight matrix. The update method of the fully connected layer parameters is similar. The gradient of the loss function with respect to the output is backpropagated to the weight matrix and bias term according to the chain rule.

[0122] The Adam optimization algorithm is used to update the parameters numerically according to the learning rate adjustment direction, thereby realizing iterative training of the graph neural network parameters. This allows the graph neural network to be continuously optimized under the training constraints of mirror sample pairs, improving its ability to distinguish confused class pairs. This training process repeats the above operation in each iteration until the change of the mirror adversarial loss value in several consecutive training cycles is less than the preset convergence threshold or the preset maximum number of iterations.

[0123] In this embodiment, step eight specifically includes:

[0124] The file representation vector of each file to be classified is input into the trained graph neural network. The node embedding representation output by the graph neural network is mapped to the classification space through a fully connected layer. The normalized probability of each category is calculated by the Softmax classifier, which serves as the confidence probability distribution of the file belonging to each category.

[0125] Based on the index corresponding to the category with the highest confidence probability distribution, determine the classification label of the file, use the highest probability value as the confidence of the classification result, and output the final classification result.

[0126] refer to Figure 2 An automatic classification system for archival data based on machine learning includes the following modules:

[0127] The data acquisition module is used to collect data on the archive itself and behavioral trajectory data;

[0128] The ontology feature extraction module is used to extract text vectors, image vectors, and structural vectors from the archival ontology data;

[0129] The behavior modeling module is used to encode the behavior trajectory data into a multi-dimensional time series in chronological order, input it into the improved TS2Vec model, and introduce a multi-view behavior projection mechanism to output behavior afterimage vectors.

[0130] The vector fusion module is used to concatenate and fuse text vectors, image vectors, structural vectors, and behavioral afterimage vectors to generate archive representation vectors;

[0131] A semantic graph construction module is used to construct a semantic graph structure based on the file representation vector and form an adjacency matrix;

[0132] The graph neural network module is used to perform node embedding propagation and generate preliminary classification results;

[0133] The misclassification identification and perturbation propagation module is used to identify misclassified samples in the preliminary classification results, construct misclassification pulse paths, propagate perturbation signals along the misclassification pulse paths in the semantic graph structure, and generate a perturbation signal intensity matrix.

[0134] The mirror adversarial training module is used to select frequently confused category pairs based on the perturbation signal strength matrix, construct mirror sample pairs, apply a mirror adversarial loss function to the mirror sample pairs for training, and obtain the trained graph neural network.

[0135] The classification reasoning module is used to input the file representation vector of each file to be classified into the trained graph neural network, generate classification labels and classification confidence scores, and output the final classification result.

[0136] Example 1:

[0137] To verify the feasibility of this invention in practice, it was applied to an archive management platform to address the problems of low classification accuracy, low utilization rate of behavioral information, and ineffective identification of easily confused categories in multimodal archive information. This platform has long faced a large amount of heterogeneous archive data with complex archive ontology data structures, including scanned text documents, table images, spreadsheets, etc., accompanied by user access, modification, approval, and borrowing behavior trajectory information. Traditional classification models often rely on text keyword matching or image OCR analysis, which struggles to fully explore the deep relationships between semantic graphs and behavioral sequences, resulting in low classification accuracy, especially in categories with frequent business overlap, where misclassification is severe.

[0138] In the practical application of this platform, the data acquisition module first collects archival ontology data and behavioral trajectory data from the archival management system. The archival ontology data includes document text content, image scan information, and structured table fields; the behavioral trajectory data includes user access, modification, approval, and borrowing behavior types, as well as operator identity, behavior occurrence time, and terminal environment information. Taking a batch of historical data containing 8764 archives as an example, the ontology feature extraction module extracts text vectors, image vectors, and structure vectors respectively. Simultaneously, an improved TS2Vec model is used to model the behavioral trajectory. Through a multi-view behavioral projection mechanism, semantic rearrangement perspective, temporal perturbation perspective, and density aggregation perspective are constructed respectively. The original temporal behavioral data undergoes perturbation semantic transformation, outputting behavioral afterimage vectors, which are finally concatenated and fused into an archival representation vector.

[0139] In the graph neural network module, a semantic graph structure is constructed based on all file representation vectors. Cosine similarity is calculated to construct edge connections and adjacency matrices. After node embedding propagation using a multi-layer graph convolutional network, preliminary classification results are output. In practice, the platform identified a large number of low-confidence nodes, especially significant confusion between the project plan and implementation schedule categories. The misclassification identification and perturbation propagation module constructs misclassified impulse paths and a perturbation signal intensity matrix based on the perturbation signal intensity. Frequently confused category pairs are obtained through threshold screening, and representative samples are selected to construct mirror sample pairs. The mirror adversarial loss function is applied for training and optimization. During the training phase, the error backpropagation algorithm is executed, combining label cross-entropy loss and contrast-preservation loss to effectively enhance the model's ability to distinguish easily confused categories.

[0140] Table 1. Comparison of classification accuracy before and after training for frequent category confusion.

[0141]

[0142] As can be seen from the data in Table 1 above, the classification accuracy of the five frequently confused category pairs significantly improved after introducing the mirror adversarial training mechanism, demonstrating the effectiveness of this invention in correcting misclassification. The original accuracy was generally between 68.1% and 75.4%, which is considered low to medium, indicating that traditional graph neural networks have limited ability to distinguish confused categories. After mirror adversarial training, the accuracy improved to over 85%, with the highest improvement at 25.9% (project plan / implementation schedule) and the lowest at 19.6% (contract template / authorization document), showing stable overall performance. The category pairs with the largest improvement often had lower original accuracy, indicating that the mirror adversarial mechanism is more effective in handling highly confused samples. By introducing highly similar but dissimilar mirror samples for adversarial training, the model's ability to distinguish boundary samples is effectively enhanced, improving its robustness and generalization ability.

[0143] This embodiment integrates the behavioral afterimage vector generated by the improved TS2Vec model with multimodal archival ontology features to construct a semantic graph structure and introduces a perturbation signal for the propagation of misclassified pulse paths, accurately identifying frequently confused category pairs. Furthermore, by combining mirror sample pair construction and mirror adversarial loss training, the graph neural network's ability to discriminate boundary samples and distinguish easily confused categories is significantly enhanced, effectively improving the accuracy and stability of archival data classification. In particular, it demonstrates good generalization ability and robustness in complex and diverse archival environments with obvious category overlap, verifying the practicality and advancement of this invention in real-world applications.

[0144] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for automatic classification of archival data based on machine learning, characterized in that, The method comprises the following steps: Step 1: collecting profile ontology data and behavior trajectory data; Step 2: performing text semantic extraction, image embedding generation and structured information analysis on the profile ontology data to generate text vectors, image vectors and structure vectors; Step 3: arranging the behavior trajectory data in chronological order, inputting the data into an improved TS2Vec model, and introducing a multi-view behavior projection mechanism to output behavior residual shadow vectors; Step 4: splicing and fusing the text vectors, image vectors, structure vectors and behavior residual shadow vectors to generate profile representation vectors; Step 5: constructing a semantic graph structure based on the profile representation vectors, each node in the semantic graph structure corresponding to a profile representation vector, performing node embedding propagation using a graph neural network to generate a preliminary classification result; Step 6: identifying misclassified samples in the preliminary classification result, constructing a misclassified pulse path, propagating a disturbance signal along the misclassified pulse path in the semantic graph structure to generate a disturbance signal strength matrix; Step 7: selecting a pair of frequently confused categories based on the disturbance signal strength matrix, constructing a mirror sample pair, and training the graph neural network based on the mirror sample pair and a mirror adversarial loss function to obtain a trained graph neural network; Step 8: inputting the profile representation vector of each profile to be classified into the trained graph neural network to generate a classification label and a classification confidence, and outputting a final classification result; The step 6 specifically comprises: Comparing the node samples with confidence lower than a preset confidence threshold in the preliminary classification result with actual labels to filter out nodes with inconsistent actual labels and predicted labels as misclassified samples; For each misclassified sample node, a misclassified pulse path is constructed based on a real category center node and a predicted category center node in the semantic graph structure, the misclassified pulse path being a disturbance signal propagation path starting from the misclassified sample node, sequentially connecting the predicted category center node, all intermediate propagation nodes on the shortest connection path between the predicted category center node and the real category center node, and the real category center node, and the intermediate propagation nodes being selected by Dijkstra algorithm in an adjacency matrix; A disturbance signal is injected into each misclassified sample node in the semantic graph structure, the initial strength of the disturbance signal being 1, and the disturbance signal propagating hop by hop along the misclassified pulse path, the current disturbance signal strength being multiplied by the edge weight value between the previous hop node and the current node as the new disturbance signal strength at each hop propagation process until the end of the misclassified pulse path; The disturbance signal propagation records in all misclassified pulse paths are counted, and a disturbance signal strength matrix is constructed based on the correspondence between the final propagated real category label and the initial predicted category label, each matrix element in the disturbance signal strength matrix representing the total disturbance signal strength accumulated between the category pair.

2. The method of claim 1, wherein the method is based on machine learning. The profile ontology data specifically comprises text content, image information and structured table data; and the behavior trajectory data specifically comprises user access, modification, approval, borrowing behavior types and operator identity, behavior occurrence time and terminal environment information. 3.The method of claim 1, wherein, The step 2 specifically comprises: The text content in the archive ontology data is subjected to character coding normalization and word segmentation processing, the word segmentation result is input into a BERT language model comprising a multi-layer Transformer encoder, context-related embedding vectors are extracted and average pooling is performed to generate a text vector; The image information in the archive ontology data is subjected to pixel matrix normalization, the processed image is input into a convolutional neural network comprising a convolutional layer, a batch normalization layer and an activation function, an intermediate layer feature map is extracted, and an image vector is obtained through global average pooling; The structured table data in the archive ontology data is subjected to cell boundary detection and row-column structure recognition, the table content is parsed into a key-value pair sequence, each key-value pair is subjected to word vector conversion and position encoding splicing, and is input into a multi-layer perceptron structure for non-linear transformation to generate a structure vector.

4. The method of claim 1, wherein the method is based on machine learning. The improved TS2Vec model specifically comprises a time step encoding module, a hierarchical time series modeling module, a dynamic perspective interleaving module and a time series compression output module; The time step encoding module receives behavior trajectory data arranged in time sequence, embeds the behavior type, operator identity, behavior occurrence time and terminal environment information in each behavior record respectively, outputs four equal-length vectors and splices them in field order to generate a time step vector sequence; The hierarchical time series modeling module receives the time step vector sequence, inputs three one-dimensional convolution channels in sequence, uses convolution kernel lengths of 3, 5 and 7 respectively, and a sliding step length of 1 to extract local change patterns, splices the outputs of the three convolution channels in the channel dimension to output a behavior feature vector sequence; The dynamic perspective interleaving module receives the behavior feature vector sequence and introduces a multi-view behavior projection mechanism to construct three disturbance perspectives of semantics, time and density; The three perspective tensors and the behavior feature vector sequence are subjected to channel-by-channel weighted summation operation, and the weighted result is subjected to non-linear activation processing using a ReLU activation function to generate a fused behavior representation tensor; The time series compression output module receives the fused behavior representation tensor, performs average pooling processing, and inputs it into a two-layer fully connected network, the first fully connected network performs dimension compression, and the second fully connected network uses a hyperbolic tangent activation function for non-linear transformation to generate a behavior residual vector.

5. The method of claim 4, wherein the machine learning-based automatic classification of archival data is based on a plurality of machine learning algorithms. The multi-view behavior projection mechanism specifically comprises a semantic rearrangement perspective construction method, a time disturbance projection method and a density aggregation projection method; The semantic rearrangement perspective construction method calculates the difference between each time step behavior feature vector and the overall average vector of the behavior feature vector sequence in the embedding space, subtracts the numerical values at corresponding positions in each dimension of the two vectors, takes the absolute value and sums them to obtain a difference value; the time steps are reordered from small to large according to the difference value, and the rearranged behavior feature vector sequence constitutes a semantic rearrangement perspective tensor; The time disturbance projection mode counts the time difference between each two adjacent time steps, calculates the average time interval of the entire behavior feature vector sequence, and sets twice the average time interval as a time segmentation threshold; when the time difference between adjacent time steps is less than the time segmentation threshold, it is determined as the same time period, and when the time difference between adjacent time steps is greater than the time segmentation threshold, it is divided into a new time period; The behavior feature vectors of each time period are summed and divided by the number of behavior feature vectors of the time period to obtain a time period average vector, which is arranged in time sequence to form a time period average vector sequence as a time disturbance perspective tensor; The density aggregation projection mode counts the occurrence times of each time step in the behavior type of the entire behavior feature vector sequence to obtain the frequency of each type of behavior; the proportion of the behavior type of each time step in the total frequency is taken as a weight factor, and the weight factor is multiplied by the behavior feature vector of the corresponding time step by element by element to obtain a weighted vector. The weighted vectors are summed and divided by the total number of time steps to obtain the behavior feature vector after density aggregation, and the behavior feature vector after density aggregation is copied as a tensor with the same length as the behavior feature vector sequence to form a frequency perspective tensor.

6. The method of claim 1, wherein the method is based on machine learning. The step five is specifically: According to all the archive representation vectors, a semantic graph structure is constructed, each archive representation vector is taken as a graph node, the cosine similarity between any two archive representation vectors is calculated, if the cosine similarity exceeds a set similarity threshold, an edge connection is established between the corresponding nodes, and the cosine similarity is taken as an edge weight value to form an adjacency matrix; The semantic graph structure is input into a graph neural network model, the graph neural network includes multiple layers of graph convolution layers, each layer of graph convolution layer receives the node embedding representation of the previous layer and the adjacency matrix, and updates the node embedding representation of each node through a neighborhood feature aggregation operation; The neighborhood feature aggregation operation performs weighted summation and splicing on the target node embedding representation and the direct connected neighbor node representation to obtain an intermediate embedding representation, and inputs the intermediate embedding representation into a ReLU activation function for nonlinear mapping to generate an updated node embedding representation; The node embedding representation output by the last layer of graph convolution layer is input into a fully connected layer and connected to a Softmax classifier to obtain the preliminary classification label and confidence of each node corresponding to the preliminary classification result.

7. The method of claim 1, wherein the method is based on machine learning. The step seven is specifically: Based on the disturbance signal intensity matrix, a class pair with a disturbance signal cumulative intensity greater than a preset intensity threshold is selected as a frequently confused class pair; From the training samples, archive samples with true labels as two different classes in the frequently confused class pair are selected respectively to obtain the archive representation vector of each archive sample, and the archive representation vector of each archive sample in one of the classes is taken as a reference to find the archive representation vector with the largest cosine similarity in the other class as a matching object to form a mirror sample pair; The mirror pair loss value is calculated for each mirror sample pair. The mirror pair loss function is composed of two parts, namely, a label cross-entropy loss and a contrast preservation loss. The label cross-entropy loss is calculated by calculating the cross-entropy value between the predicted probability distribution of each archive sample and the true label, and then averaging the cross-entropy values of the two archive samples to obtain the label cross-entropy loss value. The contrast preservation loss is calculated by calculating the square sum of the difference between the numerical values of each dimension position in the probability distribution vectors of the two archive samples to obtain the contrast preservation loss value. The cross-entropy loss value and the contrast preservation loss value are added to obtain the mirror pair loss value. The mirror pair loss value is input into the graph neural network training process, and the error back propagation algorithm is executed to train the graph neural network parameters according to the gradient direction corresponding to the mirror pair loss value, thereby obtaining the trained graph neural network. 8.The method of claim 1, wherein, The step eight is specifically: The archive representation vector of each archive to be classified is input into the trained graph neural network. The node embedding representation output by the graph neural network is mapped to a classification space through a fully connected layer, and a Softmax classifier is used to calculate the normalized probability of each category as the confidence probability distribution of the archive belonging to each category. The classification label of the archive is determined according to the index corresponding to the category with the maximum value in the confidence probability distribution, and the maximum probability value is taken as the confidence of the classification result, thereby outputting the final classification result.

9. A machine learning based automatic classification system of archival data, performing a machine learning based automatic classification method of any one of claims 1 to 8, characterized in that, The system comprises the following modules: A data acquisition module is configured to acquire archive ontology data and behavior trajectory data. An ontology feature extraction module is configured to extract text vectors, image vectors, and structure vectors from the archive ontology data. A behavior modeling module is configured to encode the behavior trajectory data into a multi-dimensional time sequence in chronological order, input the multi-dimensional time sequence into an improved TS2Vec model, and introduce a multi-view behavior projection mechanism to output a behavior residual vector. A vector fusion module is configured to splice and fuse the text vectors, image vectors, structure vectors, and behavior residual vectors to generate an archive representation vector. A semantic graph construction module is configured to construct a semantic graph structure based on the archive representation vector and form an adjacency matrix. A graph neural network module is configured to perform node embedding propagation and generate a preliminary classification result. An error identification and disturbance propagation module is configured to identify misclassified samples in the preliminary classification result, construct a misclassification pulse path, propagate a disturbance signal along the misclassification pulse path in the semantic graph structure, and generate a disturbance signal strength matrix. Specifically: The node samples with a confidence lower than a preset confidence threshold in the preliminary classification result are compared with actual labels to filter out nodes with inconsistent actual labels and predicted labels as misclassified samples. For each misclassified sample node, a misclassification pulse path is constructed based on a real category center node and a predicted category center node in the semantic graph structure. The misclassification pulse path is a disturbance signal propagation path that starts from the misclassified sample node, sequentially connects the predicted category center node, all intermediate propagation nodes on the shortest connection path between the predicted category center node and the real category center node, and the real category center node, and the intermediate propagation nodes are selected in the adjacency matrix through a Dijkstra algorithm. The disturbance signal is injected into each misclassified sample node in the semantic graph structure, the initial strength of the disturbance signal is set to 1, the disturbance signal propagates along the misclassified pulse path hop by hop, and the current disturbance signal strength is multiplied by the edge weight value between the previous hop node and the current node as the new disturbance signal strength in each hop propagation process until the end of the misclassified pulse path is reached; The records of the disturbance signal propagation in all misclassified pulse paths are counted, and a disturbance signal strength matrix is constructed based on the correspondence between the final propagated true class label and the initial predicted class label. Each matrix element in the disturbance signal strength matrix represents the total cumulative received disturbance signal strength between the class pair. The mirror confrontation training module is used to select the frequently confused class pair based on the disturbance signal strength matrix, construct the mirror sample pair, apply the mirror confrontation loss function based on the mirror sample pair for training, and obtain the trained graph neural network. The classification inference module is used to input the archive representation vector of each archive to be classified into the trained graph neural network, generate a classification label and a classification confidence, and output a final classification result.

Images