A social event prediction method based on fourier graph learning

Abstract

The application discloses a social event prediction method based on Fourier graph learning, relates to the technical field of network mining, deep learning and social computing, and comprises the following steps: acquiring a social event data stream, parsing the social event data stream into four tuples, and constructing an event graph sequence arranged in time sequence; respectively encoding a real part and an imaginary part, wherein the imaginary part is encoded, time is converted into a frequency domain through fast Fourier transform, mixed coding learning in the frequency domain is performed, and the enhanced time domain mode is restored through inverse transform; on the basis, a Fourier graph containing node and edge complex number features is constructed, and a Fourier space multiplier is introduced, the complex number features are interactively operated in the frequency domain to replace the traditional space-time convolution operation in the time domain, and the calculation complexity is significantly reduced; and finally, a complex-valued event decoder is constructed. While ensuring accurate capture of long-term and short-term time dependence, the method effectively solves the problems of large calculation overhead and slow reasoning speed of existing graph neural networks in processing large-scale data streams.

Description

Technical Field

[0001] This application relates to the fields of web mining, deep learning and social computing, and in particular to a social event prediction method based on Fourier graph learning. Background Technology

[0002] With the rapid development of internet technology, the internet has become the primary venue for human social activities and information dissemination. Massive amounts of online news and social media posts constitute a real-time updated network of internet events. Social event prediction aims to utilize this historical data to uncover the evolutionary patterns and causal relationships between events, thereby predicting potential future events. This has extremely important strategic significance for corporate risk avoidance and social governance.

[0003] In existing technologies, social event prediction is typically modeled as a temporal knowledge graph reasoning problem, employing a combination of recurrent neural networks and graph neural networks to model the temporal and semantic relationships between events. However, these technologies face the dual challenges of difficulty in capturing long-term temporal dependencies and high computational complexity when processing internet social event streams. Specifically, existing methods cannot locate events with significant impact over long time spans. Furthermore, the use of temporal spatiotemporal convolutional kernels in existing methods hinders rapid prediction when dealing with large numbers of events. Designing a mechanism that can accurately capture long-term recurring patterns while operating efficiently on large-scale data is a pressing technical challenge. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, this application provides a social event prediction method based on Fourier graph learning, which aims to solve the problems of recurrent neural networks being unable to capture long-term dependencies, graph neural networks having excessively high computational complexity, and poor model interpretability in the prior art.

[0005] To achieve the above objectives, this application adopts the following technical solution, including: A social event prediction method based on Fourier graph learning includes the following steps: S1, acquire the social event data stream within the historical time window, parse each social event into a quadruple containing the event subject, relationship type, event object, and timestamp, construct an event graph based on the quadruples with the same timestamp, and arrange them in chronological order to obtain an event graph sequence; the timestamp is the time when the event occurred; S2, for each social event quadruple, encodes the real part of the event subject, event object, and event relationship respectively; S3. Based on the real part encoding results and the event occurrence time, perform imaginary part encoding for the event subject, event object, and event relationship respectively; S4. Based on the results of real part encoding and imaginary part encoding, a Fourier diagram is constructed, and the original event diagram quadruples are transformed into complex triples containing real part encoding values ​​and imaginary part encoding values; the complex triples are composed of complex representations of event subject, event object, and event relationship; S5. Construct a Fourier space multiplier to perform interactive operations on the complex features of the event subject, relation type, and event object in the complex domain to obtain the predicted complex representations of the event subject, relation type, and event object. S6 constructs a complex-valued event decoder based on the predicted complex representations of the event subject, relation type, and event object, and outputs the prediction results of a specific event occurring at a future timestamp.

[0006] Preferably, in step S2, for each social event quadruple... The real parts of the event subject, event object, and event relationship are encoded respectively, as shown below: ; ; ; in, As the main body of the event, As the object of the event, For relational types, The timestamp is the time when the event occurred. Indicates One-Hot encoding; , , These are weight matrices for the event subject, relationship type, and event object, respectively. These are bias matrices for the event subject, relationship type, and event object, respectively. These are the real part encoding values ​​for the event subject, relation type, and event object, respectively. This represents the activation function.

[0007] Preferably, in step S3, based on the real part encoding result and combined with the event occurrence time, the imaginary part encoding is performed on the event subject, event object, and event relationship, as shown below: First, regarding historical time windows Discrete time step Perform Fourier encoding: ; ; ; in, It is a historical time series; The first time series after being transformed into the frequency domain One frequency component; Frequency index; for The first in One time step; For time indexing; The length of the historical time window; This represents the total number of sampling points for the Fourier transform. The Fast Fourier Transform (FFT) aims to transform discrete time steps within a historical time window. All data is converted to the frequency domain for modeling, resulting in a time-frequency domain representation. ; Based on the obtained time-frequency domain characterization Using weight matrix , , Bias matrix Real part encoded value Frequency domain hybrid coding learning: ; ; ; in, Indicates the activation function; These are the event subject, relationship type, and event object related to the time of event occurrence. Frequency domain hybrid coding; Will Extracted as a time-domain representation: ; ; ; in, This represents the total number of sampling points for the Fourier transform. These are the event subject, relationship type, and event object related to the time of event occurrence. The time-domain representation; Frequency index; For time indexing; , , The first part of the frequency domain hybrid coding of event subject, relation type, and event object respectively One frequency component; Frequency domain hybrid coding results are obtained using inverse fast Fourier transform. Convert to imaginary part encoded value: ; ; ; in, Indicates the inverse fast Fourier transform; These are the imaginary part encoding values ​​for the event subject, relationship type, and event object, respectively.

[0008] Preferably, in step S4, the complex triplet As shown below: ; in, These are respectively plural representations of the event subject, relationship type, and event object; The imaginary unit; These are the imaginary part encoding values ​​for the event subject, relationship type, and event object, respectively. These are the real part encoding values ​​for the event subject, relation type, and event object, respectively. As the main body of the event, As the object of the event, For relational types, The timestamp is the time when the event occurred. A Fourier plot Represented as a set of multiple complex triples: ;in, The number of complex triples contained in a Fourier plot.

[0009] Preferably, in step S5, frequency domain features are aggregated for each complex triplet in three ways: By masking complex triplets In and use the remaining The masked part is fitted using complex multiplication: ; in, The item represents pure semantic fusion. The item serves as a penalty for time pattern matching. Item and The terms represent the cross-driving effect of semantics and temporal patterns. For the prediction of complex representation of the event object; By masking complex triplets In and use the remaining The masked part is fitted using complex multiplication: ; in, The item represents pure semantic fusion. The item serves as a penalty for time pattern matching. Item and The terms represent the cross-driving effect of semantics and temporal patterns. For predicting complex representations of relation types; By masking complex triplets In and use the remaining The masked part is fitted using complex multiplication: ; in, The item represents pure semantic fusion. The item serves as a penalty for time pattern matching. Item and The terms represent the cross-driving effect of semantics and temporal patterns. The predicted complex representation of the event subject.

[0010] Preferably, in step S6, the similarity scoring function of the real and imaginary parts is used to calculate respectively. and between, and Between and and The similarity score between the real and imaginary parts is defined as follows: ; ; ; in, They are respectively and between, and Between and and Similarity score between them; and These represent the real and imaginary parts of a complex number, respectively. The modulus penalty coefficient; Indicates taking the absolute value; The similarity scores are mapped to the classification feature space, and the normalized probability distributions of all event subjects, relationship types, and event objects are output. All event subjects, relationship types, and event objects are ranked according to their confidence scores, and the event subjects, relationship types, and event objects with the highest confidence scores are selected as the quadruplet prediction results.

[0011] Preferably, it also includes: constructing a real-to-virtual part similarity matching loss based on similarity scores. Its expression is: ; The real-imaginary part similarity matching loss is combined with classification cross-entropy loss, norm regularization term loss, and temporal difference regularization loss to construct a joint loss function. Optimize the network.

[0012] Preferably, in step S1, in the event graph, the event subject and the event object are used as nodes, and edges are constructed based on the relationship type.

[0013] This application also provides a readable storage medium having a computer program stored thereon, which, when executed, implements the aforementioned social event prediction method based on Fourier graph learning.

[0014] This application also provides an electronic device, which includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned social event prediction method based on Fourier graph learning.

[0015] The advantages of this application are: (1) This application relates to a social event prediction method based on frequency domain analysis and graph neural network, which uses a Fourier graph learning framework to process large-scale, dynamic Internet social event stream data with complex spatiotemporal dependencies.

[0016] (2) This application transforms the spatiotemporal convolution in the time domain into efficient multiplication in the frequency domain by utilizing the convolution theorem. Combined with the decoupling encoding capability of complex vector space for entity semantics and temporal patterns, it significantly reduces the computational complexity of graph neural networks under large-scale data. At the same time, it introduces a semantic-aware frequency filtering mechanism, which uses the filter to adaptively extract long and short-term time dependencies in the frequency domain to overcome the forgetting problem. Combined with the real and imaginary part similarity scoring function of complex numerical event decoder, it effectively solves the problems of recurrent neural networks being unable to capture long-term dependencies and graph neural networks having large computational overhead in existing social event prediction methods, thereby improving the accuracy of social event prediction and real-time monitoring capability. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process of the social event prediction method of this application; Figure 2 This is a schematic diagram of the data processing procedure for the social event prediction method of this application; Figure 3 This is a schematic diagram comparing training time in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Depend on Figure 1 As shown, a social event prediction method based on Fourier diagrams includes the following steps: S1, Data Acquisition and Graph Sequence Construction: Acquire the Web social event data stream within the historical time window, parse each social event into a quadruple containing the event subject, relationship type, event object, and timestamp, and reorganize the quadruple into an event graph sequence arranged in chronological order based on the timestamp.

[0020] The data acquisition and graph sequence construction process in step S1 includes: Define entity collection and relation set For any social event collected from social networks, it is formalized into a quadruple. ,in As the main body of the event, As the object of the event, For relational types, The time when the event occurred (timestamp); Historical time window Discretize into Each time step, for each time step Aggregate all that happens Construct the corresponding event graph from the four-tuple of time points. , ; Arrange all event diagrams in chronological order to form the event diagram sequence. , as input data for the model.

[0021] S2, Event Real Part Encoding: For each social event quadruple... Regarding the main body of the event Event Object Event Relationship Encode the real part as a representation of the real part.

[0022] ; ; ; in, This indicates One-Hot encoding. , , These are weight matrices for the event subject, relationship type, and event object, respectively. These are bias matrices for the event subject, relationship type, and event object, respectively. These are the real part encoding values ​​for the event subject, relation type, and event object, respectively. This represents the activation function.

[0023] S3, Event Imaginary Part Encoding: Based on the event real part encoding result described in S2. Combined with the time of the event For the main body of the event Event Object Event Relationship Encode the imaginary part as its representation.

[0024] First, regarding historical time windows Discrete time step Perform Fourier encoding: ; ; ; in, It is a historical time series; The first time series after being transformed into the frequency domain One frequency component; Frequency index; for The first in One time step; For time indexing; The length of the historical time window; This represents the total number of sampling points for the Fourier transform. The Fast Fourier Transform (FFT) aims to transform discrete time steps within a historical time window. All data is converted to the frequency domain for modeling, resulting in a time-frequency domain representation. .

[0025] Based on the obtained time-frequency domain characterization The weight matrix obtained in step S2 is then used further. , , Bias matrix Real part encoded value Frequency domain hybrid coding learning: ; ; ; in, This represents the activation function. These are the event subject, relationship type, and event object related to the time of event occurrence. Frequency domain hybrid coding.

[0026] Will Extracted as a time-domain representation: ; ; ; in, This represents the total number of sampling points for the Fourier transform. These are the event subject, relationship type, and event object related to the time of event occurrence. The time-domain representation; Frequency index; For time indexing; , , The first part of the frequency domain hybrid coding of event subject, relation type, and event object respectively One frequency component; The frequency domain hybrid coding result was then used with inverse fast Fourier transform. Convert to imaginary part encoded value: ; ; ; in, Indicates the inverse fast Fourier transform; These are the imaginary part encoding values ​​for the event subject, relationship type, and event object, respectively.

[0027] S4, Fourier plot construction: Constructing a Fourier plot based on the results of real part encoding and imaginary part encoding. The graph construction process mainly involves converting the existing event graph's quadruples into quadruples. Construct a complex triple containing both real and imaginary part encoded values. The complex triplet As shown below: ; in, These are respectively plural representations of the event subject, relationship type, and event object; It is the imaginary unit.

[0028] In this Fourier plot, since a single Fourier plot can represent... This allows for the extraction of all time-frequency information, thereby improving computational efficiency. A single Fourier plot can be represented as a set of multiple complex triples: ; in, The number of complex triples contained in a Fourier plot.

[0029] S5, Frequency Domain Feature Aggregation: For the complex triples obtained in S4 A Fourier space multiplier is constructed to perform interactive operations on the complex features of the event subject, relation type, and event object in the complex domain. Specifically, for each complex triplet... Frequency domain feature aggregation is performed in three ways.

[0030] First, by masking complex triples In and use the remaining The masked part is fitted using complex multiplication: ; in, The item represents pure semantic fusion. The item serves as a penalty for time pattern matching. Item and The term represents the cross-driving effect of semantics and temporal patterns. This step utilizes frequency domain multiplication to perform convolution operations on the Fourier diagram, obtaining the predicted complex representation of the event object. .

[0031] Secondly, by masking complex triples In and use the remaining The masked part is fitted using complex multiplication: ; in, The item represents pure semantic fusion. The item serves as a penalty for time pattern matching. Item and The term represents the cross-driving effect of semantic and temporal patterns. This step utilizes frequency domain multiplication to perform convolution operations on the Fourier diagram, obtaining a complex representation of the predicted relation type. .

[0032] Finally, by masking complex triples In and use the remaining The masked part is fitted using complex multiplication: ; in, The item represents pure semantic fusion. The item serves as a penalty for time pattern matching. Item and The term represents the cross-driving effect of semantics and temporal patterns. This step utilizes frequency domain multiplication to perform convolution operations on the Fourier diagram, obtaining the predicted complex representation of the event subject. .

[0033] S6, Event Decoding and Prediction: Predictive Complex Representation of Event Objects Based on S5 Complex representation of relation types and the predicted complex representation of the event subject Calculate using the real and imaginary part similarity scoring function respectively and between, and Between and and The similarity score between the real and imaginary parts is defined as follows: ; ; ; in, They are respectively and between, and Between and and Similarity score between them; and These represent the real and imaginary parts of a complex number, respectively. The modulus penalty coefficient; This indicates taking the absolute value.

[0034] The similarity scores are mapped to the classification feature space, and the normalized probability distributions of all event subjects, relationship types, and event objects are output. High-confidence combinations are selected and spliced ​​into future prediction quadruplets, and the event occurrence confidence percentage and long-term evolution trend curve are output simultaneously. S7, Construct a real-to-virtual part similarity matching loss based on similarity scores. Its expression is: ; The real-imaginary part similarity matching loss is combined with classification cross-entropy loss, norm regularization term loss, and temporal difference regularization loss to construct a joint loss function. Optimize the network.

[0035] like Figure 2 As shown, the social event prediction method based on Fourier graph learning proposed in this application includes the following specific implementation process: First, acquire the Web social event data stream and parse it into quadruplets. The method constructs a sequence of event graphs arranged chronologically. Then, it enters a complex embedding layer to encode semantic embeddings (real parts) and temporal patterns (imaginary parts). The temporal patterns are processed by Fast Fourier Transform (FFT) into the frequency domain, and the resulting frequency-domain hybrid encoding is then restored to an enhanced temporal pattern using Inverse Fast Fourier Transform (irFFT). Based on this, a Fourier graph containing complex features of nodes and edges is constructed, and a Fourier spatial multiplier is introduced to perform interactive operations on the features in the frequency domain, replacing the traditional spatiotemporal convolution operation in the time domain, significantly reducing computational complexity. Finally, a similarity score is calculated using a real-imaginary part similarity scoring function, and the network is optimized end-to-end based on a joint loss function. This method effectively solves the problems of high computational cost and slow inference speed in existing graph neural networks when processing large-scale data streams, while ensuring accurate capture of long and short-term temporal dependencies.

[0036] like Figure 3 As shown in the figure, to verify the effectiveness of this application, the performance comparison results of the method of this application and the existing mainstream methods in terms of model training efficiency are presented in the training time comparison diagram.

[0037] In terms of model efficiency, such as Figure 3 As shown in the comparison chart of training time (unit: s / epoch), the training times of mainstream social event prediction models and models applying the method of this application are as follows: SeCoGD (313.6 s / epoch), HiSMatch (289.1 s / epoch), RE-GCN (230.2 s / epoch), and the FGL model applying the method of this application (40.9 s / epoch). The area of ​​the squares in the chart corresponds to the training time. The area occupied by the method of this application (FGL) is the smallest, indicating that its training speed is the fastest. This result verifies that the Fourier graph learning paradigm based on frequency domain multiplication can significantly reduce computational complexity while maintaining prediction performance, achieving a speed improvement of about 5 to 7 times. It has significant computational efficiency advantages and is particularly suitable for large-scale Web real-time data stream monitoring and early warning deployment applications.

[0038] The above are merely preferred embodiments of this application and are not intended to limit the scope of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application shall be included within the protection scope of this application.

Claims

1. A social event prediction method based on Fourier graph learning, characterized in that, Includes the following steps: S1, acquire the social event data stream within the historical time window, parse each social event into a quadruple containing the event subject, relationship type, event object, and timestamp, construct an event graph based on the quadruples with the same timestamp, and arrange them in chronological order to obtain an event graph sequence; the timestamp is the time when the event occurred; S2, for each social event quadruple, encodes the real part of the event subject, event object, and event relationship respectively; S3. Based on the real part encoding results and the event occurrence time, perform imaginary part encoding for the event subject, event object, and event relationship respectively; S4. Based on the results of real part encoding and imaginary part encoding, a Fourier diagram is constructed, and the original event diagram quadruples are transformed into complex triples containing real part encoding values ​​and imaginary part encoding values; the complex triples are composed of complex representations of event subject, event object, and event relationship; S5. Construct a Fourier space multiplier to perform interactive operations on the complex features of the event subject, relation type, and event object in the complex domain to obtain the predicted complex representations of the event subject, relation type, and event object. S6 constructs a complex-valued event decoder based on the predicted complex representations of the event subject, relation type, and event object, and outputs the prediction results of a specific event occurring at a future timestamp.

2. The social event prediction method based on Fourier graph learning according to claim 1, characterized in that, In step S2, for each social event quadruple... The real parts of the event subject, event object, and event relationship are encoded respectively, as shown below: in, As the main body of the event, As the object of the event, For relational types, The timestamp is the time when the event occurred. Indicates One-Hot encoding; , , These are weight matrices for the event subject, relationship type, and event object, respectively. These are bias matrices for the event subject, relationship type, and event object, respectively. These are the real part encoding values ​​for the event subject, relation type, and event object, respectively. This represents the activation function.

3. The social event prediction method based on Fourier graph learning according to claim 2, characterized in that, In step S3, based on the real part encoding result and combined with the event occurrence time, the imaginary part encoding is performed on the event subject, event object, and event relationship, as shown below: First, regarding historical time windows Discrete time step Perform Fourier encoding: in, It is a historical time series; The first time series after being transformed into the frequency domain One frequency component; Frequency index; for The first in One time step; For time indexing; The length of the historical time window; This represents the total number of sampling points for the Fourier transform. The Fast Fourier Transform (FFT) aims to transform discrete time steps within a historical time window. All data is converted to the frequency domain for modeling, resulting in a time-frequency domain representation. ; Based on the obtained time-frequency domain characterization Using weight matrix , , Bias matrix Real part encoded value Frequency domain hybrid coding learning: in, Indicates the activation function; These are the event subject, relationship type, and event object related to the time of event occurrence. Frequency domain hybrid coding; Will Extracted as a time-domain representation: in, This represents the total number of sampling points for the Fourier transform. These are the event subject, relationship type, and event object related to the time of event occurrence. The time-domain representation; Frequency index; For time indexing; , , The first part of the frequency domain hybrid coding of event subject, relation type, and event object respectively One frequency component; Frequency domain hybrid coding results are obtained using inverse fast Fourier transform. Convert to imaginary part encoded value: in, Indicates the inverse fast Fourier transform; These are the imaginary part encoding values ​​for the event subject, relationship type, and event object, respectively.

4. The social event prediction method based on Fourier graph learning according to claim 1, characterized in that, In step S4, complex triplet As shown below: in, These are respectively plural representations of the event subject, relationship type, and event object; The imaginary unit; These are the imaginary part encoding values ​​for the event subject, relationship type, and event object, respectively. These are the real part encoding values ​​for the event subject, relation type, and event object, respectively. As the main body of the event, As the object of the event, For relational types, The timestamp is the time when the event occurred. A Fourier plot Represented as a set of multiple complex triples: ;in, The number of complex triples contained in a Fourier plot.

5. The social event prediction method based on Fourier graph learning according to claim 4, characterized in that, In step S5, frequency domain features are aggregated for each complex triplet in three ways: By masking complex triplets In and use the remaining The masked part is fitted using complex multiplication: in, The item represents pure semantic fusion. The item serves as a penalty for time pattern matching. Item and The terms represent the cross-driving effect of semantics and temporal patterns. For the prediction of complex representation of the event object; By masking complex triplets In and use the remaining The masked part is fitted using complex multiplication: in, The item represents pure semantic fusion. The item serves as a penalty for time pattern matching. Item and The terms represent the cross-driving effect of semantics and temporal patterns. For predicting complex representations of relation types; By masking complex triplets In and use the remaining The masked part is fitted using complex multiplication: in, The item represents pure semantic fusion. The item serves as a penalty for time pattern matching. Item and The terms represent the cross-driving effect of semantics and temporal patterns. The predicted complex representation of the event subject.

6. The social event prediction method based on Fourier graph learning according to claim 5, characterized in that, In step S6, the similarity scoring function of the real and imaginary parts is used to calculate respectively. and between, and Between and and The similarity score between the real and imaginary parts is defined as follows: in, They are respectively and between, and Between and and Similarity score between them; and These represent the real and imaginary parts of a complex number, respectively. The modulus penalty coefficient; Indicates taking the absolute value; The similarity scores are mapped to the classification feature space, and the normalized probability distributions of all event subjects, relationship types, and event objects are output. All event subjects, relationship types, and event objects are ranked according to their confidence scores, and the event subjects, relationship types, and event objects with the highest confidence scores are selected as the quadruplet prediction results.

7. The social event prediction method based on Fourier graph learning according to claim 6, characterized in that, Also includes: Based on similarity scores, construct a real-to-virtual part similarity matching loss. Its expression is: The real-imaginary part similarity matching loss is combined with classification cross-entropy loss, norm regularization term loss, and temporal difference regularization loss to construct a joint loss function. Optimize the network.

8. The social event prediction method based on Fourier graph learning according to claim 1, characterized in that, In step S1, in the event graph, the event subject and the event object are used as nodes, and edges are constructed based on the relationship type.

9. A readable storage medium, characterized in that, It stores a computer program, which, when executed, implements a social event prediction method based on Fourier graph learning as described in any one of claims 1 to 8.

10. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a social event prediction method based on Fourier graph learning as described in any one of claims 1 to 8.

Images