Information grouping function construction method and device, electronic equipment and storage medium

Abstract

The invention provides an information grouping function construction method and device, electronic equipment and a storage medium, and the information grouping function construction method comprises the steps: obtaining to-be-queried information and a target knowledge graph; wherein the target knowledge graph is constructed based on emergencies, operation indexes and metadata in rail transit; constructing an incremental dynamic graph based on the to-be-queried information and the target knowledge graph; inputting the incremental dynamic graph and the to-be-queried information into a link prediction reasoning model to obtain a link reasoning result; wherein the link prediction reasoning model is fused with a graph attention mechanism; and generating an information grouping instruction based on a link reasoning result. According to the method, the information grouping instruction is generated through the target knowledge graph and the link prediction reasoning model, so that rapid pulling of the information is realized based on the information grouping instruction, and meanwhile, the link prediction reasoning model performs processing on the basis of the incremental dynamic graph, so that the calculation amount is reduced to a certain extent.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method, apparatus, electronic device, and storage medium for constructing information grouping functions. Background Technology

[0002] Currently, the future development direction of the network dispatching and command center is to achieve an intelligent dispatching mode that realizes dynamic, collaborative, and adaptive management of network dispatching and command through artificial intelligence technology and intelligent equipment. Among these developments, the construction of a dynamic dispatching mode based on the current dispatching scenario is a crucial direction. To achieve dynamic dispatching, it is necessary to retrieve the required information in real time from the Network Operations Control Center (NOCC), a function known as information clustering. However, under the current unified dispatching and command center control model for rail transit, the information clustering of the NOCC system generally relies on manual templates. These templates are set for specific scenarios such as emergencies, and cannot dynamically determine the information to be extracted based on improvements in operational management levels or changes in operational quantitative indicators.

[0003] For a long time, although the network dispatch and command center has built a data sharing platform, the efficiency of information extraction is still not high, the mining of existing dispatch experience and operational data is not deep, and the value of data has not been fully released. As the speed of experience accumulation and updating is getting faster and faster, the demand for information clustering is showing characteristics such as diversification, intelligence, and high quality and efficiency, but the lack of data integration and intelligent means and the low efficiency of information clustering. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, and storage medium for constructing information grouping functions, in order to solve the problem of low efficiency in information grouping in the prior art.

[0005] This invention provides a method for constructing an information grouping function, comprising: Obtain the information to be queried and the target knowledge graph; wherein, the target knowledge graph is constructed based on emergencies, operational indicators and metadata in rail transit; An incremental dynamic graph is constructed based on the information to be queried and the target knowledge graph; The incremental dynamic graph and the information to be queried are input into the link prediction inference model to obtain the link inference result; wherein, the link prediction inference model incorporates a graph attention mechanism; Information grouping instructions are generated based on the link inference results.

[0006] According to a method for constructing an information grouping function provided by the present invention, before obtaining the information to be queried and the target knowledge graph, the method further includes: Construct a first knowledge graph for emergencies in rail transit and the arguments of each emergency; wherein the arguments include operational indicators; Construct a second knowledge graph for the operational metrics and their metadata; The first knowledge graph and the second knowledge graph are aligned to obtain the target knowledge graph.

[0007] According to a method for constructing an information clustering function provided by the present invention, the link prediction inference model includes an encoder and a decoder; the step of inputting the incremental dynamic graph and the information to be queried into the link prediction inference model to obtain the link inference result includes: Based on the encoder, the incremental dynamic graph is forward-propagated to obtain encoded data; wherein, the encoder incorporates a graph attention mechanism. Based on the decoder and the encoded data, link prediction inference is performed to obtain the link inference result; wherein, the link inference result includes multiple candidate predicted links.

[0008] According to a method for constructing an information grouping function provided by the present invention, the training process of the link prediction inference model includes: Obtain a training dataset; wherein the training dataset includes multiple training samples based on training burst events, training operation metrics, and training metadata; Based on the target knowledge graph and the training samples, the initial link prediction inference model is iteratively trained to determine the model parameters of the initial link prediction inference model, and the link prediction inference model is obtained based on the model parameters; wherein, the model parameters include independent weight matrix and diagonal matrix parameters for the relationship between the sudden event, the operational indicators and the metadata, and graph attention weight parameters for the importance of the operational indicators.

[0009] According to a method for constructing an information grouping function provided by the present invention, the step of iteratively training an initial link prediction inference model based on the target knowledge graph and the training samples includes: Based on the target knowledge graph, the training samples, and the training loss, the initial link prediction inference model is iteratively trained; wherein, the training loss includes the main loss and the regularization loss, and the regularization loss is the relative entropy loss calculated when there is an argument relationship between the sudden event and the operational indicator.

[0010] According to a method for constructing an information grouping function provided by the present invention, the step of constructing an incremental dynamic graph based on the information to be queried and the target knowledge graph includes: Using the query event in the query information as the root node, the first-level neighbor node and second-level neighbor node of the query event are determined in the target knowledge graph to obtain the incremental dynamic graph.

[0011] According to a method for constructing an information clustering function provided by the present invention, the step of generating information clustering instructions based on the link inference result includes: Obtain a predefined instruction template, and generate information grouping instructions based on the instruction template and the link inference result; The information grouping instructions are embedded into the unified dispatch and command system of the network to obtain the instruction mapping scheme corresponding to dispatch and command and emergency response.

[0012] The present invention also provides an apparatus for constructing an information grouping function, comprising: The first acquisition module is configured to acquire the information to be queried and the target knowledge graph; wherein, the target knowledge graph is constructed based on emergencies, operational indicators and metadata in rail transit; The first construction module is configured to construct an incremental dynamic graph based on the information to be queried and the target knowledge graph; The input module is configured to input the incremental dynamic graph and the information to be queried into the link prediction inference model to obtain the link inference result; wherein, the link prediction inference model incorporates a graph attention mechanism; The generation module is configured to generate information grouping instructions based on the link inference results.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a method for constructing the information grouping function as described above.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for constructing information grouping functions as described above.

[0015] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements a method for constructing information grouping functions as described above.

[0016] This invention provides a method, apparatus, electronic device, and storage medium for constructing an information grouping function. It pre-constructs a target knowledge graph based on emergencies, operational indicators, and metadata in rail transit, and then constructs an incremental dynamic graph for the information to be queried. The incremental dynamic graph and the information to be queried are input into a link prediction and reasoning model to obtain link reasoning results. This link prediction and reasoning model incorporates a graph attention mechanism, enabling easy and accurate link prediction and reasoning within the incremental dynamic graph. Information grouping instructions are generated based on the link reasoning results. This invention generates information grouping instructions using a target knowledge graph and a link prediction and reasoning model, facilitating rapid information retrieval based on these instructions. Furthermore, the link prediction and reasoning model in this application processes the incremental dynamic graph, reducing computational load to some extent. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the method for constructing the information grouping function provided by the present invention.

[0019] Figure 2 This is a schematic diagram of the target knowledge graph provided by the present invention.

[0020] Figure 3 This is a schematic diagram of the structure of the information grouping function construction device provided by the present invention.

[0021] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0022] 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 with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] Figure 1 This is a flowchart illustrating a method for constructing an information grouping function according to an exemplary embodiment. For example... Figure 1As shown in an exemplary embodiment, the method for constructing an information grouping function includes steps 110 to 140, which are described in detail below.

[0024] Step 110: Obtain the information to be queried and the target knowledge graph; wherein, the target knowledge graph is constructed based on emergencies, operational indicators and metadata in rail transit.

[0025] In this embodiment of the invention, a target knowledge graph is pre-constructed. The target knowledge graph is obtained based on a first knowledge graph and a second knowledge graph. The first knowledge graph represents the relationship between emergencies in rail transit and between emergencies and arguments. Event entities in the knowledge graph have two types of argument roles: event location and operation indicators. The relationship between an event entity and the operation indicator argument is "possessing argument", and the relationship between an event entity and the event location argument is "event location".

[0026] The second knowledge graph represents the relationship between operational indicators and metadata, namely, operational indicator entities, metadata entities, and the necessary relationship between the two. The operational indicator entity attributes include the management scope, and the management scope attributes include one of four types of attributes: station, section, line, and network.

[0027] Using the node ID of the operational metric as the alignment entity, align the first knowledge graph and the second knowledge graph to form the target knowledge graph, and store it in the graph database.

[0028] The information to be queried is the information entered by the user, such as what metadata I need to monitor under the current incident.

[0029] Step 120: Construct an incremental dynamic graph based on the information to be queried and the target knowledge graph.

[0030] In this embodiment of the invention, for the sudden event to be queried presented in the information to be queried, the corresponding event entity in the target knowledge graph is used as the head node, and the target knowledge graph is expanded to generate an incremental dynamic graph. The incremental dynamic graph can be continuously updated and expanded as the situation changes.

[0031] Step 130: Input the incremental dynamic graph and the information to be queried into the link prediction inference model to obtain the link inference result; wherein, the link prediction inference model integrates a graph attention mechanism.

[0032] In this embodiment of the invention, a link prediction and reasoning model is pre-trained. The link prediction and reasoning model integrates a graph attention mechanism, which enables link prediction and reasoning to be performed on the query information based on an incremental dynamic graph, and obtains the link reasoning result for the query information.

[0033] Step 140: Generate information grouping instructions based on the link inference results.

[0034] In this embodiment of the invention, information clustering instructions are generated based on the link inference results, so as to retrieve the information corresponding to the information to be queried from the online unified scheduling and command system based on the information clustering.

[0035] In this embodiment of the invention, a target knowledge graph is pre-constructed based on emergencies, operational indicators, and metadata in rail transit. Then, an incremental dynamic graph is constructed for the information to be queried. The incremental dynamic graph and the information to be queried are input into a link prediction and reasoning model to obtain link reasoning results. The link prediction and reasoning model integrates a graph attention mechanism, enabling easy and accurate link prediction and reasoning within the incremental dynamic graph. Information grouping instructions are generated based on the link reasoning results. This invention generates information grouping instructions using the target knowledge graph and the link prediction and reasoning model, facilitating rapid information retrieval. Furthermore, the link prediction and reasoning model in this application processes the incremental dynamic graph, reducing computational load to some extent. This invention generates information grouping instructions using the target knowledge graph and the link prediction and reasoning model, and then, based on these instructions, enables adaptive intelligent-driven real-time decision-making for network dispatching and command operations, rapid iteration of business knowledge, and rapid response to information retrieval. This enhances the capabilities of network dispatching and command automation and informatization integration, multi-source information integration, and intelligent collaborative services, adapting to the dynamic adaptive needs of intelligent dispatching and comprehensively improving the integrated business data fusion capabilities of network dispatching and command.

[0036] In an exemplary embodiment of the present invention, before obtaining the information to be queried and the target knowledge graph, the method further includes: Construct a first knowledge graph for emergencies in rail transit and the arguments of each emergency; wherein the arguments include operational indicators; Construct a second knowledge graph for the operational metrics and their metadata; The first knowledge graph and the second knowledge graph are aligned to obtain the target knowledge graph.

[0037] In this embodiment of the invention, various types of operational record text source data are acquired, such as news reports, emergency response reports, historical emergency handling records, and emergency handling logs. These operational record text source data can reflect the entire process of emergency response, thereby forming an emergency corpus.

[0038] Specifically, several emergencies that can be included in the same emergency response report will form an event chain. (Event Chain) consists of several events The total number of events is ,but The events naturally possess a certain logical causal relationship.

[0039] Every emergency Each argument has several arguments, which represent the characteristics that should be noted during emergency response to sudden events. (Feature Chain) can be configured as needed to include the time and location of the event, the number of casualties, the duration of operational disruption, the resulting economic losses, the emergency measures taken, and operational metrics. Rounds can be configured as needed, such as using the event location and operational metrics as the arguments. Indicates an emergency Characteristics involved in emergency response.

[0040] There can be multiple operational metrics, such as handling sudden surges in passenger flow at stations. At the same time, it is necessary to monitor real-time passenger flow, passenger flow growth rate, passenger flow warning level, number of people gathered on the platform, current train occupancy rate, train departure interval, availability of backup capacity, average passenger waiting time, availability of flow control measures, passenger flow control level, average evacuation time at the current station, current station information coverage, personnel density in key areas, and equipment operating status. In this embodiment, traction failure events have indicators such as traction equipment status and overhead contact line status; train delay events have indicators such as train delay time; station congestion events have indicators such as station cross-sectional passenger flow, section cross-sectional passenger flow, and large passenger flow level; and passenger complaint events have indicators such as complaint intensity.

[0041] Operational metrics need to be uniformly assigned values ​​to achieve consistent metadata status. Numerical operational metrics are represented by floating-point numbers to indicate changes in status and are assigned values ​​to metadata, such as the current real-time passenger flow within the station being 4395 people. Data representing equipment status needs to undergo enumeration and numerical transformation before being assigned values ​​to metadata, such as converting equipment status into 0 / 1 / 2 status indicators to represent equipment failure / offline / online statuses, respectively.

[0042] When extracting emergencies and corresponding arguments from operational record text source data, a trained first entity extraction model can be used for extraction. The first entity extraction model includes a first extraction sub-model for extracting event entities and a second extraction sub-model for extracting argument entities.

[0043] The first extraction sub-model can be trained using large-scale Chinese-based natural language processing models such as BERT and RoBERTa. During training, it can be pre-trained using domain dictionaries and domain corpora to improve the model's recognition capabilities in the field of rail transit scheduling. The first extraction sub-model uses event extraction and relation extraction techniques to extract sudden events, arguments, and relations from operational record text source data. The first extraction sub-model extracts sudden event entities based on fixed trigger words, such as traction failures, train delays, line congestion, and passenger complaints. When setting trigger words, it is necessary to fully cover all possible trigger words. Essentially, the extraction of event entities in the first extraction sub-model is a pattern extraction or dictionary matching method, based on deterministic rules, requiring no parameters and no loss function.

[0044] The first extraction sub-model extracts all argument entities, such as the location of the event and operational indicators. This first extraction sub-model can also be built on a pre-trained language model + linear classification layer architecture, and fine-tuned using BIO-annotated corpora of operational indicators in the rail transit field. For example, using "B" as the entity start word, "I" as the entity middle or end word, and "O" as the annotation word not belonging to any entity, the statement "Currently, the next train entering the station has a full load rate of 80%" can be adjusted to "Currently, the next train entering the station (O) (B) (I) has a full load rate (I) of (O)80 (O)% (O)".

[0045] The training dataset should contain more than 10,000 high-quality labeled data entries. The pre-trained language model should be tuned using the cross-entropy loss function shown below: ; in, This represents the true label of the sample input to the pre-trained language model. This represents the output of the first extraction sub-model.

[0046] After argument entity extraction, a pre-trained language model combined with a linear classification layer architecture is used to construct an argument role classification model to classify arguments by role and clarify which arguments belong to which event. The structure of emergency event corpora generally first describes the situation of the emergency (what happened), and then describes which operational indicators were involved in the handling process. Therefore, following this natural language assumption, in the output text labeled with the location of the event and operational indicators, an additional layer of manual event type labeling is added for each operational indicator, assigning it a positive sample value of 1 (i.e., this argument belongs to the labeled emergency event). Then, a sufficient number of negative samples are constructed and assigned a value of 0. Tags are inserted before and after the arguments to be judged. , as a special indicator of argument boundaries. Event sentences are packaged according to natural language assumptions, beginning with [CLS] and sequentially listing events. All corpora are segmented using [SEP] to encompass all arguments. The data is then input into the model for classification training, and the classification results are still optimized using the cross-entropy loss function.

[0047] By calculating event similarity, the extracted specific emergencies are clustered and generalized to form the first knowledge graph of upper-level emergencies. Two vectors in dimensional space , Representing events respectively With the event The similarity is calculated using the following formula: .

[0048] The natural language patterns in the corpus of sudden events define all relationships between sudden events as causal relationships. This is to summarize the development patterns of sudden events from massive amounts of data and to help predict possible next events when a sudden event occurs.

[0049] An event may be related to multiple events, becoming either a cause or an effect of those events. For example, a traction failure may cause a train delay, a train breakdown in a section, or a rear-end collision. Therefore, it is necessary to determine the causal relationships between sudden events. This determination can be done manually or using deep learning models. The deep learning model approach involves first constructing a pre-trained language model based on a corpus of event-related relational annotations. Then, on the input text sequence, the causal relationships between the sudden events are analyzed. Event markers are inserted before and after the trigger word as special indicators of the event trigger word boundary. Event sentences are packaged according to natural language assumptions, starting with [CLS], and two event corpora are sequentially split using [SEP], encompassing all arguments to achieve complete extraction of the event trigger word context and arguments. Following the principle of sequential description of the scheduling process text, event pairs are enumerated (using various algorithms such as the sliding window locality hypothesis algorithm). Only sudden events appearing within the window selection are paired to form event pairs. A pre-trained language model + linear classification layer model is then used to determine the relationship between the event pairs and whether they belong to a causal relationship. The model is optimized using the cross-entropy loss function.

[0050] For each event pair, generate an inverse cause-effect relationship, such as Figure 2As shown. The constructed causal relationship is naturally oriented, and reverse causal relationships can be added using Python code such as networkx.DiGraph.reverse(), and the two graphs can be merged.

[0051] An expert scoring mechanism is used to assign importance scores to each relationship edge of <emergency event, argument, operational metric>, with scores assigned to each edge. For each event, the sum of the importance of all edges with argument relations is 1. Importance of all p arguments At the operational metrics level, during emergency response, it is usually necessary to distinguish the most important and fundamental one or several operational metrics. Therefore, an expert scoring mechanism is adopted to score the importance of each operational metric in each emergency. This allows the model to introduce a two-stage graph attention mechanism based on expert scoring and self-learning. Expert scoring helps the encoder of the link prediction inference model determine which operational metric is more noteworthy, while the encoder also learns the relationship between operational metrics and emergencies through vector relationships, thus maintaining a balance.

[0052] We obtain operational metrics from knowledge sources such as scheduling procedures, operation logs, and operational metric statistical rules. By analyzing the calculation logic of each operational metric (including calculation formulas and aggregation rules), we obtain the corresponding underlying metadata requirements, and then extract the metadata name field to form a structured set of logical layer business rules.

[0053] Define the schema layer of the second knowledge graph, constructing two types of entities: operational metrics and metadata, as well as the required relationships, i.e., the calculation of operational metrics requires metadata. First, extract all operational metrics and metadata entities using the second extraction sub-model, then determine the relationships between entities to clarify which metadata belongs to which operational metric. A rule-based extraction method can be used to construct a vocabulary with extraction fields, or a pre-trained language model based on supervised learning can be input. The sequence labeling principle is the same as described in the first knowledge graph construction process above, or a large language model can be used directly for intelligent extraction. Construct a set of <operational metrics, required, metadata> triples to form a structured logical association knowledge sub-graph describing the business logic: "operational metrics → (relationship: required) → metadata," thus forming the second knowledge graph supported by the underlying data. Add a management scope attribute to the operational indicator entity to clarify whether the operational indicator belongs to a station, section, line or network, and distinguish indicators with the same name but different functions.

[0054] Based on operational indicator entities, two knowledge graphs are merged through entity alignment technology to construct a two-layer knowledge graph that integrates reasoning and knowledge, namely the target knowledge graph.

[0055] Specifically, two knowledge graphs are loaded. The first knowledge graph is used to extract the location of all events, the argument names of operational metrics, and their respective IDs. The second knowledge graph is used to extract the entity names of all operational metrics and their respective IDs.

[0056] For the event locations extracted from the first knowledge graph, methods such as rules and machine learning classifiers are used to summarize and map them into four categories of enumerations: station, section, line, and network. Candidate pairs are constructed: <Operational indicators extracted from the first knowledge graph, event location mapping> and <Operational indicators extracted from the second knowledge graph, management scope>.

[0057] Candidate pairs with identical names are identified through string matching. If two operational metrics in a candidate pair are the same, it is marked as aligned with a confidence level of 1, and the node ID and mapping relationship are recorded. For the remaining entities with different names but identical semantics, a confidence threshold of 0.9 can be set. For the remaining entities with different names but identical semantics, methods such as Jaccard similarity calculation or edit distance similarity can be used for identification. The Jaccard similarity calculation formula is as follows: ; Here, C and D represent the sets of the two "operational indicators" entities in the candidate pair after word segmentation. Based on the calculated similarity, candidate pairs with similarity higher than the similarity threshold are marked as aligned, and the node ID and mapping relationship are recorded.

[0058] When entities are aligned, the target knowledge graph forms a vertical association of “event → (relationship: owning argument) → operational metric → (relationship: need) → metadata”.

[0059] Using graph databases such as Neo4j, the target knowledge graph is saved according to a pre-set saving frequency, and the constructed target graph structure is as follows: Figure 2 As shown. Simultaneously, a node ID mapping table for graph fusion is constructed.

[0060] In an exemplary embodiment of the present invention, the training process of the link prediction inference model includes: Obtain a training dataset; wherein the training dataset includes multiple training samples based on training burst events, training operation metrics, and training metadata; Based on the target knowledge graph and the training samples, the initial link prediction inference model is iteratively trained to determine the model parameters of the initial link prediction inference model, and the link prediction inference model is obtained based on the model parameters; wherein, the model parameters include independent weight matrix and diagonal matrix parameters for the relationship between the sudden event, the operational indicators and the metadata, and graph attention weight parameters for the importance of the operational indicators.

[0061] In this embodiment of the invention, a link prediction and inference model is pre-trained, and the initial link prediction and inference model is trained offline using a constructed target knowledge graph. The link prediction and inference model includes an encoder and a decoder, and the encoder adopts a relational graph convolutional network model. When training the link prediction and inference model, following the update frequency of the target knowledge graph, the decoder is used to train the model at a low frequency until convergence according to a pre-set training frequency, and outputs a model with fixed model parameters, thus obtaining the link prediction and inference model.

[0062] Relational graph convolutional network models can distinguish different types of relationships in a target knowledge graph, thereby assigning higher scores to link prediction probabilities that better match expectations during queries. By simultaneously encoding nodes and different types of relationships to achieve link prediction, this model is well-suited for graph reasoning scenarios involving information clusters in rail transit.

[0063] In this embodiment of the invention, a relational graph convolutional network model with a fused graph attention mechanism is constructed by assigning expert scores to each <emergency event, argument, operational metric> relational edge in the first knowledge graph, generating edge-level weight coefficients through the graph attention mechanism, and then weighting and summing them during the aggregation stage of the relational graph convolutional network model to achieve joint encoding of edge weights and relational matrix. When the decoder queries, it can infer the most important operational metrics based on different weight scores, and prune invalid information during the information extraction stage.

[0064] In this embodiment of the invention, a two-layer stacked relational graph convolutional network model is defined for urban rail transit information clusters to adapt to the two-hop reachability pattern of events, operational indicators, and metadata of the target knowledge graph, while reducing computational load and improving response speed.

[0065] The parameters to be trained in the link prediction inference model include an independent weight matrix representing the relationships between sudden events, operational metrics, and metadata, i.e., causal relationships. Cause and effect relationship Having argument relations Needs and Relationships and self-circulation relationship The independent weight matrix, in addition to the graph attention weight parameters based on the graph attention mechanism and expert prior knowledge jointly learned, and the diagonal matrix parameters of the aforementioned five relationships given by the decoder. , , , , .

[0066] During the training phase, a training dataset is first constructed, comprising a positive sample set and a negative sample set. The positive samples in the positive sample set and the negative samples in the negative sample set constitute the training samples. The positive sample set is obtained by layer-wise sampling of positive sample triples from the target knowledge graph according to five existing relations, ensuring a balanced number of samples for each relation. The samples are then concatenated and deduplicated. The triples to be sampled include the following types: <Event entity, causal relationship, event entity>; <Event entity, causal relationship, event entity>; <Event entity, possessing arguments, operational metric entity>; <Operational metrics entity, requires metadata entity>; <Entity, self-looping, entity>.

[0067] The negative sample set is generated using a negative sampling method that randomly replaces the first and last entities, resulting in triplet negative samples that do not exist in the target knowledge graph, thus ensuring sample balance. Positive samples are labeled with a 1, and negative samples are labeled with a 0. The encoder is trained through forward propagation with message passing, linear transformation, computational graph attention, normalization, weighted aggregation, and updates to embed nodes. During weighted aggregation, only nodes with argument relations and their corresponding weighted average aggregation methods using fused graph attention parameters are employed. For other nodes and relations that do not require an attention mechanism, a normalized aggregation method based on normalization constants is used. The update method for nodes with argument relations can be represented as follows: ; in, Vector representation of a node; It represents all specific relation types (including the five relations that appear in the target knowledge graph). In relation type At that time, node The set of all neighbors, where, for each type of relationship They all use an independent weight matrix Perform the transformation; In this embodiment of the invention, the normalization constant is... (i.e., relation type is) At that time, the number of all neighbor nodes (balancing the differences in the number of neighbors with different relationships, and treating the contribution of all neighbors to the central node as equal) is changed to a graph attention score, which is an adaptive and learnable weight. (i.e., relation type is) At that time, node neighboring nodes The graph attention weight score) makes the aggregation process incorporating the attention mechanism change from an aggregation of all edges with equal weights to a weighted sum, where the weights are determined by the features of the node itself and its neighbors. By different types of relationships Neighbor information obtained from down-aggregation and self-join transformation The summation result is passed through a non-linear activation function (ReLU). , obtain node In the A new representation of layers.

[0068] Among them, by the central node and neighboring nodes In relationship The graph attention weight score is obtained by performing calculations on the following representations. : ; in, Representing relations The attention vector is a learnable parameter used to calculate the attention weights; the graph attention weights are calculated using the LeakyReLU nonlinear function. This indicates concatenating two node vectors that have undergone linear transformation. Indicates the central node The representation vector After the relationship-specific weight matrix The result after linear transformation Indicates the central node neighboring nodes Through the same relation-specific weight matrix The result of the linear transformation is to enable the subsequent dot product calculation to obtain a joint representation of the two types of nodes simultaneously; Then, the Softmax function is used to normalize the scores of all neighbors to obtain the final graph attention weight parameters, the sum of which is 1. .

[0069] In this embodiment of the invention, the link prediction task is a binary classification task that determines whether a certain type of relationship exists between two nodes. The entire link prediction task is completed through a set of multiple independent binary classification tasks. The decoder in this embodiment uses the DistMult decoder, which trains an independent, learnable diagonal matrix for each type of relationship. , , , , Each diagonal matrix is ​​only sensitive to its corresponding relation, in order to distinguish what data needs to be monitored (involving having arguments, needing relations), what event will happen next (involving causal relations), or to find the root cause of an event (involving effect-cause relations), etc., which involve different relations.

[0070] In an exemplary embodiment of the present invention, the iterative training of the initial link prediction inference model based on the target knowledge graph and the training samples includes: Based on the target knowledge graph, the training samples, and the training loss, the initial link prediction inference model is iteratively trained; wherein, the training loss includes the main loss and the regularization loss, and the regularization loss is the relative entropy loss calculated when there is an argument relationship between the sudden event and the operational indicator.

[0071] In this embodiment of the invention, during the training phase, each relation edge in the predicted link needs to undergo independent binary classification of the five relations mentioned above, and the result with the smallest loss function is taken as the prediction result. However, during the training process, only a single positive / negative sample triple < entity 1, entity 2, relation is classified. Is it a specific relationship? This task involves aggregating the loss functions for all positive / negative sample triples for model parameter tuning. The loss function is only tuned during the training phase and becomes ineffective during the prediction phase, where the decoder scoring function is used.

[0072] In this embodiment of the invention, among the aforementioned five types of relations, only relations with arguments are considered. The relative entropy (Kullback-Leibler Divergence, KL) loss between graph attention weights and expert scores is calculated, and this loss is added as a regularization term to the loss function of the training loss of the link prediction inference model.

[0073] The training loss function consists of two parts: the main loss function of the model and the regularization loss function. The regularization loss is calculated only for nodes with argument relations and their nodes, while the main loss is calculated using the cross-entropy method for other nodes and relations.

[0074] In this embodiment of the invention, a binary classification cross-entropy loss function is used to calculate the main loss for link prediction. The loss function for each relation edge is as follows: ; in, This indicates that the relationship type in the link prediction results is... The probability is denoted as , where 0 indicates the relationship does not exist and 1 indicates its existence. The average of all relation edges in the entire link prediction task is calculated: ; Based on the residence loss, this embodiment of the invention calculates the graph attention weights learned by the model for specific relation edges that have arguments. Scoring by experts The KL loss between them. The graph attention weights are transformed into a probability distribution as follows: ; Original expert scoring vector After being calculated using the Softmax function, it is converted into a probability distribution: .

[0075] To enable the model to learn the graph attention probability distribution As close as possible to the probability distribution of expert scores KL loss is used to measure the central node In relation to All neighboring nodes The sum of the differences between the two probability distributions, for example, the central node (event A) has several operational indicator arguments at the same time, calculate the total difference between the weights learned by the graph attention distribution of all operational indicator arguments and the probability distribution of expert scoring.

[0076] .

[0077] To reduce the impact of simple summation on the main loss function, after calculating the KL loss for all relationships between all nodes and their neighbors, the calculated KL losses are summed and then averaged, as expressed by the following formula: ; in, This represents the product of the number of nodes and the number of relations; its value equals the total KL loss to be calculated. It represents the set of all nodes in the graph.

[0078] Then, this loss is added as a regularization term to the model's total loss function, using hyperparameters. To control the strength of regularization, i.e., the degree of influence of KL loss on the main loss function, soft guided graph attention weights. Approximately expert scores The expression for the training loss is shown below.

[0079] ; hyperparameters The value needs to be adjusted based on the model's link prediction performance on the validation set, taking the value that yields the best performance. value.

[0080] During backpropagation, all parameters of all models are automatically updated through gradient descent, following the two gradient descent paths of main loss and regularization loss. This includes gradually updating decoder parameters, graph attention weights, node representations, and independent weight matrices for all different relation categories.

[0081] The initial link prediction inference model is iteratively trained using training samples until the model converges, thus obtaining the link prediction inference model.

[0082] In an exemplary embodiment of the present invention, the link prediction inference model includes an encoder and a decoder; the step of inputting the incremental dynamic graph and the query information into the link prediction inference model to obtain the link inference result includes: Based on the encoder, the incremental dynamic graph is forward-propagated to obtain encoded data; wherein, the encoder incorporates a graph attention mechanism. Based on the decoder and the encoded data, link prediction inference is performed to obtain the link inference result; wherein, the link inference result includes multiple candidate predicted links.

[0083] In this embodiment of the invention, a trained encoder is used to perform forward propagation coding of message passing, linear transformation, computational graph attention, normalization, weighted aggregation, and updates on the incremental dynamic graph.

[0084] Message passing refers to obtaining the first-level neighbor nodes, second-level neighbor nodes, and vector representations of relationships of the head node in an incremental dynamic graph, starting from the head node, using single-hop or two-hop methods, based on the input query information.

[0085] Aggregation involves normalizing and weighting the vector representations of all sudden event nodes, event arguments (operational metrics), and first- and second-order neighbor nodes of metadata on the incremental dynamic graph. For each type of causal relationship, effect-cause relationship, and need relationship, a corresponding independent weight matrix is ​​used for the neighbors. Perform linear transformations separately, then sum them according to the normalization constant. For neighboring nodes with argument relations, use their corresponding independent weight matrices. After performing a linear transformation, a weighted sum is calculated based on the graph attention weight parameters. Finally, the aggregation results under all different relationships are added together and then updated after being combined with the node's own information.

[0086] The update involves adding the current vector representation of all nodes to the aggregated neighbor information, inputting it into the ReLU activation function, and generating a new vector representation for all nodes in the first layer.

[0087] The encoder has a two-layer stacked structure. Therefore, the output representations of all nodes in the first layer are used as the inputs to the second layer, and the above message passing, aggregation, and update process is repeated. Through these two stacked layers, the final representation of each node incorporates information from its two-hop neighborhood.

[0088] In this embodiment of the invention, taking a positive edge of causal relationship as an example, the embedding is updated by aggregating the first-level neighbor nodes of all head nodes in the first layer of stacking: (1) “Event 1 → (Relationship: Cause and Effect) → Event 2”; (2) "Event 1 → (Relationship: Owning Argument) → Operational Metric 1"; In the second-level stack, aggregate all second-level neighbor nodes of the head node, and use the updated first-level embeddings to aggregate second-level neighbors in parallel, updating the embeddings as follows: (1) “Event 1 → (relationship: causality) → Event 2 → (relationship: causality) → Event 3”; (2) "Event 1 → (relationship: causality) → Event 2 → (relationship: ownership argument) → Operational indicator 2"; (3) “Event 1 → (relation: own argument) → Operational indicator 1 → (relation: need) → Metadata 1”.

[0089] Similarly, the reverse edges of the cause-effect relationship are used to complete the two-layer embedding in parallel.

[0090] In this embodiment of the invention, the encoded data output by the encoder is input into the decoder. The decoder uses the information clustering requirement as a link prediction task, scores and ranks the candidate predicted links of "to be queried sudden event → metadata". For the NOCC information clustering requirement, a triple query of <head node, relation, query requirement> is constructed, and a scoring function is used for the triple. Rating: ; in, , These represent the head node obtained from the encoder and the vector of the query request, respectively. The query request is derived from the information to be queried. It is a relationship diagonal matrix, This indicates that the score is obtained by weighting the dot product across all dimensions.

[0091] The decoder uses different diagonal matrices to calculate confidence scores for queries involving different relationships, resulting in candidate entities with high scores. A list of results is returned in descending order of confidence score, representing the probability of the triple being true, with node entities identified by their IDs. Candidate prediction links are links formed with the query event as the head node. The decoder outputs the confidence score of the triples corresponding to each candidate prediction link and sorts the candidate prediction links in descending order of confidence score. The top-K candidate prediction links are selected to obtain a link sequence list, which also records the IDs of each node.

[0092] A rule-based post-processing module is constructed to return the link sequence list to the target knowledge graph, query and find the name and real attributes of the node based on the entity ID, and output the corresponding link and node.

[0093] The link prediction inference model provided in this embodiment of the invention can realize the following three types of queries.

[0094] (1) For the query “What metadata do I need to monitor under the current incident?”, input the two triples <“the XX event that is currently occurring”, which have arguments and query requirements> and <operational indicators, requirements, query requirements>. The link prediction reasoning model performs cross-layer reasoning of information groups according to the path of “current event → (relation: arguments) → operation indicators → (relation: requirements) → metadata”. (2) For the query “Under the current accident, what metadata do I need to monitor for the next event?”, input <“the XX event that is currently happening”, causality, query requirements>, the link prediction reasoning model will perform cross-layer reasoning of information grouping according to the path of “current event → (relationship: causality) → predicted event → (relationship: possessing arguments) → operational indicators → (relationship: need) → metadata”; (3) For the query “Under the current accident, find the root cause of the accident and find the metadata that may cause the problem?”, input <query requirements, cause and effect, “the XX event that is currently occurring”>, the link prediction reasoning model will perform cross-layer reasoning of information grouping according to the path of “current event → (relationship: cause and effect) → predicted event → (relationship: possessing arguments) → operational indicators → (relationship: need) → metadata”.

[0095] In an exemplary embodiment of the present invention, the step of constructing an incremental dynamic graph based on the information to be queried and the target knowledge graph includes: Using the query event in the query information as the root node, the first-level neighbor node and second-level neighbor node of the query event are determined in the target knowledge graph to obtain the incremental dynamic graph.

[0096] In this embodiment of the invention, to reduce computational complexity, minimize interference from irrelevant information, save computational resources, and improve response speed, an incremental dynamic graph update mechanism is constructed when a sudden event occurs. From the vast target knowledge graph, all nodes and relationships related to the current queried sudden event are extracted and pieced together to form a temporary, small-scale incremental dynamic graph. The link prediction and reasoning model then performs reasoning and computation on this small-scale incremental dynamic graph.

[0097] When an event occurs, the seed node is first located. Based on the input query information, one or more uniquely corresponding sudden event nodes are found in the first knowledge graph of the target knowledge graph by matching the unique attributes of the event (such as the event name, event ID, fault range name or / device ID, etc.) as head nodes and activated. The head node set is then constructed as the smallest dynamic subgraph.

[0098] When expanding the subgraph, a multi-hop neighborhood expansion method is adopted. That is, based on the minimum dynamic subgraph, a two-layer stacked encoder is adapted to perform a breadth-first search around the activated head node. That is, single-hop and two-hop collection of all neighbor nodes (i.e., first-level neighbor nodes and second-level neighbor nodes) and the relationship edges between nodes are collected. This includes nodes of sudden events, neighbor event nodes and their directly connected operational indicators and metadata nodes, to obtain an incremental dynamic graph.

[0099] When the situation changes, such as Traction failure events occur frequently. Train delays occur frequently. "Station congestion" incidents occur frequently. "Passenger complaint" events occur constantly, and the incremental dynamic graph is updated incrementally. That is, based on the development trend of newly input emergencies to be queried, new emergencies to be queried nodes and their related operational indicators and metadata nodes are activated and added to the established incremental dynamic graph to form an incremental dynamic graph that can be used for reasoning.

[0100] In an exemplary embodiment of the present invention, generating information grouping instructions based on the link inference result includes: Obtain a predefined instruction template, and generate information grouping instructions based on the instruction template and the link inference result; The information grouping instructions are embedded into the unified dispatch and command system of the network to obtain the instruction mapping scheme corresponding to dispatch and command and emergency response.

[0101] In this embodiment of the invention, a pre-compiled instruction template is used to fill in the returned links and nodes. The instruction template is converted into information grouping instructions in REST / SQL format according to user requirements, and submitted to the unified network scheduling and command system through the API gateway. The system is then pushed to the workflow engine to output the instruction mapping scheme required for scheduling, command, and emergency response.

[0102] The following describes the apparatus for constructing the information grouping function provided by the present invention. The apparatus for constructing the information grouping function described below can be referred to in correspondence with the method for constructing the information grouping function described above. It should be noted that the apparatus provided in the following embodiments and the method provided in the above embodiments belong to the same concept, and the specific way in which each module and unit performs operations has been described in detail in the method embodiments, and will not be repeated here.

[0103] In one exemplary embodiment of the present invention, please refer to Figure 3 , Figure 3 This is an apparatus for constructing an information grouping function according to an exemplary embodiment, comprising the following modules.

[0104] The first acquisition module 310 is configured to acquire the information to be queried and the target knowledge graph; wherein, the target knowledge graph is constructed based on emergencies, operational indicators and metadata in rail transit; The first construction module 320 is configured to construct an incremental dynamic graph based on the information to be queried and the target knowledge graph. Input module 330 is configured to input the incremental dynamic graph and the information to be queried into the link prediction inference model to obtain the link inference result; wherein, the link prediction inference model incorporates a graph attention mechanism; The generation module 340 is configured to generate information grouping instructions based on the link inference results.

[0105] In an exemplary embodiment of the present invention, the information grouping function construction apparatus further includes: The second construction module is configured to construct a first knowledge graph for emergencies in rail transit and arguments for each emergency; wherein the arguments include operational indicators. The third construction module is configured to construct a second knowledge graph for the operational metrics and the metadata of the operational metrics; The entity alignment module is configured to perform entity alignment operations on the first knowledge graph and the second knowledge graph to obtain the target knowledge graph.

[0106] In an exemplary embodiment of the present invention, the link prediction inference model includes an encoder and a decoder; the input module 330 includes: The encoding submodule is configured to perform forward propagation encoding on the incremental dynamic graph based on the encoder to obtain encoded data; wherein the encoder incorporates a graph attention mechanism. The link prediction inference submodule is configured to perform link prediction inference based on the decoder and the encoded data to obtain the link inference result; wherein the link inference result includes multiple candidate predicted links.

[0107] In an exemplary embodiment of the present invention, the apparatus for constructing an information grouping function includes: The second acquisition module is configured to acquire a training dataset; wherein the training dataset includes multiple training samples based on training burst events, training operation indicators and training metadata; The training module is configured to iteratively train the initial link prediction inference model based on the target knowledge graph and the training samples, determine the model parameters of the initial link prediction inference model, and obtain the link prediction inference model based on the model parameters; wherein, the model parameters include independent weight matrix and diagonal matrix parameters for the relationship between the sudden event, the operational indicators and the metadata, and graph attention weight parameters for the importance of the operational indicators.

[0108] In one exemplary embodiment of the present invention, the training module includes: The training submodule is configured to iteratively train the initial link prediction inference model based on the target knowledge graph, the training samples, and the training loss; wherein, the training loss includes the main loss and the regularization term loss, and the regularization term loss is the relative entropy loss calculated when there is an argument relationship between the sudden event and the operational indicator.

[0109] In an exemplary embodiment of the present invention, the first construction module 320 includes: A submodule is constructed and configured to take the query event in the query information as the root node, determine the first-level neighbor node and second-level neighbor node of the query event in the target knowledge graph, and obtain the incremental dynamic graph.

[0110] In an exemplary embodiment of the present invention, the generation module 340 includes: The acquisition submodule is configured to acquire a predefined instruction template and generate information grouping instructions based on the instruction template and the link inference result. The embedded submodule is configured to embed the information grouping instructions into the unified network dispatch and command system to obtain the instruction mapping scheme corresponding to dispatch and command and emergency response.

[0111] Figure 4An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, communications interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a method for constructing an information grouping function. This method includes: acquiring the information to be queried and a target knowledge graph; wherein the target knowledge graph is constructed based on emergencies, operational indicators, and metadata in rail transit. An incremental dynamic graph is constructed based on the information to be queried and the target knowledge graph; The incremental dynamic graph and the information to be queried are input into the link prediction inference model to obtain the link inference result; wherein, the link prediction inference model incorporates a graph attention mechanism; Information grouping instructions are generated based on the link inference results.

[0112] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0113] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the method for constructing the information grouping function provided by the above methods. The method includes: acquiring information to be queried and a target knowledge graph; wherein the target knowledge graph is constructed based on emergencies, operational indicators and metadata in rail transit. An incremental dynamic graph is constructed based on the information to be queried and the target knowledge graph; The incremental dynamic graph and the information to be queried are input into the link prediction inference model to obtain the link inference result; wherein, the link prediction inference model incorporates a graph attention mechanism; Information grouping instructions are generated based on the link inference results.

[0114] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for constructing information grouping functions provided by the above methods. The method includes: acquiring information to be queried and a target knowledge graph; wherein the target knowledge graph is constructed based on emergencies, operational indicators, and metadata in rail transit. An incremental dynamic graph is constructed based on the information to be queried and the target knowledge graph; The incremental dynamic graph and the information to be queried are input into the link prediction inference model to obtain the link inference result; wherein, the link prediction inference model incorporates a graph attention mechanism; Information grouping instructions are generated based on the link inference results.

[0115] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0116] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing an information grouping function, characterized in that, include: Obtain the information to be queried and the target knowledge graph; wherein, the target knowledge graph is constructed based on emergencies, operational indicators and metadata in rail transit; An incremental dynamic graph is constructed based on the information to be queried and the target knowledge graph; The incremental dynamic graph and the information to be queried are input into the link prediction inference model to obtain the link inference result; wherein, the link prediction inference model incorporates a graph attention mechanism; Information grouping instructions are generated based on the link inference results.

2. The method for constructing the information grouping function according to claim 1, characterized in that, Before obtaining the information to be queried and the target knowledge graph, the method further includes: Construct a first knowledge graph for emergencies in rail transit and the arguments of each emergency; wherein the arguments include operational indicators; Construct a second knowledge graph for the operational metrics and their metadata; The first knowledge graph and the second knowledge graph are aligned to obtain the target knowledge graph.

3. The method for constructing the information grouping function according to claim 1, characterized in that, The link prediction inference model includes an encoder and a decoder; the step of inputting the incremental dynamic graph and the query information into the link prediction inference model to obtain the link inference result includes: Based on the encoder, the incremental dynamic graph is forward-propagated to obtain encoded data; wherein, the encoder incorporates a graph attention mechanism. Based on the decoder and the encoded data, link prediction inference is performed to obtain the link inference result; wherein, the link inference result includes multiple candidate predicted links.

4. The method for constructing the information grouping function according to claim 1, characterized in that, The training process of the link prediction inference model includes: Obtain a training dataset; wherein the training dataset includes multiple training samples based on training burst events, training operation metrics, and training metadata; Based on the target knowledge graph and the training samples, the initial link prediction inference model is iteratively trained to determine the model parameters of the initial link prediction inference model, and the link prediction inference model is obtained based on the model parameters; wherein, the model parameters include independent weight matrix and diagonal matrix parameters for the relationship between the sudden event, the operational indicators and the metadata, and graph attention weight parameters for the importance of the operational indicators.

5. The method for constructing the information grouping function according to claim 4, characterized in that, The iterative training of the initial link prediction inference model based on the target knowledge graph and the training samples includes: Based on the target knowledge graph, the training samples, and the training loss, the initial link prediction inference model is iteratively trained; wherein, the training loss includes the main loss and the regularization loss, and the regularization loss is the relative entropy loss calculated when there is an argument relationship between the sudden event and the operational indicator.

6. The method for constructing the information grouping function according to any one of claims 1 to 5, characterized in that, The construction of an incremental dynamic knowledge graph based on the information to be queried and the target knowledge graph includes: Using the query event in the query information as the root node, the first-level neighbor node and second-level neighbor node of the query event are determined in the target knowledge graph to obtain the incremental dynamic graph.

7. The method for constructing the information grouping function according to any one of claims 1 to 5, characterized in that, The generation of information clustering instructions based on the link inference results includes: Obtain a predefined instruction template, and generate information grouping instructions based on the instruction template and the link inference result; The information grouping instructions are embedded into the unified dispatch and command system of the network to obtain the instruction mapping scheme corresponding to dispatch and command and emergency response.

8. A device for constructing an information grouping function, characterized in that, include: The first acquisition module is configured to acquire the information to be queried and the target knowledge graph; wherein, the target knowledge graph is constructed based on emergencies, operational indicators and metadata in rail transit; The first construction module is configured to construct an incremental dynamic graph based on the information to be queried and the target knowledge graph; The input module is configured to input the incremental dynamic graph and the information to be queried into the link prediction inference model to obtain the link inference result; wherein, the link prediction inference model incorporates a graph attention mechanism; The generation module is configured to generate information grouping instructions based on the link inference results.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method for constructing the information grouping function as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method for constructing the information grouping function as described in any one of claims 1 to 7.

Images