A fuzzy cognitive graph neural network modeling method and device for multivariate time series data classification

Abstract

The application belongs to the technical field of data processing, and proposes a fuzzy cognitive graph neural network modeling method and device for multivariate time series data classification. Specifically, multivariate time series intensive care data and corresponding label data are collected, and a representation learning model is applied to extract multivariate time series representation data with distinguishability and representativeness. Different length time windows are used to divide the representation data, and global maximum pooling operation is used to obtain corresponding instance-level representation data. A similarity-based scale screening principle is used to determine a multi-scale representation data set for integrated modeling. For the screening results, the feature importance is quantitatively identified and reordered using the extreme gradient boosting decision tree. The reordered data is divided into multiple feature subsets, a hierarchical width network model is established, and classification prediction under the corresponding scale is realized. The classification prediction results of all base models are fused using a weighted average method to obtain the final decision output.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically the field of artificial intelligence (artificial neural network) pattern data fitting, recognition, and classification. It relates to a fuzzy cognitive map (FCM) neural network modeling method and apparatus for multivariate time-series data classification. Background Technology

[0002] Multivariate time-series data analysis and classification based on artificial intelligence methods can provide relatively sufficient feature correlation and reference information. Multivariate time-series data is characterized by complex distribution, diverse correlation information, and diverse data sources, placing certain requirements on the artificial intelligence model used to perform classification tasks. This invention proposes a modeling method for fuzzy cognitive graph neural networks to achieve feature correlation analysis and classification processing of time-series data, while extracting semantic information from the classification results to achieve the purpose of assisting classification.

[0003] Fuzzy cognitive graphs are models used to solve the problems of modeling and analyzing complex systems. Each feature in the system is represented as a concept node, and the relationships between concept nodes are represented by directed edges, with the weights of the edges indicating the degree of influence of one concept node on another. Compared to traditional neural network models, fuzzy cognitive graphs can describe the mutual influence and correlation between concept nodes at different times, enabling a more comprehensive description and understanding of the characteristics and behaviors of complex models. Therefore, they have found wide application in interpretability analysis in machine learning and artificial intelligence.

[0004] Long Short-Term Memory (LSTM) networks, a special type of neural network, offer a feasible method for processing complex, high-dimensional, and large-scale time-series data. LTM networks possess powerful memory capabilities, enabling them to more effectively handle long-term dependencies. Their core components are cell states and a three-gate structure: cell states... Used for storing and transmitting information, it is an internal state spanning the entire sequence and updated through iterations of time steps; the three gate structures are the forget gate. Input gate and output gate Through a unique gate structure, it selectively forgets and retains the input sequence data, thereby achieving control and regulation of information.

[0005] The application of multivariate time-series data is becoming increasingly widespread, resulting in diverse data distributions, rich correlations, and a wide range of data sources. These characteristics make multivariate time-series data complex, high-dimensional, and massive in scale. For this type of data, different features interact with each other, exhibiting complex correlations and temporal patterns, posing significant challenges to data analysis and processing. Faced with such complex datasets, there is an urgent need to effectively analyze data, perform classification tasks, and extract dynamic patterns. By combining advanced technologies such as deep learning and machine learning, we can better uncover the key features and patterns hidden behind the data, providing more accurate and reliable support for decision-making in various fields. Therefore, how to fully utilize multivariate time-series data and apply artificial intelligence technologies for effective analysis and application has become an urgent task for current research and practice. Summary of the Invention

[0006] The purpose of this invention is to provide a fuzzy cognitive graph neural network modeling method for time series data classification. While ensuring the generalization ability and modeling performance of the neural network, it analyzes the correlation of concept nodes at different times through the concept node temporal weight learning strategy established by the fuzzy cognitive graph, realizes the classification prediction of multivariate time series data, and improves the interpretability of the model.

[0007] The technical solution of the present invention: A fuzzy cognitive graph neural network modeling method for multivariate time-series data classification includes the following steps: S1. Collect data and process the data to form a multivariate time series data sample set; Collect multivariate time series data to form a multivariate time series data sample set, and represent the input sequence of the data sample set as follows: It contains feature information describing multivariate time series; it represents the output sequence of the data sample set as... It includes label information describing the judgment of multivariate time-series data; among which... For the number of samples, The duration of characteristic observations in a multivariate time series. The feature dimension of the sample. For the number of tags; S2. Based on a multivariate time series data sample set, dynamic time warping is used to perform similarity analysis on time series features. Based on the analysis results, the time series features are divided into multiple time series feature groups. Dynamic time warping is used to perform similarity analysis on time series features of a multivariate time series data sample set. For a given multivariate time series data... , , , Indicates the first The first sample The feature in the first The eigenvalues ​​at time t, where For the number of samples, The duration of characteristic observations in a multivariate time series. Let be the feature dimension of the sample, using Indicates the first The first sample The values ​​of each feature at all times; calculate Two univariate features and Similarity between them, time series and DTW distance between The calculation formula is as follows: in Indicates a point in time and The Euclidean distance, and , .

[0008] The above formula can be used to calculate the first... Sample The similarity between all pairwise time series variables in the dataset is determined by setting a threshold. , threshold By comparing the DTW distance with the feature variable sequence, the first... Sample Similarity matrix , , Indicates the first In the nth sample The first feature and the second The specific formula for calculating whether features are similar is as follows: If the first In the nth sample The first feature and the second Similar features Then, for the features of this sample and characteristics If it is classified into a class, if the first In the nth sample The first feature and the second The features are not similar, that is Then, for the features of this sample and characteristics They are divided into different categories.

[0009] Similarly, all samples were clustered using the DTW algorithm, containing... Cumulative similarity matrix of multivariate time series data for each sample , , This indicates the result after comprehensively considering the feature similarity of all samples. Whether a feature is similar to other features, specifically, Indicates the first The first feature and the second Whether each feature is similar, accumulate the similarity matrix. The specific calculations are as follows: Subsequently, a hierarchical clustering method was used to analyze the accumulated similarity matrix. Clustering is performed on the various features in the dataset to obtain feature groups suitable for the entire dataset. First, the accumulated similarity matrix is... Each feature is treated as a separate cluster, forming the initial clustering group. Then, the similarity between each cluster is calculated using Euclidean distance, and the clusters with the highest similarity are merged. At the same time, the distances between all features of two clusters are calculated, and the average of these distances is used as the distance between the two clusters to update the similarity matrix. By repeatedly calculating, the number of clusters is gradually reduced to form a hierarchical structure.

[0010] S3. Group by features and define the duration of feature effect. Concept nodes are generated through a temporal fuzzy cognitive graph; Based on the partitioning results of the multivariate time series data, the original multivariate time series data sample set was... After feature similarity analysis, they were divided into Group After feature segmentation, new concept nodes are constructed by assigning different weights to different features. The weights reflect the contribution of each original feature to the concept node. Furthermore, since time series data are interconnected and influence each other, it is necessary to consider past data to reveal the relationships between features at different times. The influence of each concept node at each time step on the concept node generated at the current time step. Assume that the cross-time dimension interactions between sequences continue. The expression for a concept node with time memory at any given moment is as follows: in Indicates sample In the At the [time]th moment Concept nodes generated in each group Indicates the generation of concept nodes The weight matrix is ​​composed of the importance of the grouping features of the original data. , Indicates sample In the At the [time]th moment The original data in each group, Indicates the generation of samples Concept nodes hour, Concept node before time interval The importance weight of the concept node at the current moment, The function is the sigmoid function, which maps the output range to [0, 1], converting the predicted value into a probability. S4. Construct a bidirectional long short-term memory network with peephole connection, use the updated concept nodes as the input of the model, and use the loss function to train the weights to form a learning model with temporal weights of concept nodes. A classification model is designed using a bidirectional long short-term memory (LSTM) network capable of processing both past and future information, and a peephole-type LSM network capable of capturing long-term dependencies. This model introduces peephole connections to the traditional LSM network model, incorporating the cell state from the previous time step into the forget gate and input gate, enhancing the retention of historical information. Simultaneously, the output gate receives the updated cell state, further improving information transfer efficiency. Furthermore, for multivariate time-series data where long-term dependencies may exist between features, subsequent data is used as another input to more deeply explore the interactions between different features. Therefore, a bidirectional approach is introduced, using both forward and backward calculations to predict the output at each time step. The forward layer performs forward computation from time 1 to time t, and the result is saved as the output of the forward hidden layer. The backward layer performs backward computation from time t to time 1, and the result is saved as the output of the backward hidden layer. Finally, the final output at each time step is obtained by weighting the outputs of the forward and backward layers. Concept nodes... The inputs to the model are fed into the forward and reverse layers respectively, and the final prediction result is obtained by weighting the outputs of the two layers. Taking the forward layer as an example, the forget gate in the forward layer functions by using the forward input at the current time step. Output of the previous time step in the forward layer and the cell state at the previous moment in the positive layer The function determines which information to forget from the cell state, and uses the sigmoid function to control the output of the forget gate within the interval [0,1]. The function expression for the forward forget gate is: in, , This represents the weights and biases that need to be trained in the forward forgetting gate. For sigmoid activation function, These are the weights between the input layer and the forward layer. The bias between the input layer and the forward layer; The function of a forward input gate is to pass the forward input at the current time. The output of the previous time step of the forward layer and the cell state at the previous moment in the positive layer The forward input gate determines whether the current input information is retained in the memory cell, thus controlling the effect of the current input information on the memory cell. The function expression for the forward input gate is as follows: in, , This represents the weights and biases that need to be trained in the forward forgetting gate; At the same time, construct the current cell state. Candidate vectors that need to be added The candidate vector values ​​are mapped to the range [-1, 1] using the hyperbolic tangent function. The specific calculation formula is as follows: in, , This represents the weights and biases that need to be trained when generating positive candidate vectors; Output based on the current time of the forget gate The cell state at the previous moment The product of the input gate and the output of the input gate. and the candidate vector at the current time The product of these can be used to obtain the memory cell at this moment, which is the core of the LSTM structure responsible for storing and transmitting information: The last gate structure is the output gate, which first passes through a... The transformation determines what information is output. , and then combine By transforming the cell state, a definite output is obtained. The function expression for the output gate is: in, , This represents the weights and biases that need to be trained in the forward output gate; In summary, the first The hidden layer output at each time step can be obtained from the input at the current time step. Output of the hidden state at the previous time step and the cell state output at the previous moment Calculated together, using Summary of the first The internal computation process of the peephole LSTM at each moment: Therefore, the reverse layer is... Output at time The calculation formula is as follows: in and Weights between the input layer and the reverse layer Bias.

[0011] There is no connection between the forward and backward layers, and their information transmission directions are different. The calculated states from both layers are passed to the output layer to obtain the final output at the current time. .

[0012] in These are the weights between the feedforward layer and the output. These are the weights between the back layer and the output. For bias; The final predicted value of the classification model is the first... Output at time The error between this error and the true value is calculated using the proposed loss function to minimize the objective function. Training determines the model parameters.

[0013] in It is the focus factor, used to adjust the weight of easy and difficult samples. The larger the value, the more attention is paid to the samples that are difficult to classify. This is a balancing parameter used to balance the weights of positive and negative samples. It is typically set as the ratio between the number of samples in the minority class and the number of samples in the multi-class class. For the sample The final predicted value after classification model, Then it is a sample The true label; use This represents the parameters that need to be trained in the model, including , , , , , , , , , , , , , , , , , , , , , , , , The Adaptive Moment Estimation (Adam) algorithm is used as the parameter optimization algorithm to update the parameters in the classification model.

[0014] S5. Based on time series weights, different feature importance thresholds and time importance thresholds are set to select important time features, and an AFS decision tree is constructed to form dynamic semantic rules for multivariate time series data. Extracting rules during model building helps explain the model's prediction process. In many applications, especially those involving decision-making in human life or health, understanding how the model makes predictions is essential. Rule extraction transforms the complex operation of the model into easily understood rules or patterns, improving the model's interpretability. This patent utilizes an AFS decision tree method for extracting dynamic semantic rules from dynamic multivariate data.

[0015] In multivariate time series data analysis, the most recent data often plays a crucial role in data prediction and intervention. Based on this consideration, it is particularly important to focus on the influence weight of the current-time concept node on subsequent-time concept nodes obtained from the time-series fuzzy cognitive graph, thereby identifying the significant moments and features in the multivariate time series.

[0016] Set rule threshold and ,in For feature importance threshold, The threshold for time-matter importance. The concept node at the most recent time step. By the Input data and past data for each feature group at each time point The concept of time is derived through the combined effect of nodes. For the most recent time point... Concept Nodes If the feature weights satisfy the following conditions, then it indicates that the features at the most recent time are... Right now Selected: in Indicates the first The first concept node The important weights of each feature Indicates the composition of the first The number of features of each concept node.

[0017] If a historical concept node satisfies the following conditions, then it indicates that the concept node... In the The value at time is Selected: in Indicates the generation of samples Concept nodes hour, Concept node before time interval The importance weight of the concept node at the current moment, Indicates the duration of the characteristic effect.

[0018] For the selected historical concept nodes The current moment feature value and historical node that meet the threshold can still be selected according to the above formula. This process is iterated until no new historical concept node is selected.

[0019] Therefore, the features of the conceptual nodes that influence the most recent moment are: One, construct rule features ,feature Each factor contains a feature value of time-series information, that is... The meaning is a certain feature value at a certain moment, therefore, according to the rule features From the dataset A new feature set is obtained from this to extract rules. and sample set .

[0020] in Indicates the first One sample, The regular features representing this sample. This represents the first sample. Each rule feature The class label representing this sample. Represents the dataset The number of samples in the sample.

[0021] Before constructing the decision tree, continuous attributes are fuzzified, that is, conceptualized, to obtain semantic information. This is done for rule features. The simple concepts of "small," "medium," and "large" as characteristics are represented as follows: The set of simple concepts after fuzzing all continuous features. It can be represented as: feature Belongs to the The expression for the membership degree of a simple concept is as follows: in Indicates sample Features Values, Representation of features The maximum value obtained among all samples, Representation of features The minimum value obtained among all samples, Representation of features The mean is obtained from all samples. Representation of features The largest difference between it and its average level is used to measure the dispersion or deviation of the data. Representation of features The smallest difference between the data and its average level is used to measure the concentration or stability of the data.

[0022] By logically connecting multiple single attributes, more complex attributes can be formed, and the AFS logical operation rules can be used to generate a set of complex concepts. Its definition is as follows: in, A non-empty set of indices for a simple concept. express The extraction, Indicates all The conjunction of .

[0023] A consistent membership function can be obtained based on simple concepts and the AFS structure. Assume... It is a complex concept, and its membership function Represented as: in, For a simple concept The weighting function can be determined based on the data distribution. It is the number of times the sample was observed.

[0024] The core of constructing an AFS decision tree lies in consistently selecting the node with the highest information gain for splitting, thereby maximizing the growth rate of information in complex concepts. Split the current node Find the most suitable combination of simple concepts to ensure that the nodes after splitting are Compared with the node before splitting The information gain difference is the largest, that is: in, Indicates the current node Information gain For the current node The amount of information, for Split the current node The weighted information content is then used to obtain a simple concept for segmenting the current node. ,in and The calculation method is as follows: in, , For the first N In layer nodes membership degree , Indicates the first The class outputs a set of samples. Indicates when fuzzy concepts Split the current node At that time, it belonged to the first Consistent membership degree of class samples.

[0025] Therefore, we can obtain the following: Maximizing the fuzzy concept, from Delete ,Right now: If the consistency membership function of a fuzzy concept satisfies , If the set threshold is not met, then the AFS logical operation is performed. "Construct new tree nodes" If the information gain of the current node is negative, that is... Or in fuzzy attributes Split If the obtained subset is an empty set, Stop expanding child nodes, and you get the AFS decision tree. At this point, the fuzzy attribute from the root node to the leaf node is a semantic rule. .

[0026] Table 1 Confusion Matrix Tab. 1 Confusion matrix

[0027] The designed model is primarily intended for performing classification tasks and extracting dynamic rules from multivariate time-series data. Therefore, to evaluate the performance of the binary classification algorithm, TP represents the number of samples where both predicted and true values ​​are positive, TN represents the number of samples where both predicted and true values ​​are negative, FN represents the number of samples where the predicted value is negative and the true value is positive, and FP represents the number of samples where the predicted value is positive and the true value is negative. Table 1 shows the confusion matrix for different attributes of the predicted and true values. Only these four scenarios will occur in a binary classification problem, therefore the following metrics are defined based on these four results.

[0028] (1) Accuracy: The percentage of correct predictions.

[0029] (2) Precision: The proportion of samples that are actually positive and are also predicted to be positive out of all samples that are predicted to be positive.

[0030] (3) Recall: The proportion of samples that are actually positive and are also predicted to be positive out of all samples that are actually positive. The purpose of recall is to ensure that all positive samples are predicted as much as possible, and it is often used in medical data classification.

[0031] (4) F1-score: the harmonic mean of precision and recall, used to comprehensively evaluate the performance of the model.

[0032] (5) ROC curve and AUROC The ROC (Receiver Operating Characteristic Curve) is plotted with the False Positive Rate (FPR) on the horizontal axis and the True Positive Rate (TPR) on the vertical axis. The ROC curve is a visualization of model performance at different classification thresholds. AUROC (Area Under the ROC Curve) is the area under the ROC curve, used to quantify the model's classification performance. A higher AUROC value indicates better model performance.

[0033] (6) PRC curve and AUPRC The PRC curve is plotted with recall on the horizontal axis and precision on the vertical axis. The closer the PRC curve is to the upper right corner, the better the model performance. AUPRC (Area Under the Precision-Recall Curve) is the area under the PRC curve, used to comprehensively evaluate the model's average precision at different recall rates. The higher the AUPRC value, the better the model performance. Unlike AUC, AUPRC is more suitable for handling imbalanced datasets and tasks that emphasize positive example recognition.

[0034] The present invention also provides an apparatus for multivariate time series data analysis and modeling, comprising: The acquisition unit is used to acquire data and process the data to form multi-dimensional time-series data samples.

[0035] The first construction unit is used for multivariate time series data sample sets. It uses dynamic time warping to perform similarity analysis on time series features, and divides the features based on the analysis results to form multiple feature groups.

[0036] The second construction unit is used to generate concept nodes based on feature grouping and to define the duration of feature effects. Concept nodes are updated through a temporal fuzzy cognitive graph; The third building unit is used to construct a bidirectional long short-term memory network with peephole connection. The updated concept nodes are used as the input of the model, and the weights are trained using the loss function to form a learning model with temporal weights of concept nodes. The rule extraction unit, based on temporal weights, sets different feature importance thresholds and time importance thresholds to select important time features, constructs an AFS decision tree, and thus forms dynamic semantic rules for multivariate time-series data.

[0037] The beneficial effects of this invention are as follows: This invention proposes a novel modeling method for a variant of long short-term memory networks based on temporal fuzzy cognitive graphs. This method can achieve classification tasks and dynamic rule extraction through temporal feature similarity analysis, feature learning, and selection of important features from multivariate time-series data. Addressing the characteristics of multivariate time-series data, such as complex distribution, diverse associated data, and multiple data channels, it forms a complex, high-dimensional, and large-scale sample set. The proposed variant of the Long Short-Term Memory (LSTM) network based on a temporal fuzzy cognitive graph can group multidimensional temporal features through temporal feature similarity analysis, reducing redundant features, improving model efficiency and performance, and generating input for subsequent models. Features from different groups are fused into novel concept nodes, and the concept nodes are updated using the temporal weights of the temporal fuzzy cognitive graph. This aims to reduce the size of the fuzzy cognitive graph, the number of concept nodes, and the computational complexity of processing, better utilize the correlation and similarity between features, more clearly understand the relationships and interactions between similar features, and enhance the interpretability of model prediction results. The updated concept nodes are used as input to a bidirectional LSM network with peephole connections, inputting into both the forward and reverse layers. The final prediction result is obtained by weighting the outputs of the forward and reverse layers. The classification accuracy and temporal concept node relevance weights of the model are trained using the S-FL loss function. Meanwhile, this method incorporates the influence weights of the current time-to-time concept nodes on subsequent time-to-time concept nodes, thereby filtering out significant moments and features in multivariate time series. By extracting rules, the complex operation process of the model can be transformed into easily understandable rules, which can improve the interpretability of multivariate time series data classification models and provide a feasible solution for multivariate time series data classification and model interpretability. Attached Figure Description

[0038] Figure 1 This is a flowchart of a fuzzy cognitive graph neural modeling method for multivariate time-series data classification provided by the present invention; Figure 2 This is a schematic diagram of a variant of a long short-term memory network based on a temporal fuzzy cognitive graph provided by the present invention; Figure 3 This invention provides a dynamic rule extraction process based on AFS decision trees; Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0040] In view of this, embodiments of the present invention provide a fuzzy cognitive graph neural network modeling method for sepsis data classification. Please refer to... Figure 1 The flowchart of a fuzzy cognitive graph neural network provided in this embodiment of the invention includes: S1. Collect data and process the data to form a sepsis data sample set; S2. Based on the sepsis data sample set, similarity analysis of time series features is performed using dynamic time warping. Based on the analysis results, the features are divided into multiple feature groups. S3. Generate concept nodes based on feature grouping and define the duration of feature effect. Concept nodes are updated through a temporal fuzzy cognitive graph; S4. Construct a bidirectional long short-term memory network with peephole connection, use the updated concept nodes as the input of the model, and use the loss function to train the weights to form a learning model with temporal weights of concept nodes. S5. Based on temporal weights, different feature importance thresholds and time importance thresholds are set to select important time features and construct an AFS decision tree, thereby forming dynamic semantic rules for sepsis data.

[0041] Specifically, S1 includes: Take the German sepsis dataset Sepsisexp as an example. This dataset was collected based on the data management system running at the University Medical Center Mannheim (UMM) in Germany. It extracted 43 features from the intensive care and anesthesia systems: respiratory volume per minute (MVV), platelet count (PLT), blood glucose (GLU), heart rate (HR), Horowitz index (HOR), base excess (BE), white blood cell count (WBC), blood urea / creatinine ratio (BUN / CR), chloride ion (CL), body temperature (TEMP), temperature difference (TEMPDIFF), calcium (CA), partial pressure of carbon dioxide (PaCO2), lactate (LAC), potassium (K), respiratory rate (RR), bicarbonate ion (HCO3), mixed venous oxygen saturation (SvO2), and arterial pH. The following parameters were used to measure the patient's ICU stay: pH, C-reactive protein (CRP), urine volume (VOL), bilirubin (BIL), hemoglobin (HB), blood urea nitrogen (BUN), heart volume per minute (CI), alanine aminotransferase (ALT), creatinine (CREA), lymphocytes (LYM), aspartate aminotransferase (AST), diastolic blood pressure (DBP), sodium (NA), ejection fraction (EF), fractional oxygen inhaled (FiO2), pancreatic lipase (AMYL), ejection resistance index (SVRI), mean arterial pressure (MAP), procalcitonin (PT), and age. The timeline of patients during their ICU stay was discretized into uniform 30-minute steps. All values ​​were standardized using z-scores and assigned a disease status label (0 or 1). Four training files were provided in the database, each containing 319 patients: 64 with disease and 255 without disease, resulting in an imbalance rate of 3.98%. A data sample set is formed by collecting biochemical indicator test data related to sepsis. The input sequence of the data sample set is represented as follows: It contains characteristic information describing sepsis, and the output sequence is represented as follows: It contains labeling information describing the diagnosis of sepsis; among which For the number of samples, The duration of characteristic observations in a multivariate time series. The feature dimension of the sample. This represents the number of tags.

[0042] Furthermore, S2 is used to perform similarity analysis on time series features. Based on the analysis results, the features are divided into multiple feature groups, specifically including: Based on the sepsis data sample set described in S1, dynamic time warping is used to perform similarity analysis on time series features for a given multivariate time series data. , , , Indicates the first The first sample The feature in the first The eigenvalues ​​at time t, where For the number of samples, The duration of characteristic observations in a multivariate time series. Let be the feature dimension of the sample, using Indicates the first The first sample The values ​​of each feature at all times. Calculate Two univariate features and Similarity between them, time series and DTW distance between The calculation formula is as follows.

[0043] in Indicates a point in time and The Euclidean distance, and , .

[0044] The above can be used to calculate the first... Sample The similarity between all pairwise time series variables in the dataset is determined by setting a threshold. , threshold By comparing the DTW distance with the feature variable sequence, the first... Sample Similarity matrix , , Indicates the first In the nth sample The first feature and the second The specific formula for calculating whether features are similar is as follows: If the first In the nth sample The first feature and the second Similar features Then, for the features of this sample and characteristics If it is classified into a class, if the first In the nth sample The first feature and the second The features are not similar, that is Then, for the features of this sample and characteristics They are divided into different categories.

[0045] Similarly, all samples were clustered using the DTW algorithm, containing... Cumulative similarity matrix of multivariate time series data for each sample , , This indicates the result after comprehensively considering the feature similarity of all samples. Whether a feature is similar to other features, specifically, Indicates the first The first feature and the second Whether each feature is similar, accumulate the similarity matrix. The specific calculations are as follows: Subsequently, a hierarchical clustering method was used to analyze the accumulated similarity matrix. Clustering is performed on the various features in the dataset to obtain feature groups suitable for the entire dataset. First, the accumulated similarity matrix is... Each feature is treated as a separate cluster, forming the initial clustering group. Then, the similarity between each cluster is calculated using Euclidean distance, and the clusters with the highest similarity are merged. At the same time, the distances between all features of two clusters are calculated, and the average of these distances is used as the distance between the two clusters to update the similarity matrix. By repeatedly calculating, the number of clusters is gradually reduced to form a hierarchical structure.

[0046] Furthermore, concept nodes are generated by grouping based on features using S3, and the duration of feature effects is defined. Updating concept nodes through a temporal fuzzy cognitive graph includes: Based on the partitioning results of the sepsis dataset obtained in S2, the original multivariate time-series data sample set was... After feature similarity analysis, they were divided into Group After feature segmentation, new concept nodes are constructed by assigning different weights to different features. The weights reflect the contribution of each original feature to the concept node. Furthermore, since time series data are interconnected and influence each other, it is necessary to consider past data to reveal the relationships between features at different times. The influence of each concept node at each time step on the concept node generated at the current time step. Assume that the cross-time dimension interactions between sequences continue. The expression for a concept node with time memory at any given moment is as follows: in Indicates sample In the At the [time]th moment Concept nodes generated in each group Indicates the generation of concept nodes The weight matrix is ​​composed of the importance of the grouping features of the original data. , Indicates sample In the At the [time]th moment The original data in each group, Indicates the generation of samples Concept nodes hour, Concept node before time interval The importance weight of the concept node at the current moment, The function is the sigmoid function, which maps the output range to [0, 1], converting the predicted value into a probability. Furthermore, a bidirectional long short-term memory network with peephole connectivity is constructed using S4. The updated concept nodes are used as input to the model, and the weights are trained using a loss function to form a learning model with temporal weights for concept nodes. For detailed procedures, please refer to [link to documentation]. Figure 2 Specifically, it includes: According to S3, the updated concept nodes As input to the model, a classification model is designed using a bidirectional long short-term memory (LSTM) network capable of simultaneously processing past and future information and a peephole-type LSM network capable of capturing long-term dependencies. This model introduces peephole connections to the traditional LSM network model, incorporating the cell state from the previous time step into the forget gate and input gate, enhancing the memory of historical information. Simultaneously, the output gate receives the updated cell state, further improving the efficiency of information transmission. Furthermore, for multivariate time-series data where long-term dependencies may exist between features, subsequent data is used as another input to more deeply explore the interactions between different features. Therefore, a bidirectional approach is introduced, using both forward and backward calculations to predict the output at each time step. The forward layer performs forward computation from time 1 to time t, and the result is saved as the output of the forward hidden layer. The backward layer performs backward computation from time t to time 1, and the result is saved as the output of the backward hidden layer. Finally, the final output at each time step is obtained by weighting the outputs of the forward and backward layers. Concept nodes... The inputs to the model are fed into the forward and reverse layers respectively, and the final prediction result is obtained by weighting the outputs of the two layers. Taking the forward layer as an example, the forget gate in the forward layer functions by using the forward input at the current time step. Output of the previous time step in the forward layer and the cell state at the previous moment in the positive layer The function determines which information to forget from the cell state, and uses the sigmoid function to control the output of the forget gate within the interval [0,1]. The function expression for the forward forget gate is: in, , This represents the weights and biases that need to be trained in the forward forgetting gate. For sigmoid activation function, These are the weights between the input layer and the forward layer. The bias between the input layer and the forward layer; The function of a forward input gate is to pass the forward input at the current time. The output of the previous time step of the forward layer and the cell state at the previous moment in the positive layer The forward input gate determines whether the current input information is retained in the memory cell, thus controlling the effect of the current input information on the memory cell. The function expression for the forward input gate is as follows: in, , This represents the weights and biases that need to be trained in the forward forgetting gate; At the same time, construct the current cell state. Candidate vectors that need to be added The candidate vector values ​​are mapped to the range [-1, 1] using the hyperbolic tangent function. The specific calculation formula is as follows: in, , This represents the weights and biases that need to be trained when generating positive candidate vectors; Output based on the current time of the forget gate The cell state at the previous moment The product of the input gate and the output of the input gate. and the candidate vector at the current time The product of these can be used to obtain the memory cell at this moment, which is the core of the LSTM structure responsible for storing and transmitting information: The last gate structure is the output gate, which first passes through a... The transformation determines what information is output. , and then combine By transforming the cell state, a definite output is obtained. The function expression for the output gate is: in, , This represents the weights and biases that need to be trained in the forward output gate; In summary, the first The hidden layer output at each time step can be obtained from the input at the current time step. Output of the hidden state at the previous time step and the cell state output at the previous moment Calculated together, using Summary of the first The internal computation process of the peephole LSTM at each moment: Therefore, the reverse layer is... Output at time The calculation formula is as follows: in and Weights between the input layer and the reverse layer Bias.

[0047] There is no connection between the forward and backward layers, and their information transmission directions are different. The calculated states from both layers are passed to the output layer to obtain the final output at the current time. .

[0048] in These are the weights between the feedforward layer and the output. These are the weights between the back layer and the output. For bias; The final predicted value of the classification model is the first... Output at time The error between this error and the true value is calculated using the proposed loss function to minimize the objective function. Training determines the model parameters.

[0049] in It is the focus factor, used to adjust the weight of easy and difficult samples. The larger the value, the more attention is paid to the samples that are difficult to classify. This is a balancing parameter used to balance the weights of positive and negative samples. It is typically set as the ratio between the number of samples in the minority class and the number of samples in the multi-class class. For the sample The final predicted value after classification model, Then it is a sample The true label; use This represents the parameters that need to be trained in the model, including , , , , , , , , , , , , , , , , , , , , , , , , The Adaptive Moment Estimation (Adam) algorithm is used as the parameter optimization algorithm to update the parameters in the classification model.

[0050] Specifically, the main steps of the Adam algorithm are as follows: (1) Initialize parameters: First, initialize the model parameters. At the same time, the initial first moment estimate and second moment estimates All are set to 0; (2) Calculate the gradient: Backpropagation of the loss function to the parameters is one of the key steps in optimizing the model parameters. The chain rule is used to calculate the gradient of the loss function with respect to the model parameters. ; in Indicates the first Model parameters before the next iteration. Indicates in the parameter The loss function is as follows.

[0051] (3) Calculate the first moment estimate and second moment estimates Calculate the first moment estimate of the gradient of each parameter. and second moment estimates ; in Indicates the current iteration step. The coefficients represent the partial first moment estimates. The coefficients representing the partial second moment estimates are... and The values ​​are all between 0 and 1. Represents the gradient of the parameters; (4) Update parameters: update the unbiased first-order matrix values ​​of the parameters. and unbiased second-order matrix values The calculations are performed to update the model's training parameters, dynamically adjust the learning rate, and update each parameter.

[0052] in This is a constant set to avoid the denominator being 0. If it is the learning rate, then the first... The parameters are updated in the next iteration. .

[0053] (5) Repeat steps 2 to 4 until the set number of iterations is reached. The weights will no longer be updated, and model training will end.

[0054] In this way, the Adam algorithm can adaptively adjust the learning rate during training, which helps to accelerate convergence and improve the training effect of the model. The designed S-FL loss function and the Adam algorithm can ultimately yield the classification results. and the parameters that need to be trained in the model .

[0055] Furthermore, based on S5 and utilizing temporal weights, different feature importance thresholds and time importance thresholds are set to select important time-time features, and an AFS decision tree is constructed to form dynamic semantic rules for sepsis data, as specifically described below: Set rule threshold and ,in For feature importance threshold, The threshold for time-matter importance. The concept node at the most recent time step. By the Input data and past data for each feature group at each time point The concept of time is derived through the combined effect of nodes. For the most recent time point... Concept Nodes If the feature weights satisfy the following conditions, then it indicates that the features at the most recent time are... Right now Selected: in Indicates the first The first concept node The important weights of each feature Indicates the composition of the first The number of features of each concept node.

[0056] If a historical concept node satisfies the following conditions, then it indicates that the concept node... In the The value at time is Selected: in Indicates the generation of samples Concept nodes hour, Concept node before time interval The importance weight of the concept node at the current moment, Indicates the duration of the characteristic effect.

[0057] For the selected historical concept nodes The current moment feature value and historical node that meet the threshold can still be selected according to the above formula. This process is iterated until no new historical concept node is selected.

[0058] Therefore, the features of the conceptual nodes that influence the most recent moment are: One, construct rule features ,feature Each factor contains a feature value of time-series information, that is... The meaning is a certain feature value at a certain moment, therefore, according to the rule features From the dataset A new feature set is obtained from this to extract rules. and sample set .

[0059] in Indicates the first One sample, The regular features representing this sample. This represents the first sample. Each rule feature The class label representing this sample. Represents the dataset The number of samples in the sample.

[0060] Before constructing the decision tree, continuous attributes are fuzzified, that is, conceptualized, to obtain semantic information. This is done for rule features. The simple concepts of "small," "medium," and "large" as characteristics are represented as follows: The set of simple concepts after fuzzing all continuous features. It can be represented as: feature Belongs to the The expression for the membership degree of a simple concept is as follows: in Indicates sample Features Values, Representation of features The maximum value obtained among all samples, Representation of features The minimum value obtained among all samples, Representation of features The mean is obtained from all samples. Representation of features The largest difference between it and its average level is used to measure the dispersion or deviation of the data. Representation of features The smallest difference between the data and its average level is used to measure the concentration or stability of the data.

[0061] By logically connecting multiple single attributes, more complex attributes can be formed, and the AFS logical operation rules can be used to generate a set of complex concepts. Its definition is as follows: in, A non-empty set of indices for a simple concept. express The extraction, Indicates all The conjunction of .

[0062] For example, if the key temporal features selected for sepsis are {the most recent Horowitz index, the most recent systemic vascular resistance index, the most recent calcium level, the potassium levels two moments ago, and the creatinine levels three moments ago}, then the simple concept set is: ={Recent Horowitz index is low, recent Horowitz index is moderate, recent Horowitz index is high, recent systemic vascular resistance index is low, recent systemic vascular resistance index is moderate, recent systemic vascular resistance index is high, recent calcium level is low, recent calcium level is moderate, recent calcium level is high, potassium level two moments ago is low, potassium level two moments ago is moderate, potassium level two moments ago is high, creatinine level three moments ago is low, creatinine level three moments ago is moderate, creatinine level three moments ago is high}, denoted as , in This indicates "patients with a low Horowitz index and a moderate systemic vascular resistance index at the most recent time". This indicates "patients with a low systemic vascular resistance index and moderate calcium levels at recent times," while Here, "+" indicates disjunction. This can refer to "a patient with a low Horowitz index and a moderate systemic vascular resistance index at the most recent time" or "a patient with a low systemic vascular resistance index and a moderate calcium level at the most recent time".

[0063] A consistent membership function can be obtained based on simple concepts and the AFS structure. Assume... It is a complex concept, and its membership function Represented as: in, For a simple concept The weighting function can be determined based on the data distribution. It is the number of times the sample was observed.

[0064] The core of constructing an AFS decision tree lies in consistently selecting the node with the highest information gain for splitting, thereby maximizing the growth rate of information in complex concepts. Split the current node Find the most suitable combination of simple concepts to ensure that the nodes after splitting are Compared with the node before splitting The information gain difference is the largest, that is: in, Indicates the current node Information gain For the current node The amount of information, for Split the current node The weighted information content is then used to obtain a simple concept for segmenting the current node. ,in and The calculation method is as follows: in, , For the first N In layer nodes membership degree , Indicates the first The class outputs a set of samples. Indicates when fuzzy concepts Split the current node At that time, it belonged to the first Consistent membership degree of class samples.

[0065] Therefore, we can obtain the following: Maximizing the fuzzy concept, from Delete ,Right now: If the consistency membership function of a fuzzy concept satisfies , If the set threshold is not met, then the AFS logical operation is performed. "Construct new tree nodes" If the information gain of the current node is negative, that is... Or in fuzzy attributes Split If the obtained subset is an empty set, Stop expanding child nodes, and you get the AFS decision tree. At this point, the fuzzy attribute from the root node to the leaf node is a semantic rule. .

[0066] Please refer to the detailed process. Figure 3 The figure shows a dynamic rule extraction process based on AFS decision tree provided by the present invention.

[0067] The various embodiments of the present invention are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. The focus of each embodiment is to highlight the differences from other embodiments. In particular, the system embodiments are relatively brief because they are basically similar to the method embodiments, and relevant information can be found in the detailed description of the method embodiments.

[0068] It should be clarified that the above description represents only embodiments of the present invention and does not limit the invention. For those skilled in the art, the present invention has many possibilities for modification and variation. Any modifications, equivalent substitutions, improvements, etc., that conform to the spirit and principles of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A fuzzy cognitive graph neural network modeling method for multivariate time-series data classification, characterized in that, Includes the following steps: S1. Collect data and process the data to form a multivariate time series data sample set; Collect multivariate time series data to form a multivariate time series data sample set, and represent the input sequence of the data sample set as follows: It contains feature information describing multivariate time series; it represents the output sequence of the data sample set as... It includes label information describing the judgment of multivariate time-series data; among which... For the number of samples, The duration of characteristic observations in a multivariate time series. The feature dimension of the sample. For the number of tags; S2. Based on a multivariate time series data sample set, dynamic time warping is used to perform similarity analysis on time series features. Based on the analysis results, the time series features are divided into multiple time series feature groups. Dynamic time warping is used to perform similarity analysis on time series features of a multivariate time series data sample set. For a given multivariate time series data... , , , Indicates the first The first sample The feature in the first The eigenvalues ​​at time t, where For the number of samples, The duration of characteristic observations in a multivariate time series. Let be the feature dimension of the sample, using Indicates the first The first sample The values ​​of each feature at all times; calculate Two univariate features and Similarity between them, time series and DTW distance between The calculation formula is as follows: in, Indicates a point in time and The Euclidean distance, and , ; Calculate the first using the above formula. Sample The similarity between all pairwise time series variables in the dataset is determined by setting a threshold. , threshold By comparing the DTW distance with the feature variable sequence, the first... Sample Similarity matrix , , Indicates the first In the nth sample The first feature and the second The specific formula for calculating whether features are similar is as follows: If the first In the nth sample The first feature and the second Similar features Then, for the features of this sample and characteristics If it is classified into a class, if the first In the nth sample The first feature and the second The features are not similar, that is Then, for the features of this sample and characteristics They were divided into different categories; Similarly, all samples were clustered using the DTW algorithm, containing... Cumulative similarity matrix of multivariate time series data for each sample , , This indicates the result after comprehensively considering the feature similarity of all samples. Whether a feature is similar to other features, specifically, Indicates the first The first feature and the second Whether each feature is similar, accumulate the similarity matrix. The specific calculations are as follows: Subsequently, a hierarchical clustering method was used to analyze the accumulated similarity matrix. Clustering is performed on each feature in the dataset to obtain feature groups suitable for the entire dataset; first, the accumulated similarity matrix is... Each feature is treated as a separate cluster, forming the initial clustering group; then, the similarity between each cluster is calculated using Euclidean distance, the clusters with the highest similarity are merged, and the distance between all features between two clusters is calculated. The average of these distances is used as the distance between the two clusters, and the similarity matrix is ​​updated; by repeatedly calculating, the number of clusters is gradually reduced to form a hierarchical structure. S3. Group by features and define the duration of feature effect. Concept nodes are generated through a temporal fuzzy cognitive graph; Based on the partitioning results of the multivariate time series data, the original multivariate time series data sample set was... After feature similarity analysis, they were divided into Group After feature segmentation, new concept nodes are constructed by assigning different weights to different features. The weights reflect the contribution of each original feature to the concept node. Furthermore, since time series data are interconnected and influence each other, to reveal the relationships between features at different times, it is also necessary to consider past data. The influence of each concept node at each time step on the concept node generated at the current time step; assuming that the cross-time dimension interaction between sequences continues. The expression for a concept node with time memory at any given moment is as follows: in, Indicates sample In the At the [time]th moment Concept nodes generated in each group Indicates the generation of concept nodes The weight matrix is ​​composed of the importance of the grouping features of the original data. , Indicates sample In the At the [time]th moment The original data in each group, Indicates the generation of samples Concept nodes hour, Concept node before time interval The importance weight of the concept node at the current moment, The function is the sigmoid function, which maps the output range to [0, 1], converting the predicted value into a probability: S4. Construct a bidirectional long short-term memory network with peephole connection, use the updated concept nodes as the input of the model, and use the loss function to train the weights to form a learning model with temporal weights of concept nodes. A classification model is designed using a bidirectional long short-term memory (LSTM) network capable of processing both past and future information and a peephole-type LSM network capable of capturing long-term dependencies. This model introduces peephole connections to the traditional LSM network model, incorporating the cell state from the previous time step into the forget gate and input gate to enhance the memory of historical information. Simultaneously, the output gate receives the updated cell state, further improving information transmission efficiency. Furthermore, for multivariate time-series data where long-term dependencies may exist between features, subsequent data is used as another input to more deeply explore the interactions between different features. Therefore, a bidirectional approach is introduced, using both forward and backward calculations to predict the output at each time step. The forward layer performs forward computation from time 1 to time t, saving the result as the output of the forward hidden layer. The backward layer performs backward computation from time t to time 1, saving the result as the output of the backward hidden layer. Finally, the final output at each time step is obtained by weighting the outputs of the forward and backward layers. Concept nodes... As input to the model, the forward and reverse layers are fed into each other, and the final prediction result is obtained by weighting the outputs of the two layers. For the forward layer, the forget gate functions by passing the current forward input. Output of the previous time step in the forward layer and the cell state at the previous moment in the positive layer The function determines which information to forget from the cell state, and uses the sigmoid function to control the output of the forget gate within the interval [0,1]. The function expression for the forward forget gate is: in, , This represents the weights and biases that need to be trained in the forward forgetting gate. It is the sigmoid activation function. These are the weights between the input layer and the forward layer. The bias between the input layer and the forward layer; The function of a forward input gate is to pass the forward input at the current time. The output of the previous time step of the forward layer and the cell state at the previous moment in the positive layer This determines whether the current input information is retained in the memory cell, that is, it controls the effect of the information input at this moment on the memory cell; the function expression of the forward input gate is as follows: in, , This represents the weights and biases that need to be trained in the forward forgetting gate; At the same time, construct the current cell state. Candidate vectors that need to be added The candidate vector values ​​are mapped to the range [-1, 1] using the hyperbolic tangent function. The specific calculation formula is as follows: in, , This represents the weights and biases that need to be trained when generating positive candidate vectors; Output based on the current time of the forget gate The cell state at the previous moment The product of the input gate and the output of the input gate. and the candidate vector at the current time The product of these two factors yields the memory cell at this moment, which is the core of the LSTM structure responsible for storing and transmitting information: The last gate structure is the output gate, which first uses a sigmoid transform to determine what information to output. , and then combine By transforming the cell state, a definite output is obtained. The function expression for the output gate is: in, , This represents the weights and biases that need to be trained in the forward output gate; In summary, the first The hidden layer output at each time step can be obtained from the input at the current time step. Output of the hidden state at the previous time step and the cell state output at the previous moment Calculated together, using Summary of the first The internal computation process of the peephole LSTM at each moment: Therefore, the reverse layer is... Output at time The calculation formula is as follows: in and Weights between the input layer and the reverse layer Bias; There is no connection between the forward and backward layers, and their information transmission directions are different. The calculated states from both layers are passed to the output layer to obtain the final output at the current time. ; in These are the weights between the feedforward layer and the output. These are the weights between the back layer and the output. For bias; The final prediction value of the classification model is the first Output at time The error between this error and the true value is calculated using the proposed loss function to minimize the objective function. Training determines the model parameters; in, It is the focus factor, used to adjust the weight of easy and difficult samples. The larger the value, the more attention is paid to the difficult samples. This is a balancing parameter used to balance the weights of positive and negative samples. It is set as the ratio between the number of samples in the minority class and the number of samples in the multi-class class. For the sample The final predicted value after classification model, Then it is a sample The true label; use This represents the parameters that need to be trained in the model, including , , , , , , , , , , , , , , , , , , , , , , , , ; An adaptive moment estimation algorithm is used as a parameter optimization algorithm to update the parameters in the classification model. S5. Based on time series weights, different feature importance thresholds and time importance thresholds are set to select important time features, and an AFS decision tree is constructed to form dynamic semantic rules for multivariate time series data. Set rule threshold and ,in For feature importance threshold, The importance threshold for each moment; the concept node at the most recent moment. By the Input data and past data for each feature group at each time step The concept of time is obtained through the combined effect of nodes; for the most recent time, the first... Concept Nodes If the feature weights satisfy the following conditions, then it indicates that the features at the most recent time are... Right now Selected: in Indicates the first The first concept node The important weights of each feature Indicates the composition of the first The number of features of each concept node; If a historical concept node satisfies the following conditions, then it indicates that the concept node... In the The value at time is Selected: in, Indicates the generation of samples Concept nodes hour, Concept node before time interval The importance weight of the concept node at the current moment, Indicates the duration of the characteristic effect; For the selected historical concept nodes The current moment feature value and historical node that meet the threshold can still be selected according to the above formula. This process is iterated until no new historical concept node is selected. Therefore, the features of the conceptual nodes that influence the most recent moment are: One, construct rule features ,feature Each factor contains a feature value of time-series information, that is... The meaning is a certain feature value at a certain moment, therefore, according to the rule features Available from dataset A new feature set is obtained from this to extract rules. and sample set ; in , , indicating the first One sample, The regular features representing this sample. This represents the first sample. Each rule feature The class label representing this sample. Represents the dataset The number of samples in; Before constructing the decision tree, continuous attributes are fuzzified, that is, conceptualized, in order to obtain semantic information; for rule features... The simple concepts that take "small," "medium," and "large" as characteristics are represented as follows: The set of simple concepts after fuzzing all continuous features. It can be represented as: feature , , belongs to the The expression for the membership degree of a simple concept is as follows: : in Indicates sample Features Values, Representation of features The maximum value obtained among all samples, Representation of features The minimum value obtained among all samples, Representation of features The mean is obtained from all samples. Representation of features The largest difference between it and its average level is used to measure the dispersion or deviation of the data. Representation of features The smallest difference between the data and its average level is used to measure the concentration or stability of the data. By logically connecting multiple single attributes, more complex attributes can be formed, and the AFS logical operation rules can be used to generate a set of complex concepts. Its definition is as follows: in, A non-empty set of indices for a simple concept; express The extraction, Indicates all Conjunction; A consistent membership function can be obtained based on simple concepts and the AFS structure; assuming... It is a complex concept, and its membership function Represented as: in, For a simple concept The weighting function can be determined based on the data distribution. It is the number of times the sample was observed; The core of constructing an AFS decision tree lies in consistently selecting the node with the highest information gain for splitting, thereby maximizing the growth rate of information in complex concepts. , Split the current node Find the most suitable combination of simple concepts to ensure that the nodes after splitting are Compared with the node before splitting The information gain difference is the largest, that is: in, Indicates the current node Information gain For the current node The amount of information, for Split the current node The weighted information content is then used to obtain a simple concept for segmenting the current node. ,in and The calculation method is as follows: in, , For the first N In layer nodes membership degree , Indicates the first The class outputs a set of samples. Indicates when fuzzy concepts Split the current node At that time, it belonged to the first Consistent membership degree of class samples; From this, we can obtain the following: Maximizing the fuzzy concept, from Delete ,Right now: If the consistency membership function of a fuzzy concept satisfies , If the set threshold is not met, then the AFS logical operation is performed. "Construct new tree nodes" If the information gain of the current node is negative, that is... Or in fuzzy attributes Split If the obtained subset is an empty set, Stop expanding child nodes to obtain the AFS decision tree; at this point, the fuzzy attribute from the root node to the leaf node is a semantic rule. ; Table 1 Confusion Matrix The main purpose of the designed model is to perform classification tasks and dynamic rule extraction on multivariate time-series data. Therefore, to evaluate the performance of the binary classification algorithm, TP represents the number of samples where both the predicted and true values ​​are positive, TN represents the number of samples where both the predicted and true values ​​are negative, FN represents the number of samples where the predicted value is negative and the true value is positive, and FP represents the number of samples where the predicted value is positive and the true value is negative. Table 1 shows the confusion matrix for different attributes of the predicted and true values. In the binary classification problem, only the above four cases will occur, so the following indicators are defined based on these four results. (1) Accuracy: The percentage of correct predictions; (2) Precision: The proportion of samples that are actually positive and are also predicted to be positive out of all samples that are predicted to be positive; (3) Recall: The proportion of samples that are actually positive and are also predicted to be positive out of all samples that are actually positive; (4) F1-score: the harmonic mean of precision and recall, used to comprehensively evaluate the performance of the model; (5) ROC curve and AUROC The ROC curve is plotted with the false positive rate on the horizontal axis and the true positive rate on the vertical axis. The ROC curve is a visualization of the model's performance at different classification thresholds. AUROC is the area under the ROC curve, used to quantify the model's classification performance. The larger the AUROC value, the better the model's performance. (6) PRC curve and AUPRC The PRC curve is plotted with recall on the horizontal axis and precision on the vertical axis. The closer the PRC curve is to the upper right corner, the better the model performance. AUPRC is the area under the PRC curve, used to comprehensively evaluate the average precision of the model at different recall rates. The higher the AUPRC value, the better the model performance.

2. An apparatus for a multivariate time-series data analysis and modeling method, used to implement a fuzzy cognitive graph neural network modeling method for multivariate time-series data classification as described in claim 1, characterized in that, include: The acquisition unit is used to acquire data and process the data to form a multi-dimensional time-series data sample; The first construction unit is used for multivariate time series data sample sets. It uses dynamic time warping to perform similarity analysis on time series features, and divides the features according to the analysis results to form multiple feature groups. The second construction unit is used to generate concept nodes based on feature grouping and to define the duration of feature effects. Concept nodes are updated through a temporal fuzzy cognitive graph; The third building unit is used to construct a bidirectional long short-term memory network with peephole connection. The updated concept nodes are used as the input of the model, and the weights are trained using the loss function to form a learning model with temporal weights of concept nodes. The rule extraction unit, based on temporal weights, sets different feature importance thresholds and time importance thresholds to select important time features, constructs an AFS decision tree, and thus forms dynamic semantic rules for multivariate time-series data.

Images