A rare social event prediction method based on intermediate event learning
By constructing an event transition matrix and a conditional hierarchical graph convolutional neural network, filtering intermediary events, and dynamically adjusting information compression, the dependency and time drift problems of rare social event predictions in existing technologies are solved, achieving high-precision and stable prediction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEFEI UNIV OF TECH
- Filing Date
- 2026-05-25
- Publication Date
- 2026-06-19
Smart Images
Figure CN122241447A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer science, data mining, and computational social science, and in particular to a method for predicting rare social events based on mediation event learning. Background Technology
[0002] Social events are typically recorded in massive web data streams. Among these, rare social events, though occurring very infrequently, often have significant social impact. Predicting such events is crucial for social governance, but it also faces the enormous challenge of being difficult to predict due to their scarcity.
[0003] Existing technologies typically treat rare event prediction as an imbalanced classification problem, primarily employing two types of methods: data-level rebalancing (e.g., oversampling, undersampling) and algorithm-level rebalancing (e.g., cost-sensitive learning, contrastive learning). However, these techniques often suffer from the drawback of ignoring dependencies between events; that is, existing methods usually treat rare events in isolation, neglecting the interactions between different event streams. Furthermore, these methods often struggle to handle intra-event dependencies, as Web data streams exhibit significant temporal non-stationarity, meaning event patterns drift over time. Existing static rebalancing methods rely on fixed correlations and cannot adapt to dynamically changing distributions, resulting in poor generalization ability. Therefore, designing a mechanism that can effectively utilize common events as mediators to alleviate data scarcity while adaptively handling temporal distribution drift is a pressing technical challenge. Summary of the Invention
[0004] To overcome the shortcomings of the prior art, the present invention provides a rare social event prediction method based on mediator event learning, which aims to solve the problems of ignoring the dependence of precursor signals and being unable to adapt to non-stationary distributions in the prior art.
[0005] To achieve the above objectives, the present invention adopts the following technical solution, including: A method for predicting rare social events based on mediation event learning includes the following steps: S1, Data Acquisition and Transformation Matrix Generation: Acquire social event data streams within historical time windows, construct event stream sequences containing time series corresponding to various types of social events; calculate the structure score, co-occurrence score, and semantic score between different types of social events, and generate an event transition matrix representing the probability of mutual transformation between different types of social events; S2, Mediating Event Filtering: Based on the event transition matrix, a Markov chain-based routing algorithm is used to identify event paths leading to the target rare social event and filter mediating events; S3, Hierarchical Graph Structure Construction: A hierarchical spatiotemporal graph is constructed using mediating events and other events besides mediating events as two types of nodes; features are extracted from the time series of social events to generate node embedding vectors, and the graph representation is calculated by combining the event transition matrix. S4. Conditional hierarchical information propagation: Construct a conditional hierarchical graph convolutional neural network to predict the probability distribution of the occurrence of mediating events and generate gating vectors. Use the gating vectors to control the information flow of the graph convolutional neural network, complete the conditional graph convolution based on the probability of the occurrence of mediating events, and obtain the controlled graph representation. S5. Adaptive Regularization and Prediction: Construct an adaptive information regularizer, introduce conditional information bottleneck constraints, and dynamically adjust the balance between information compression and retention; combine prediction constraints to perform loss learning, optimize model parameters, and complete the prediction of rare social events.
[0006] Preferably, the specific method for calculating the structure score in step S1 is as follows: based on the CAMEO event classification body, the CAMEO classification system is modeled as an undirected graph structure, and the shortest path distance between any two different types of social events in the tree hierarchy is calculated using the undirected graph structure. The structure score is obtained by quantifying the shortest path distance, and the path distance is negatively correlated with the structure score.
[0007] Preferably, the specific method for calculating the co-occurrence score in step S1 is as follows: the joint entropy is used to measure the co-occurrence probability of two different types of social events within a preset time window, and the co-occurrence score is obtained. The joint entropy is negatively correlated with the co-occurrence score.
[0008] Preferably, the specific method for calculating the semantic score in step S1 is as follows: a pre-trained language model is used to extract semantic vectors from texts describing different types of social events, and the semantic vector similarity is calculated using cosine similarity to obtain the semantic score.
[0009] Preferably, the specific method for generating the event transition matrix in step S1 is as follows: set the weights and bias terms corresponding to the structure score, co-occurrence score, and semantic score, fuse the three types of scores, calculate the event transition probability between any two different types of social events, and traverse all event pairs to generate the event transition matrix.
[0010] Preferably, the specific process of the Markov chain-based routing algorithm in step S2 is as follows: Performing double and triple multiplication on the event transition matrix yields the second-order and third-order transition matrices, respectively. Based on the second-order and third-order transition matrices, the Soft-Dijkstra algorithm is used to perform a differentiable search on the event transition matrix to select the most probable path to the target rare social event. Social events on the highest probability path are marked as mediating events, and a set of mediating events is constructed.
[0011] Preferably, the specific process of constructing the layered graph structure in step S3 is as follows: We use an embedding layer to extract features from time series of social events, and employ a trend and seasonality decomposition strategy to process the time series of social events. We use a low-order polynomial to fit the trend term of the time series and a Fourier series to fit the seasonal term of the time series, thus transforming the time-domain time series into a node embedding vector. Graph representations are computed using graph convolutional neural networks based on node embedding vectors and event transition matrices.
[0012] Preferably, the specific process of conditional hierarchical information propagation in step S4 is as follows: A lightweight parameter header module is set up to estimate the propagation coefficients of the conditional hierarchical graph convolutional neural network, and the shape parameters of the beta-Bernoulli distribution are predicted based on the graph representation. We use the Kumaraswamy distribution and the Gumbel-Sigmoid reparameterization technique to sample and generate gated vectors from the prediction distribution; By applying the gated vector to the graph representation through the Hadamard product, a controlled graph representation is obtained.
[0013] Preferably, the specific process of adaptive regularization and prediction in step S5 is as follows: An adaptive information regularizer is constructed, a conditional information bottleneck constraint is introduced, and a first loss function is calculated based on graph representation and controlled graph representation. The first loss function is used to constrain the degree of feature information compression. Set up a classifier, input the graph representation into the classifier to obtain the prediction result, combine it with the real event label, use the standard cross-entropy loss to calculate the second loss function, and complete the prediction constraint loss learning; By combining the first and second loss functions with hyperparameter weighting, the total loss function is obtained. The model parameters are then iteratively optimized using the backpropagation algorithm to achieve the prediction of rare social events.
[0014] The present invention also provides an electronic device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned method for predicting rare social events based on mediation event learning.
[0015] The advantages of this invention are: (1) The method of the present invention decomposes the extremely unbalanced prediction task into two relatively balanced stages, and effectively solves the problems of existing methods ignoring the dependence of precursor signals and being unable to adapt to non-stationary distributions by explicitly modeling the evolutionary dependence between events and adaptively adjusting the compression strategy.
[0016] (2) This invention solves the problem of extreme class imbalance by explicitly modeling mediation events, and uses common mediation events to help mine evolutionary patterns and alleviate the problem of scarce samples of rare social events.
[0017] (3) This invention constructs an event transition matrix from structural, co-occurrence, and semantic multi-view perspectives, and models cross-type dependencies and temporal evolution relationships between events, thus overcoming the shortcomings of traditional methods that isolate events and ignore event interactions.
[0018] (4) The present invention sets an adaptive information regularizer and a conditional information bottleneck constraint, which can dynamically adjust the feature compression degree, actively adapt to the time drift characteristics of Web data stream, and has strong time stability.
[0019] (5) The present invention adopts differentiable routing and reparameterized gating mechanism to achieve refined information propagation control, and can still maintain high prediction accuracy and small performance fluctuation in long delay and distributed drift scenarios.
[0020] (6) This invention is highly versatile and can be applied to various social event monitoring scenarios. Attached Figure Description
[0021] Figure 1 This is a schematic diagram of the overall process of the rare social event prediction method of the present invention.
[0022] Figure 2 This is a schematic diagram of the overall process of predicting rare social events according to the present invention.
[0023] Figure 3 This is a schematic diagram comparing the timing generalization performance of the present invention. Detailed Implementation
[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0025] like Figure 1 As shown, a method for predicting rare social events based on mediation event learning includes the following steps: S1, Data Acquisition and Transition Matrix Generation: Acquire social event data streams within a historical time window and construct event stream sequences; the event stream sequences contain time series corresponding to various types of social events. Calculate the structure score, co-occurrence score, and semantic score between different types of social events in the event stream sequences, and generate an event transition matrix representing the probability of mutual transformation between different types of social events. .
[0026] The process of generating the event transition matrix in step S1 includes: First, the hierarchical distance between two different types of social events is calculated based on the CAMEO event classification system, thereby calculating the structure score. .
[0027] The Conflict and Mediation Event Observations (CAMEO) coding system is a structural feature used to construct a priori hierarchical structure for events. CAMEO is widely used as a standardized taxonomy in computational social sciences and political event analysis. Its core feature lies in the systematic organization and coding of various social behaviors using a tree-like hierarchical structure. In this system, each specific event type is assigned a unique numerical code, and the depth of the code reflects the granularity and refinement of the event description. For example, code "02" represents the top-level category "Appeal"; code "17.2" represents the specific subcategory "Impose Sanctions"; and its lower-level node "17.2.3" represents the more granular "Impose Curfew". In the method of this invention, the CAMEO classification system is modeled as an undirected graph structure. This undirected graph structure is used to calculate the shortest path distance between any two event codes in the tree hierarchy, which serves as a quantitative indicator to measure the degree of structural correlation between events (i.e., the aforementioned structural score). The shorter the path distance, the closer the two events are in the classification tree (for example, the distance between "implementing sanctions" and "implementing curfew" is only 1), and the stronger their semantic correlation, thus having a higher potential transition probability in the evolution of events; conversely, events that are far apart (such as "implementing sanctions" and "calling for action") have a lower structural correlation.
[0028] Then, the co-occurrence probability of two different types of social events within a given time window is measured using joint entropy, and the co-occurrence score is calculated. Co-occurrence score This method measures the degree of co-occurrence between two events in time series from a probabilistic and statistical perspective. The lower the joint entropy, the more certain the co-occurrence pattern of the two events is. The calculation formula is as follows: ; Where, assuming These are two different types of social events. This represents the time series corresponding to these two types of social events. It is based on the joint probability of the time series of two types of social events. , This formula is used to refer to the entire state space of two types of social event time series. In practical calculations, this probability is usually estimated by statistically analyzing the frequency of the two time series at the same time step. This formula measures the uncertainty of the two event series as a whole system.
[0029] Finally, a pre-trained language model is used to calculate the semantic similarity between two texts describing different types of social events, and a semantic score is obtained. : ; Where, assuming These are two different types of social events. This is the cosine similarity calculation function. For the event With the event The descriptive text semantic vector is generated by the pre-trained Sentence-BERT model.
[0030] The above score generation events are combined using the following formula. With the event Event transition probability between , representing events To the event The possibility of evolution: ; in, Structural scores Co-occurrence score semantic score The weight, This is the bias term. At this point... The event transition matrix is considered to represent the probability of transformation between different types of social events. The Middle Line number The values corresponding to the columns. Repeating the above process for all pairwise event pairs in the event flow sequence yields the complete event transition matrix. : ; in, The total number of social event types in the event flow sequence is represented by the matrix. Each element in the matrix has a value between 0 and 1, representing the probability of event transformation.
[0031] S2, Mediation event filtering: Based on the event transition matrix described in step S1 This study utilizes a Markov chain-based routing algorithm to identify event paths leading to rare social events, filters intermediary events (precursor intermediary events), and constructs a set of intermediary events. .
[0032] The specific process of the Markov chain-based routing algorithm in step S2 includes: Using the event transition matrix Perform multiplication by itself, and calculate the results of the second and third multiplications respectively; The matrix after quadratic self-multiplication (second-order transition matrix) is expressed as: ,exist The Middle Line number The elements of a column can be written as: ; The matrix after three self-multiplications (third-order transition matrix) is expressed as: ,exist The Middle Line number The elements of a column can be written as: ; The second-order transition matrix and the third-order transition matrix represent the transition probabilities of two-step events and three-step events, respectively.
[0033] Regarding the obtained , and The Soft-Dijkstra algorithm is used to perform a differentiable search on the event transition matrix to select the path with the highest probability to the target rare social event.
[0034] Select the event types on the highest probability path, label these events as mediating events, and construct a set of mediating events leading to the target rare social event. .
[0035] S3, Hierarchical graph structure construction: based on the set of mediation events selected in step S2. and except sets Other event sets consisting of all other event types besides A hierarchical spatiotemporal graph containing two types of nodes is constructed, where nodes represent social event types (including sets of mediated events). Mediation event types and other event sets Other event types in (the context).
[0036] The specific process of constructing the layered graph structure in step S3 includes: First, feature extraction is performed on the time series data of various social events using an embedding layer. Considering that social event data typically contains long-term trends and periodic fluctuations, this invention employs a trend and seasonality decomposition strategy, using low-order polynomials to fit the trend term and Fourier series to fit the seasonality term. In this way, the time series data in the time domain are transformed into a high-dimensional spatiotemporal node representation. ; Where Decompose is the decomposition strategy function for long-term trends and cyclical fluctuations. As a time series of social events, The node embedding vector.
[0037] Based on the obtained high-dimensional node embedding vector Compared with the event transition matrix obtained in step S1 Furthermore, the convolution formula for graph convolutional neural networks is obtained: ; in, The formula for the convolution kernel of a graph convolutional neural network. A graphical representation of all social events.
[0038] S4, Conditional Hierarchical Information Propagation: Construct a conditional hierarchical graph convolutional neural network. First, estimate the probability distribution of the occurrence of mediating events and generate a gating vector. Then, apply the gating vector to the feature propagation process to control the information flow of the graph convolutional neural network and realize conditional graph convolution based on mediation probability.
[0039] The specific steps of conditional hierarchical information propagation in step S4 include: Lightweight parameter header modules are used to predict the propagation coefficients of conditional hierarchical graph convolutional neural networks, based on graph representation. Predicting the shape parameters of the beta-Bernoulli distribution ; The gate vector is obtained by sampling from the prediction distribution using the Kumaraswamy distribution and the Gumbel-Sigmoid reparameterization technique. The Hadamard product is used to apply the gated vector to the graph feature representation, as shown in the formula: ; in, It is a graph representation controlled by a gated vector (controlled graph representation). This is for the Hadamard product operation.
[0040] S5, Adaptive Regularization and Prediction: Constructing an adaptive information regularizer, introducing conditional information bottleneck constraints, and based on graph representation. and graph representation controlled by gated vectors Dynamically adjust the balance between information compression and retention.
[0041] The construction of the adaptive information regularizer includes the following loss function. : ; in, For graph representation and graph representation controlled by gated vectors The bottleneck constraint calculation is based on the conditional information.
[0042] Additionally, classifiers are used for graph representations. To make a prediction, the predicted results are compared with the actual results using a loss function that imposes prediction constraints. : ; in, For standard cross-entropy loss, For classifiers, This represents the actual results in the dataset.
[0043] The final optimized total loss function is: ; in, This is a hyperparameter used to control the combination coefficients of the loss function.
[0044] like Figure 2 As shown, this invention proposes a rare social event prediction method based on mediation event learning. Its specific implementation process includes: First, acquiring a Web social event data stream, combining the entire set of events with the target rare social event set to generate an event transition matrix containing structural, co-occurrence, and semantic multi-view information, and using a Markov chain-based routing algorithm to calculate high-probability paths to the target rare social event, thereby filtering out key mediation events; then, entering the hierarchical graph construction stage, extracting time-frequency domain features and generating adaptive edges for nodes; based on this, introducing a conditional hierarchical graph convolutional neural network, using the Beta-Bernoulli distribution to estimate the probability distribution of mediation events, and using the Kumaraswamy distribution and Gumbel-Sigmoid technique for differentiable sampling to generate a gate vector, using this gate vector to perform gated normalization processing on the underlying features, realizing conditional information propagation based on mediation probability; finally, constructing an adaptive information regularizer, introducing prediction constraints and conditional information bottleneck constraints, and dynamically adjusting the information bottleneck to adapt to temporal changes. This method decomposes the extremely imbalanced prediction task into two relatively balanced stages, and effectively solves the problems of existing methods ignoring the dependence of precursor signals and being unable to adapt to non-stationary distributions by explicitly modeling the evolutionary dependencies between events and adaptively adjusting the compression strategy.
[0045] like Figure 3 As shown, to verify the effectiveness of this invention, a comparison diagram of temporal generalization performance illustrates the performance robustness comparison between the method of this invention and existing mainstream optimization strategies in the face of time distribution drift; regarding model robustness and generalization ability, as... Figure 3 As shown, box plots illustrate the performance distribution of different methods in both standard and long-latency testing environments. Analysis reveals that when facing the challenge of data distribution drift caused by long latency, the performance of each model is as follows: In long-latency tests, the standard Binary Cross-Entropy (BCE) and Static Information Bottleneck (IB) strategies exhibit a significant leftward shift in box position (a substantial decrease in median) and a wider box width (increased variance and more discrete points), indicating a severe decline in predictive performance and extreme instability. In contrast, the adaptive information bottleneck strategy applying the method of this invention maintains a high median and does not show significant expansion in box width. This result verifies that the adaptive regularization paradigm based on conditional information bottleneck proposed in this invention can adaptively adjust the feature compression strategy according to the dynamically evolving social context. While maintaining high predictive accuracy, it significantly enhances the model's ability to resist time non-stationarity, demonstrating significant generalization advantages and making it particularly suitable for long-term, dynamically changing Web social event monitoring and early warning deployment applications.
[0046] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for predicting rare social events based on mediation event learning, characterized in that, Includes the following steps: S1, Data Acquisition and Transformation Matrix Generation: Acquire social event data streams within historical time windows, construct event stream sequences containing time series corresponding to various types of social events; calculate the structure score, co-occurrence score, and semantic score between different types of social events, and generate an event transition matrix representing the probability of mutual transformation between different types of social events; S2, Mediating Event Filtering: Based on the event transition matrix, a Markov chain-based routing algorithm is used to identify event paths leading to the target rare social event and filter mediating events; S3, Hierarchical Graph Structure Construction: Using mediating events and other events besides mediating events as two types of nodes, a hierarchical spatiotemporal graph is constructed. Feature extraction is performed on the time series of social events to generate node embedding vectors, and graph representations are calculated by combining them with the event transition matrix. S4. Conditional hierarchical information propagation: Construct a conditional hierarchical graph convolutional neural network to predict the probability distribution of the occurrence of mediating events and generate gating vectors. Use the gating vectors to control the information flow of the graph convolutional neural network, complete the conditional graph convolution based on the probability of the occurrence of mediating events, and obtain the controlled graph representation. S5. Adaptive Regularization and Prediction: Construct an adaptive information regularizer, introduce conditional information bottleneck constraints, and dynamically adjust the balance between information compression and retention; combine prediction constraints to perform loss learning, optimize model parameters, and complete the prediction of rare social events.
2. The method for predicting rare social events based on mediation event learning according to claim 1, characterized in that, The specific method for calculating the structure score in step S1 is as follows: Based on the CAMEO event classification system, the CAMEO classification system is modeled as an undirected graph structure. The shortest path distance between any two different types of social events in the tree hierarchy is calculated using this undirected graph structure. The structure score is obtained by quantifying the shortest path distance. The path distance is negatively correlated with the structure score.
3. The method for predicting rare social events based on mediation event learning according to claim 1, characterized in that, The specific method for calculating the co-occurrence score in step S1 is as follows: joint entropy is used to measure the probability of co-occurrence of two different types of social events within a preset time window, and the co-occurrence score is obtained. The joint entropy is negatively correlated with the co-occurrence score.
4. The method for predicting rare social events based on mediation event learning according to claim 1, characterized in that, The specific method for calculating the semantic score in step S1 is as follows: a pre-trained language model is used to extract semantic vectors from texts describing different types of social events, and the semantic vector similarity is calculated using cosine similarity to obtain the semantic score.
5. The method for predicting rare social events based on mediation event learning according to claim 1, characterized in that, The specific method for generating the event transition matrix in step S1 is as follows: set the weights and bias terms corresponding to the structure score, co-occurrence score, and semantic score, fuse the three types of scores, calculate the event transition probability between any two different types of social events, and traverse all event pairs to generate the event transition matrix.
6. The method for predicting rare social events based on mediation event learning according to claim 1, characterized in that, The specific process of the Markov chain-based routing algorithm in step S2 is as follows: Performing double and triple multiplication on the event transition matrix yields the second-order and third-order transition matrices, respectively. Based on the second-order and third-order transition matrices, the Soft-Dijkstra algorithm is used to perform a differentiable search on the event transition matrix to select the most probable path to the target rare social event. Social events on the highest probability path are marked as mediating events, and a set of mediating events is constructed.
7. The method for predicting rare social events based on mediation event learning according to claim 1, characterized in that, The specific process of constructing the layered graph structure in step S3 is as follows: We use an embedding layer to extract features from time series of social events, and employ a trend and seasonality decomposition strategy to process the time series of social events. We use a low-order polynomial to fit the trend term of the time series and a Fourier series to fit the seasonal term of the time series, thus transforming the time-domain time series into a node embedding vector. Graph representations are computed using graph convolutional neural networks based on node embedding vectors and event transition matrices.
8. The method for predicting rare social events based on mediation event learning according to claim 1, characterized in that, The specific process of conditional hierarchical information propagation in step S4 is as follows: A lightweight parameter header module is set up to estimate the propagation coefficients of the conditional hierarchical graph convolutional neural network, and the shape parameters of the beta-Bernoulli distribution are predicted based on the graph representation. We use the Kumaraswamy distribution and the Gumbel-Sigmoid reparameterization technique to sample and generate gated vectors from the prediction distribution; By applying the gated vector to the graph representation through the Hadamard product, a controlled graph representation is obtained.
9. The method for predicting rare social events based on mediation event learning according to claim 1, characterized in that, The specific process of adaptive regularization and prediction in step S5 is as follows: An adaptive information regularizer is constructed, a conditional information bottleneck constraint is introduced, and a first loss function is calculated based on graph representation and controlled graph representation. The first loss function is used to constrain the degree of feature information compression. Set up a classifier, input the graph representation into the classifier to obtain the prediction result, combine it with the real event label, use the standard cross-entropy loss to calculate the second loss function, and complete the prediction constraint loss learning; By combining the first and second loss functions with hyperparameter weighting, the total loss function is obtained. The model parameters are then iteratively optimized using the backpropagation algorithm to achieve the prediction of rare social events.
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 rare social event prediction method based on mediation event learning as described in any one of claims 1 to 9.