Conflict event risk prediction method based on time sequence knowledge graph

Abstract

The invention relates to a conflict event risk prediction method based on a time sequence knowledge graph, and the method comprises the steps: constructing a conflict event knowledge graph according to conflict event data; according to the conflict event knowledge graph, generating a graph snapshot in corresponding time; obtaining a time sequence prediction model; the continuous map snapshots serve as input of a time sequence prediction model, the continuous map snapshots have missing entities or relations, and the time sequence prediction model outputs prediction classification values; calculating a classification loss value, and updating model parameters according to the classification loss value; and carrying out conflict event risk prediction at the next moment. The method has the beneficial effects that the time sequence knowledge graph and the historical characteristics of the conflict events are combined, so that the model can accurately predict the occurrence possibility of risk events at future moments, and related prevention is carried out on the conflict events.

Description

Technical Field

[0001] The present invention relates to the technical field of risk prediction, and more specifically, to a method for predicting conflict event risks based on a time series knowledge graph. Background Art

[0002] With the increase in conflict events, how to accurately predict the risks of conflict events has become a hot topic of research. Early conflict prediction theories focused on explanatory research, which was limited by subjective judgment and experience, data acquisition capabilities, and the passage of time, resulting in poor performance in prediction. Existing risk event prediction methods include static knowledge graph prediction and time series knowledge graph prediction. Static knowledge graph prediction lacks real-time performance and cannot capture dynamic changes, and cannot better predict risk events based on timeliness. In the existing time series graph model, the RE-GCN model has a high computational complexity for graph convolution operations and does not fully utilize graph node features. The RE-NET model takes too long to train and has limited ability to process large-scale data. In addition, existing methods often lack effective use of event history and time characteristics, resulting in low prediction accuracy.

[0003] Therefore, it is necessary to propose a new conflict risk event prediction method that utilizes the characteristics of time series knowledge graphs to perform efficient information screening and identification, quantitatively analyze conflict risks, learn the evolution laws of past conflict events, and output specific risk probability values ​​to solve the technical problems of existing methods. Summary of the invention

[0004] The purpose of the present invention is to address the deficiencies of the prior art and propose a conflict event risk prediction method based on a time series knowledge graph.

[0005] In the first aspect, a conflict event risk prediction method based on a time series knowledge graph is provided, comprising:

[0006] Step 1: Construct a conflict event knowledge graph based on conflict event data;

[0007] Step 2: Generate a graph snapshot at the corresponding time based on the conflict event knowledge graph; and obtain a time series prediction model; use continuous graph snapshots as input to the time series prediction model, where the continuous graph snapshots have missing entities or relationships, and the time series prediction model outputs a predicted classification value, which is used to indicate the probability of correctly predicting the missing entity or relationship;

[0008] Step 3: Calculate the classification loss value according to the predicted classification value, and update the model parameters according to the classification loss value, and repeatedly perform steps 2 and 3 until the classification loss value tends to be stable;

[0009] Step 4: Predict the risk of conflict events at the next moment.

[0010] Preferably, step 1 comprises:

[0011] Step 1.1, obtaining conflict event data from multiple sources, and preprocessing the conflict event data to obtain text data;

[0012] Step 1.2, extracting knowledge from the text data, wherein the knowledge extraction includes entity recognition and relationship extraction to form a time series knowledge graph; the time series knowledge graph is in a four-tuple format (s, r, o, t), where s represents the head entity, r represents the relationship, o represents the tail entity, and t represents the timestamp;

[0013] Step 1.3: Align the entity representations in the conflict event data from multiple sources, complete knowledge fusion, construct a conflict event knowledge graph, and store and visualize the conflict event knowledge graph.

[0014] Preferably, in step 1.1, conflict event data from multiple sources are obtained from public data sets; the conflict event data include semi-structured CSV data and unstructured text data; and the preprocessing includes: data cleaning, data standardization and data filtering.

[0015] Preferably, in step 2, generating a graph snapshot at a corresponding time according to the conflict event knowledge graph includes:

[0016] According to different timestamps, the entire conflict event knowledge graph is divided into a series of graph snapshots, each of which records all facts within the current timestamp t.

[0017] Preferably, the time series prediction model is composed of a time series evolution module, a graph aggregation module, a time encoding module, a history repetition module and a time-aware decoder module;

[0018] Among them, the temporal evolution module is used to output the aggregated graph entity embedding matrix based on the input continuous graph snapshots; the module is used to aggregate the discrete graph snapshots into one graph; the time encoding module is used to provide the order and interval information about different timestamps in the input data; the historical repetition module is used to provide global constraints; and the time-aware decoder module is used to fuse feature information and calculate the predicted classification value.

[0019] Preferably, the time encoding module encodes the time interval of the event, and the formula is as follows:

[0020]

[0021] Among them, w1,...,w d are learnable frequency parameters, p1,...,p dis a learnable phase parameter, d is the dimension of the vector embedding, t′ is the timestamp of the neighbor node in the aggregation graph, t is the timestamp of the query node, tt′ is the time interval between two nodes, and vt′ is the vector representation after time encoding.

[0022] Preferably, the formula for fusing feature information in the time-aware decoder module is:

[0023]

[0024] in, is the evolving embedding representation of entity s, is the embedding representation of entity s in the aggregate graph, g e ∈R d is a learnable gate vector parameter, ⊙ is the element-wise product, and σ is the sigmoid activation function; in this module, h Fin To embed a single entity, h is calculated for all single entities. Fin After that, we get the unified entity embedding matrix H fin .

[0025] In a second aspect, a conflict event risk prediction system based on a time series knowledge graph is provided, which is used to execute any method described in the first aspect, including:

[0026] A construction module, used to construct a conflict event knowledge graph based on conflict event data;

[0027] A generation module is used to generate a graph snapshot at a corresponding time according to the conflict event knowledge graph; and obtain a time series prediction model; use continuous graph snapshots as input to the time series prediction model, wherein the continuous graph snapshots have missing entities or relationships, and the time series prediction model outputs a predicted classification value, wherein the predicted classification value is used to indicate the probability of correctly predicting the missing entity or relationship;

[0028] A calculation module, used for calculating the classification loss according to the predicted classification value;

[0029] The prediction module is used to predict the risk of conflict events at the next moment.

[0030] According to a third aspect, a computer storage medium is provided, wherein a computer program is stored in the computer storage medium; when the computer program is executed on a computer, the computer executes any method described in the first aspect.

[0031] In a fourth aspect, an electronic device is provided, including:

[0032] Memory, used to store computer programs;

[0033] A processor is used to execute the computer program to implement any method as described in the first aspect.

[0034] The beneficial effects of the present invention are as follows: the present invention first obtains public data resources and preprocesses based on the time series of the data to construct a time series knowledge graph related to conflict events, and the graph contains a plurality of groups of four-tuples (head entity, relationship, tail entity, time) of time periods. Secondly, a time series prediction model is constructed, which includes five modules, a time series evolution module, a graph aggregation module, a time encoding module, a historical repetition module, and a time-aware decoder module. The model takes a continuous graph snapshot of length k as the input of the model. After internal calculation of the model, the decoder outputs the entity or relationship that can make the missing fact true at the next moment. Among them, the time series evolution module is used to capture the structural features and evolutionary features of the graph snapshot over time, and the aggregation graph module aggregates the discrete graph snapshots into a graph, associates the entities across timestamps, enriches the interactive information between graphs, and automatically assigns weight coefficients to each edge through the attention mechanism, thereby learning the intrinsic expression semantics of entities and relationships. The model combines the time series knowledge graph and the historical features of conflict events, so that the model can more accurately predict the possibility of risk events in the future and prevent conflict events. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A flowchart of a conflict event risk prediction method based on a time series knowledge graph provided in an embodiment of the present invention;

[0036] Figure 2 A flowchart of another conflict event risk prediction method based on a time series knowledge graph provided by an embodiment of the present invention;

[0037] Figure 3 A schematic diagram of the structure of a conflict event risk prediction system based on a time series knowledge graph provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The present invention is further described below in conjunction with embodiments. The description of the following embodiments is only used to help understand the present invention. It should be noted that for ordinary persons in the art, without departing from the principle of the present invention, the present invention can also be modified in some ways, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

[0039] Embodiment 1:

[0040] In order to solve the problems of the prior art, Embodiment 1 of the present application provides a method for predicting the risk of conflict events based on a time series knowledge graph, including:

[0041] Step 1: Build a conflict event knowledge graph based on conflict event data.

[0042] Step 1 includes:

[0043] Step 1.1: Obtain conflict event data from multiple sources, and pre-process the conflict event data to obtain text data.

[0044] For example, in step 1.1, conflict event data from multiple sources are obtained from public data sets; the conflict event data includes semi-structured CSV data and unstructured text data; and the preprocessing includes: data cleaning, data standardization, and data filtering.

[0045] Step 1.2: extract knowledge from the text data, including entity recognition and relationship extraction, to form a time series knowledge graph; the time series knowledge graph is in a four-tuple format (s, r, o, t), where s represents the head entity, r represents the relationship, o represents the tail entity, and t represents the timestamp.

[0046] Step 1.3: Align the entity representations in the conflict event data from multiple sources, complete knowledge fusion, construct a conflict event knowledge graph, and store and visualize the conflict event knowledge graph.

[0047] Step 2: Generate a graph snapshot at the corresponding time based on the conflict event knowledge graph; and obtain a time series prediction model; use continuous graph snapshots as input to the time series prediction model, where the continuous graph snapshots have missing entities or relationships, and the time series prediction model outputs a predicted classification value, which is used to represent the probability that the predicted missing entity or relationship is correct.

[0048] In step 2, generating a graph snapshot at a corresponding time based on the conflict event knowledge graph includes:

[0049] According to different timestamps, the entire conflict event knowledge graph is divided into a series of graph snapshots, each of which records all facts within the current timestamp t.

[0050] Furthermore, the time series prediction model consists of a time series evolution module, a graph aggregation module, a time encoding module, a history repetition module and a time-aware decoder module;

[0051] Among them, the temporal evolution module is used to output the aggregated graph entity embedding matrix based on the input continuous graph snapshots; the module is used to aggregate the discrete graph snapshots into one graph; the time encoding module is used to provide the order and interval information about different timestamps in the input data; the historical repetition module is used to provide global constraints; and the time-aware decoder module is used to fuse feature information and calculate the predicted classification value.

[0052] Step 3: Calculate the classification loss value according to the predicted classification value, and update the model parameters according to the classification loss value, and repeatedly perform steps 2 and 3 until the classification loss value tends to be stable.

[0053] Step 4: Predict the risk of conflict events at the next moment.

[0054] Embodiment 2:

[0055] Based on Example 1, Example 2 of the present application provides a more specific method for predicting the risk of conflict events based on a time series knowledge graph, such as Figure 2 As shown, including:

[0056] Step 1: Obtain international risk conflict data from multiple sources from public datasets and store them in the same data format.

[0057] Step 2: preprocess the semi-structured CSV data and unstructured text data obtained in step 1; the data preprocessing includes: data cleaning, data standardization and data filtering; and define the top-level abstract concepts and model the entity relationships.

[0058] Step 2 includes:

[0059] Step 2.1: Obtain conflict event data from multiple data sources (such as GDELT and ACLED) and remove redundant, erroneous, and incomplete information. Discard data with missing key fields, and then filter out valid data based on the positive and negative values ​​of GoldsteinScale and the classification of QuadClass.

[0060] Step 2.2: Standardize the event information in different data sources to ensure that the format and content of the data are consistent.

[0061] Step 2.3: Filter out low-quality events through predefined rules and algorithms, and retain high-quality conflict events. Calculate the index R using the conflict index calculation formula and remove unimportant events.

[0062] Step 3: perform knowledge extraction on the text data, including entity recognition and relationship extraction, to form a temporal knowledge graph in a four-tuple format (s, r, o, t), where s represents the head entity, r represents the relationship, o represents the tail entity, and t represents the timestamp.

[0063] Step 3 includes:

[0064] Step 3.1, use Jieba word segmentation to perform named entity recognition on the text;

[0065] Step 3.2: Use the PURE model for end-to-end relationship extraction, retrain the PURE model on the dataset, label and classify the entities, and use the trained model for knowledge extraction from text data.

[0066] Step 4: Align different entity representations in multi-source data, complete knowledge fusion, and construct the required conflict event knowledge graph.

[0067] Step 4 includes:

[0068] Step 4.1: Integrate event information from different data sources into a unified time series knowledge graph framework to form a high-quality time series knowledge graph quadruple (s, r, o, t).

[0069] Step 4.2: Build a graph snapshot sequence containing multiple timestamps, where each snapshot records all event facts within the current timestamp.

[0070] Step 5: Store and visualize the graph data in a unified form.

[0071] Step 5 includes

[0072] Step 5.1: Use Neo4j to process the processed conflict event data

[0073] Step 5.2: Concatenate the time information and the relationship content into strings to store the time information indirectly. Through the Cypher query statement, you can quickly locate the node of interest.

[0074] Step 6. Divide the quadruple according to the timestamp and generate the graph snapshot Gt at the corresponding time t. The conflict event risk prediction task is transformed into: given a missing quadruple (s, r, ?, t) or (?, r, o, t) or (s, ?, o, t), according to the proposed conflict event risk prediction model based on graph attention network, multiple consecutive timestamps are used as input. The model predicts the probability of the missing entity or relationship being correct.

[0075] Step 6 includes:

[0076] Step 6.1: Preprocess the quadruple. According to different timestamps, the entire graph can be divided into a series of graph snapshots. Each snapshot records all facts within the current timestamp t, specifically G = {G1, G2, ..., GT}.

[0077] Step 6.2: Continuous graph sequence G with a time length of k t-k:t-1 As input, the aggregate graph entity embedding matrix H is obtained after entering the time series evolution module g .

[0078] In step 6.2, the temporal evolution module focuses on capturing the characteristics of the graph evolution over time and the structural features within the aggregated graph snapshot to form an embedded representation of entities and relationships. Internally, a GCN aggregates the entity relationship representation of the current graph snapshot, and two GRUs learn the temporal evolution characteristics of entities and relationships respectively.

[0079] Furthermore, the two-layer stacked GCN fuses the embedding representations of relations and entities together through a one-dimensional convolution, as shown in the following formula:

[0080]

[0081] in, and is the embedding representation of the head entity s and the tail entity o at the lth layer at time t, rt is the embedding representation of the relationship, and is a learnable weight parameter, c o is a normalization factor, whose value is the in-degree of node o, || represents vector concatenation, ψ represents a one-dimensional convolution operator, and the embedding vectors of entities and relationships are fused through convolution operations, which is more expressive than simple vector addition operations. σ represents the RReLU activation function, which results in the entity embedding representation of the l+1th layer. After all entities are aggregated and propagated through multiple layers of GCN, the entity embedding matrix is ​​obtained.

[0082] The entity GRU component of this module is used to update the entity embedding component. The formula is as follows:

[0083]

[0084] Among them, is the entity embedding matrix at time t and time t-1, d is the dimension of vector embedding, H t ∈ is the entity embedding matrix after multi-layer GCN aggregation at time t-1.

[0085] The embedding representation of the relation is also updated through the relational GRU component, and its formula is as follows:

[0086] r′ t =[pooling(H t-1 ,H r,t )||r]

[0087] R t =GRU(R t-1 ,R' t )

[0088] Among them, H r,t is the embedding matrix of all entities associated with relation r at time t, r is the embedding vector of the relation, r' t Yes Ht-1 With H r,t Calculate all r' through the intermediate vector representation after mean pooling t Composition R' t , R t By R t-1 and R' t Obtained through updating of the relational GRU component.

[0089] Step 6.3: Generate a query-oriented aggregate graph based on the graph sequence; the aggregate graph module receives multiple graph snapshots with consecutive timestamps as input, which include relevant information of the conflict event. The aggregate graph contains all the quadruples related to the query entity and some neighbor nodes obtained by weighted random sampling, which are used as input of GAT to obtain the entity representation UEt of the aggregate graph.

[0090] In step 6.3, the aggregation graph module aggregates discrete graph snapshots into a graph, associates entities across timestamps, and enriches the interactive information between graphs. First, all facts related to the query entity are found based on all current query entities, and facts that are closer to the query entity in time are retained through the time-aware exponential weighted sampling algorithm. After obtaining the query-oriented aggregation graph, the aggregation graph is used as the input of the graph attention network (GAT), and GAT automatically assigns attention coefficients to the edges, and finally the entity embedding matrix of the aggregation graph is obtained for subsequent processing.

[0091] Step 6.4: This module needs to perform decoding prediction. Based on the ConvTransE decoder and combined with the time vector generated by the time encoding module, the original time-aware decoder and the historical time-aware decoder are constructed for score prediction.

[0092] In step 6.4, the time encoding module provides the model with information about the order and interval of different timestamps in the input data, which helps the model better understand the temporal structure in the data. The module encodes the time interval of events, and the formula is as follows:

[0093]

[0094] Among them, w1,...,w d are learnable frequency parameters, p1,...,p d is a learnable phase parameter, d is the dimension of the vector embedding, t′ is the timestamp of the neighbor node in the aggregation graph, t is the timestamp of the query node, tt′ is the time interval between two nodes, and vt′ is the vector representation after time encoding.

[0095] Step 6.5: This module borrows the replication mode of CyGNet to impose global history constraints on the prediction results.

[0096] In step 6.5, the historical repetition module provides global constraints for the prediction model. This module constrains the prediction scope to historically repeated entities and relations, focusing on generating the most likely repeated entities and relations. Given a query (s, r, ?, t), this module accepts a graph sequence G 1:t-1 As input, statistics G 1:t-1 All the tail entities in the relationship r with entity s are recorded in a sparse matrix. If the entity at the corresponding position has appeared in history, it is recorded as 1, otherwise it is recorded as 0; get the historical entity set

[0097] Step 6.6 aggregates the entity representations obtained in step 5.3 and step 5.4 through the gating unit, and finally calculates the predicted probability value through the ConvTransE decoder.

[0098] In step 6.6, feature information is fused through the gating function and the time-aware decoder module. The calculation formula is as follows:

[0099]

[0100] in, is the evolving embedding representation of entity s, is the embedding representation of entity s in the aggregate graph, g e ∈R d is a set of learnable gating vector parameters, ⊙ is the element-wise product, and σ is the sigmoid activation function. Calculate h for all entities Fin After that, we can get a unified entity embedding matrix H fin .

[0101] In addition, in step 6.6, after obtaining the embedded representation of nodes and relationships, they need to be decoded and predicted. This module is also based on the ConvTransE decoder. Combined with the time vector generated by the time encoding module, the original time-aware decoder and the historical time-aware decoder are constructed for score prediction. The calculation formulas of the original time-aware decoder score function PR and the historical time-aware decoder score function PH are as follows:

[0102] ConvTransE(s,r,t)=ReLU(vec(M conv )W6)

[0103] P R (o|s,r,t)=softmax(ConvTransE(s,r,t)H fin

[0104]

[0105] in, is a learnable parameter matrix, vec is a feature map, is the historical candidate entity matrix. The final prediction score P score The weighted sum of the two decoder outputs is obtained by a hyperparameter α, and the calculation formula is as follows:

[0106] P score =α×P H (o|s,r,t)+(1-α)×P R (o|s,r,t)

[0107] Step 6.7 uses the cross entropy loss function to calculate the classification loss of entities and relations.

[0108] Step 7: Calculate the classification loss value according to the predicted classification value, and update the model parameters according to the classification loss value, and repeatedly perform steps 6 and 7 until the classification loss value tends to be stable.

[0109] In step 7, the cross entropy loss function is used to calculate the loss, and its formula is as follows:

[0110]

[0111] in, and They are the label values ​​corresponding to the entity and relationship respectively. If the prediction is correct, it is 1, and if the prediction is wrong, it is 0.

[0112] Step 8. Predict the risk of conflict events at a certain time T+1: Given a missing quadruple (s, r, ?, t) or (?, r, o, t) or (s, ?, o, t), the model predicts the missing entity or relationship and completes the risk prediction.

[0113] It should be noted that the parts in this embodiment that are the same or similar to those in Embodiment 1 can be referenced to each other and will not be described in detail in this application.

[0114] Embodiment 3:

[0115] Based on Example 1, Example 3 of the present application provides a conflict event risk prediction system based on a time series knowledge graph, such as Figure 3 As shown, including:

[0116] A construction module, used to construct a conflict event knowledge graph based on conflict event data;

[0117] A generation module is used to generate a graph snapshot at a corresponding time according to the conflict event knowledge graph; and obtain a time series prediction model; use continuous graph snapshots as input to the time series prediction model, wherein the continuous graph snapshots have missing entities or relationships, and the time series prediction model outputs a predicted classification value, wherein the predicted classification value is used to indicate the probability of correctly predicting the missing entity or relationship;

[0118] A calculation module, used for calculating the classification loss according to the predicted classification value;

[0119] The prediction module is used to predict the risk of conflict events at the next moment.

[0120] Specifically, the system provided in this embodiment is a system corresponding to the method provided in Embodiment 1. Therefore, the parts in this embodiment that are the same or similar to those in Embodiment 1 can be referenced to each other and will not be repeated in this application.

Claims

1. A conflict event risk prediction method based on time series knowledge graph, characterized in that: include: Step 1: Construct a conflict event knowledge graph based on conflict event data; Step 2: Generate a graph snapshot at the corresponding time based on the conflict event knowledge graph; and obtain a time series prediction model; use continuous graph snapshots as input to the time series prediction model, where the continuous graph snapshots have missing entities or relationships, and the time series prediction model outputs a predicted classification value, which is used to indicate the probability of correctly predicting the missing entity or relationship; Step 3: Calculate the classification loss value according to the predicted classification value, and update the model parameters according to the classification loss value, and repeatedly perform steps 2 and 3 until the classification loss value tends to be stable; Step 4: Predict the risk of conflict events at the next moment.

2. The method for predicting conflict event risk based on time series knowledge graph according to claim 1 is characterized in that: Step 1 includes: Step 1.1, obtaining conflict event data from multiple sources, and preprocessing the conflict event data to obtain text data; Step 1.2, extracting knowledge from the text data, wherein the knowledge extraction includes entity recognition and relationship extraction to form a time series knowledge graph; the time series knowledge graph is in a four-tuple format (s, r, o, t), where s represents the head entity, r represents the relationship, o represents the tail entity, and t represents the timestamp; Step 1.3: Align the entity representations in the conflict event data from multiple sources, complete knowledge fusion, construct a conflict event knowledge graph, and store and visualize the conflict event knowledge graph.

3. The method for predicting conflict event risk based on time series knowledge graph according to claim 2 is characterized in that: In step 1.1, conflict event data from multiple sources are obtained from public data sets; the conflict event data includes semi-structured CSV data and unstructured text data; the preprocessing includes: data cleaning, data standardization and data filtering.

4. The method for predicting conflict event risk based on time series knowledge graph according to claim 2 is characterized in that: In step 2, generating a graph snapshot at a corresponding time based on the conflict event knowledge graph includes: According to different timestamps, the entire conflict event knowledge graph is divided into a series of graph snapshots, each of which records all facts within the current timestamp t.

5. The method for predicting conflict event risk based on time series knowledge graph according to claim 4 is characterized in that: The time series prediction model consists of a time series evolution module, a graph aggregation module, a time encoding module, a history repetition module and a time-aware decoder module; Among them, the temporal evolution module is used to output the aggregated graph entity embedding matrix based on the input continuous graph snapshots; the module is used to aggregate the discrete graph snapshots into one graph; the time encoding module is used to provide the order and interval information about different timestamps in the input data; the historical repetition module is used to provide global constraints; and the time-aware decoder module is used to fuse feature information and calculate the predicted classification value.

6. The method for predicting conflict event risk based on time series knowledge graph according to claim 5 is characterized in that: The time encoding module encodes the time interval of events. The formula is as follows: Among them, w1,...,w d are learnable frequency parameters, p1,...,p d is a learnable phase parameter, d is the dimension of the vector embedding, t′ is the timestamp of the neighbor node in the aggregation graph, t is the timestamp of the query node, tt′ is the time interval between two nodes, and vt′ is the vector representation after time encoding.

7. The method for predicting conflict event risk based on time series knowledge graph according to claim 6 is characterized in that: The formula for fusing feature information in the time-aware decoder module is: in, is the evolving embedding representation of entity s, is the embedding representation of entity s in the aggregate graph, g e is a learnable gate vector parameter, ⊙ is the element-wise product, and σ is the sigmoid activation function; in this module, h Fin To embed a single entity, h is calculated for all single entities. Fin After that, we get the unified entity embedding matrix H fin .

8. A conflict event risk prediction system based on time series knowledge graph, characterized in that: Used to perform the method according to any one of claims 1 to 7, comprising: A construction module, used to construct a conflict event knowledge graph based on conflict event data; A generation module is used to generate a graph snapshot at a corresponding time according to the conflict event knowledge graph; and obtain a time series prediction model; use continuous graph snapshots as input to the time series prediction model, wherein the continuous graph snapshots have missing entities or relationships, and the time series prediction model outputs a predicted classification value, wherein the predicted classification value is used to indicate the probability of correctly predicting the missing entity or relationship; A calculation module, used for calculating the classification loss according to the predicted classification value; The prediction module is used to predict the risk of conflict events at the next moment.

9. A computer storage medium, characterized in that The computer storage medium stores a computer program; when the computer program is executed on a computer, the computer executes any one of the methods described in claims 1 to 7.

10. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the method according to any one of claims 1 to 7.

Images