A news event prediction method based on heterogeneous evolutionary event clustering

Abstract

The application discloses a news event prediction method based on heterogeneous evolution event clustering, comprising the following steps: generating event representation based on the preliminary updated entity representation and relationship representation in the constructed entity graph, regarding the event as a node, regarding the heterogeneous relationship between events as an edge, and constructing an event graph; obtaining event clusters by fuzzy clustering and constructing an event cluster graph; optimizing the event cluster representation according to the distance and similarity between event clusters on the event cluster graph by using a self-supervised optimization algorithm; capturing the implicit correlation between event clusters by using an implicit relationship encoder, and then updating the representation of the event cluster, the representation of the event, the entity and the relationship representation in sequence after sparsification and information aggregation; and predicting by a convolution-based news event model. The application effectively models the pair correlation, high-order correlation and multi-step time sequence evolution between events, and has important application value in international situation analysis, social governance and intelligent decision support.

Description

Technical Field

[0001] This invention belongs to the field of news event prediction technology, specifically relating to a news event prediction method based on heterogeneous evolutionary event clustering. Background Technology

[0002] News event databases, such as the Global Database of Events, Language, and Tone (GDELT) and the Integrated Crisis Early Warning System (ICEWS), play a crucial role in modern information society. These massive databases accumulate vast amounts of news event data with precise timestamps, providing researchers, policymakers, and business analysts with a wealth of information. These news events are typically represented in the form of a four-tuple (head entity, relation, tail entity, timestamp), where entities encompass various objects such as countries, regions, citizens, and companies, and relations cover various types such as cooperation, access, statement release, condemnation, and refusal.

[0003] The task of news event prediction is based on this rich historical news event data. Its core objective is to accurately predict future relationships between entities by cleverly mapping entities and relationships into low-dimensional, dense real-valued vectors and using vector computation. This cutting-edge research direction has demonstrated enormous application potential and value in many fields, including international situation analysis, social governance, and intelligent decision support. For example, in international situation analysis, in-depth analysis and prediction of news events between countries can help analysts better grasp the direction of the international situation; in social governance, it helps to identify potential social problems in a timely manner, providing a scientific basis for policy making; and in the field of intelligent decision support, news event prediction can provide strong data support for various decisions, improving the scientific nature and accuracy of decision-making.

[0004] Deep learning-based news event prediction methods typically utilize graph neural networks (GNNs) and recurrent neural networks (RNNs) and their variants to deeply model the graph structure and temporal dependencies in news events, thus achieving good performance in news event prediction tasks. Among these methods, GNNs can fully leverage graph structures built on quadruples to capture pairwise correlations between entities, revealing the potential complex relationships between them. In recent years, with the continuous advancement of research, the latest methods have introduced derived structures, such as entity groups, communities, hypergraphs, and evolutionary clusters, to further capture higher-order correlations and their temporal evolution among multiple entities—that is, correlations existing simultaneously among three or more entities. These methods usually rely on representation learning by decomposing news events into entities and relations. While they can reflect the characteristics of events to some extent, they cannot fully capture the heterogeneous correlations between events, i.e., pairwise correlations, higher-order correlations, and their multi-step temporal evolution.

[0005] Therefore, the limitations of existing methods restrict the accuracy and comprehensiveness of news event prediction to some extent, and more effective news event prediction methods need to be explored further. Summary of the Invention

[0006] In view of the above, the purpose of this invention is to provide a news event prediction method based on heterogeneous evolutionary event clustering. To effectively model the heterogeneous correlation between events, an entity graph is constructed based on news events occurring at each timestamp. An event graph is constructed using a tensor-based relation-aware event encoding module, and an event cluster graph is constructed using an event-aware multi-step evolutionary clustering module. An event cluster graph information transmission module is introduced for representation learning to model the pairwise correlation, higher-order correlation, and multi-step temporal evolution between events, thereby achieving more accurate news event prediction. This method has significant application value in international situation analysis, social governance, and intelligent decision support.

[0007] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0008] This invention provides a news event prediction method based on heterogeneous evolution event clustering, comprising the following steps:

[0009] Entity graphs are constructed based on news events, and information is aggregated on the entity graphs using relation-aware graph convolutional neural networks to obtain preliminary updated entity and relation representations.

[0010] In the event graph construction module, event representations are generated by the tensor-based relation-aware event encoding module based on the initially updated entity and relation representations. Then, the event graph is constructed based on the event representations by treating events as nodes and the heterogeneous relationships between events as edges.

[0011] In the multi-step evolutionary clustering module of event perception, the soft membership matrix of event to event cluster is obtained based on the event graph through fuzzy clustering algorithm. Event clusters with different historical timestamps are aligned and fused to construct an event cluster graph, and a time smoothing loss is constructed based on the distance between event clusters.

[0012] A self-supervised optimization algorithm is used to optimize the event cluster representation based on the distance and similarity between event clusters on the event cluster graph, and a self-supervised smoothing loss is constructed based on the similarity between event clusters before and after optimization.

[0013] In the event cluster graph information transmission module, the implicit correlation between event clusters is captured by the implicit relation encoder and then sparsified and aggregated to update the representation of the event clusters. Then, the representation of the events is updated based on the soft membership matrix of the events to the event clusters, and then the entity and relation representations are updated.

[0014] The entity and relation representations updated with different timestamps are input into a time-dependent encoder based on an attention mechanism to obtain a comprehensive entity and relation representation. All possible future relationships between entities are predicted by a convolution-based news event model. A total prediction loss, including prediction loss, temporal smoothing loss, and self-supervised smoothing loss, is constructed to train the entire network. The trained news event model is then used to predict news events.

[0015] Preferably, the entity representation and relation representation based on the initial update are used to generate event representations through a tensor-based relation-aware event encoding module, including:

[0016] First, tensor T is used to encode the element-wise interaction between entity representation and relation representation. The calculation formula is as follows:

[0017]

[0018] Among them, v s,i The interaction v between the head entity s and the relation r. s The element, v o,i The interaction v between tail entity o and relation r. o The element, T ijk This represents the elements of tensor T, where the subscripts i, j, and k indicate that T ijk The index values ​​of h in the three directions of tensor T s,j h is the j-th element of the head entity representation. o,j h is the j-th element of the tail entity representation. r,k Let k be the k-th element of the relation representation;

[0019] Subsequently, the interaction v between the head entity and the relationship is combined. s Interaction between entities and relations v o The event representation ev is obtained, and the calculation formula is as follows:

[0020] ev = W s v s +W o v o

[0021] Among them, W s and W o These are learnable parameters.

[0022] Preferably, the event-based representation constructs an event graph by treating events as nodes and heterogeneous relationships between events as edges, including:

[0023] Based on the event representation ev, the node set of the event graph for timestamp t is represented as: The subscript i is the index representing the event, using To represent an event The set of entities contained therein, i.e. in and These represent the head entity and the tail entity, respectively.

[0024] Two types of heterogeneous edges, shared entity edges and proximity edges, are constructed to describe the relationships between events, as defined below:

[0025] Shared entity edge If two events with indices i and j share at least one entity, then there is a shared entity edge between these two events, represented as:

[0026]

[0027] Adjacent edge If, for two events with indices i and j, one event represents a location within the k-nearest neighbor set of the other event, then there exists a neighboring edge between these two events, represented as:

[0028]

[0029] Where kNN(·) represents the k-nearest neighbor algorithm, Indicates equivalence, V indicates or;

[0030] The two types of heterogeneous edges together form the edge set in the event graph, namely Subsequently, an event graph is constructed based on the node set and edge set. It also utilizes a relation-aware graph convolutional neural network to aggregate information on the event graph.

[0031] Preferably, the step of aligning and fusing event clusters with different historical timestamps to construct an event cluster graph includes:

[0032] By aggregating the entities contained in the events within each event cluster, we obtain the entity distribution matrix CE of the event cluster at time t. t The calculation formula is as follows:

[0033] CE t =(U t ) T M t D -1

[0034]

[0035] Among them, M t Represents the event-entity mapping matrix. M represents t The i-th element, if the event If entity e is included, its value is 1; otherwise, it is 0. D represents a diagonal matrix, where each element represents the total number of times an entity appears in all events, used for normalization. t The soft membership matrix represents the number of events in a cluster of events. Indicates an event The set of entities contained therein, with the superscript T indicating transpose;

[0036] Subsequently, an event cluster-aware Hungarian matching algorithm is used to ensure smooth alignment of event clusters with different timestamps. It first constructs an affinity matrix 'a' to quantify the similarity between event clusters with different timestamps. For example, to align event clusters with timestamps t-1 and t, the calculation formula is as follows:

[0037]

[0038] Where A is the element in the j-th row and k-th column of A. j,k The entity distribution matrix of event cluster j with timestamp t-1 The entity distribution matrix of event cluster k with timestamp t. The similarity between them, cos(·) represents the cosine similarity function;

[0039] After the affinity matrix A is constructed, the optimal alignment is sought using the event cluster-aware Hungarian matching algorithm to maximize the sum of similarities between aligned event clusters. The calculation formula is as follows:

[0040]

[0041] Where π represents the permutation function for achieving smooth alignment of event clusters with different timestamps. τ (j) represents the index of the event cluster with time stamp τ that is aligned with event cluster j of time stamp t, N c Indicates the total number of event clusters;

[0042] After aligning the event clusters based on historical timestamps, the aligned event clusters are then average-pooled to obtain the fused event cluster representation. The calculation formula is as follows:

[0043]

[0044] Where, N t This represents the number of timestamps up to and including timestamp t. Represents the representation of event cluster j with timestamp t. Represents the event cluster with timestamp τ that is aligned with event cluster j of timestamp t;

[0045] Subsequently, an event cluster graph is constructed. in, The set of nodes representing timestamp t represents the aligned and merged event cluster. The edge set representing timestamp t represents the implicit correlation between event clusters.

[0046] Preferably, the formula for calculating the time smoothing loss is as follows:

[0047]

[0048] in, This represents the time smoothing loss.

[0049] Preferably, the step of optimizing the event cluster representation using a self-supervised optimization algorithm based on the distance and similarity between event clusters on the event cluster graph includes:

[0050] A pairwise distance matrix L is constructed based on the Euclidean distance between event clusters, and a distance weight matrix K is constructed. The calculation formula is as follows:

[0051]

[0052]

[0053] Wherein, the element L in the j-th row and k-th column of L j,k This represents the j-th event cluster with timestamp t in the event cluster graph after alignment and merging. and the k-th event cluster The Euclidean distance between them, K j,k Let represent the element in row j and column k of K, where α is a hyperparameter used to control the magnitude of the distance weight adjustment.

[0054] Subsequently, an event cluster similarity matrix F is constructed based on the cosine similarity between event clusters, and a similarity adjustment matrix S is constructed. The calculation formula is as follows:

[0055]

[0056] Among them, F j,k S represents the element in row j and column k of F. j,k Let S represent the element in the j-th row and k-th column, and β is a hyperparameter used to control the magnitude of similarity adjustment;

[0057] Self-supervised optimization of event clusters is achieved by integrating distance-weighted and similarity-adjusted methods, calculated as follows:

[0058]

[0059] in, N represents the self-supervised optimized event cluster. c This represents the total number of event clusters.

[0060] Preferably, the formula for calculating the self-supervised smoothing loss is as follows:

[0061]

[0062] in, This represents the self-supervised smoothing loss.

[0063] Preferably, the convolution-based news event model adopts the ConvTransE model.

[0064] Preferably, the formula for calculating the predicted loss is as follows:

[0065]

[0066] in, N represents the predicted loss. r y represents the number of relations. i,j p represents the label indicating the relationship j between entities in sample i. i,j This represents the predicted probability that entities in sample i will have relationship j.

[0067] Preferably, the total prediction loss is expressed as:

[0068]

[0069] in, This represents the total prediction loss in a training batch. Let F represent the prediction loss for the v-th sample in the batch, and let F represent the number of samples in each batch. This represents the time smoothing loss for the v-th sample. Let represent the self-supervised smoothing loss for the v-th sample, and λ and γ represent the weights.

[0070] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0071] (1) This invention introduces an event graph construction module, uses a tensor-based relation-aware event encoding module to generate event representations, and constructs heterogeneous edges between events to model the pairwise correlation between events.

[0072] (2) This invention introduces an event-aware multi-step evolution clustering module, uses a fuzzy clustering algorithm to obtain the soft membership matrix of events to event clusters, and uses an event cluster-aware unsupervised alignment mechanism to align and fuse event clusters with different historical timestamps to construct an event cluster graph, so as to model the high-order correlation between events and their temporal evolution. Attached Figure Description

[0073] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0074] Figure 1 This is an overall flowchart of the news event prediction method based on heterogeneous evolution event clustering provided in the embodiments of the present invention;

[0075] Figure 2 This is an overall framework diagram of the news event prediction method based on heterogeneous evolution event clustering provided in the embodiments of the present invention. Detailed Implementation

[0076] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0077] The inventive concept of this invention is as follows: Addressing the difficulty of effectively modeling heterogeneous correlations between events in news event prediction tasks using existing technologies—namely, pairwise correlations, higher-order correlations, and their multi-step temporal evolution—this invention provides a news event prediction method based on heterogeneous evolutionary event clustering. First, an entity graph is constructed based on news events occurring at each timestamp, and information aggregation is performed on the entity graph using a relation-aware graph convolutional neural network. Then, in the event graph construction module, a relation-aware event encoding module based on tensors generates event representations and constructs the event graph. Next, in the event-aware multi-step evolutionary clustering module, a fuzzy clustering algorithm is used to obtain the soft membership matrix of events to event clusters, and an event cluster-aware unsupervised alignment mechanism is applied to align and fuse event clusters from different historical timestamps. Then, an event cluster graph is constructed based on the aligned and fused event clusters at each timestamp, and a self-supervised optimization algorithm is used to optimize the node representations in the event cluster graph. Finally, in the event cluster graph information transmission module, implicit relation encoding is used... The encoder captures the implicit correlations between event clusters and sparsifies these correlations using a learnable sparse algorithm. Then, a graph neural network aggregates information on the event cluster graph to update the event cluster representations. The event representations are updated based on the soft membership matrix of events to event clusters, which in turn updates the entity and relation representations. Next, a temporal residual gate combines the updated entity and relation representations with the input from the previous timestamp as the model input for the next timestamp. The updated entity and relation representations from different timestamps are then input into an attention-based temporal dependency encoder to obtain a comprehensive entity and relation representation. Finally, a convolution-based news event model predicts all possible future relationships between entities. This approach effectively models pairwise correlations, higher-order correlations, and their multi-step temporal evolution between events, contributing to further improvements in the accuracy of news event prediction.

[0078] Figure 1 This is an overall flowchart of the news event prediction method based on heterogeneous evolutionary event clustering provided in this embodiment of the invention. Figure 2 This is an overall framework diagram of the news event prediction method based on heterogeneous evolutionary event clustering provided in this embodiment of the invention. For example... Figure 1 and Figure 2 As shown in the embodiment, a news event prediction method based on heterogeneous evolution event clustering is provided, including the following steps:

[0079] Step 1: Preprocess the news event data and extract relevant news events. Divide the processed data into a training dataset using a sliding time window of size T, where each sample contains news events occurring at T timestamps from the news event data.

[0080] News events with timestamps in the news event data are represented as quadruples (s, r, o, t), where s represents the head entity, r represents the relation, o represents the tail entity, and t represents the timestamp. s, o ∈ ε, where ε represents the set of entities, mainly including entities such as countries, regions, citizens, and companies. in The set of representative relationships mainly includes relationships such as cooperation, visits, statements, condemnation, and rejection. in Represents a set of timestamps.

[0081] The timestamps of the quadruplets represent the times when the news events occurred. The quadruplets are arranged in chronological order, and then the processed data is divided using a sliding time window of size T to obtain the training dataset. This represents the set of news events occurring at timestamp t, where each sample contains a set of news events occurring at T timestamps in the news data.

[0082] Step 2: Randomly initialize the learnable entity representation e based on the in-degree of the nodes. i Randomly initialized learnable relational representation r i The subscript i represents the index of the entity representation or relation representation.

[0083] Step 3: Divide the training dataset into batches of a fixed batch size, with a total of N batches.

[0084] Step 4: Select a batch of training samples with index k sequentially from the training dataset, where k∈{1,…,N}. Repeat steps 5-16 for each training sample in this batch.

[0085] Step 5: Construct an entity graph based on the news events that occurred at each timestamp. in, The set of nodes representing timestamp t represents an entity; The edge set representing timestamp t represents the relationship between entities.

[0086] Step 6: For timestamps t = 1, 2, ..., T-1, iteratively execute steps 7-13.

[0087] Step 7: Use a relation-aware graph convolutional neural network to aggregate information on the entity graph.

[0088] In this step, a relation-aware graph convolutional neural network is used to aggregate information on the entity graph. The calculation formula is as follows:

[0089]

[0090] in, This represents the tail entity after information aggregation. Since tail entities in a news event can also serve as head entities in another event, this is also considered as the head entity after information aggregation. and Let W1 and W2 represent the representations of the head entity s, tail entity o, and relation r at time stamp t, respectively, at level l. W1 and W2 are learnable parameters, and d... o The value represents the in-degree of the tail entity, and RReLU(·) represents the activation function. In the following content, for brevity and to avoid ambiguity, the superscript 'l' of variables will be omitted. The relationship between t and timestamps is shown below. It is obtained from the entity representation related to r and the representation of r at the previous timestamp, and the calculation formula is as follows:

[0091]

[0092] Where [;] represents concatenation operation, and pooling[·] represents average pooling operation. This represents the entity associated with time stamp t and r, where This can be understood as i representing the head entity in (i,r,o,t) and i representing the tail entity in (s,r,i,t).

[0093] Step 8: In the event graph construction module, the event representation is generated using the tensor-based relation-aware event encoding module to construct the event graph. in, The set of nodes representing timestamp t, representing the event; The edge set representing timestamp t represents heterogeneous edges between events.

[0094] In this step, the element-wise interaction between the entity representation and the relation representation is first encoded using tensor T, and the calculation formula is as follows:

[0095]

[0096] Among them, v s,i The interaction v between the head entity s and the relation r. s The element, v o,i The interaction v between tail entity o and relation r. o The element, T ijk This represents the elements of tensor T, where the subscripts i, j, and k indicate that T ijk The index values ​​of h in the three directions of tensor T s,j h is the j-th element of the head entity representation. o,j h is the j-th element of the tail entity representation. r,k Let k be the k-th element of the relation representation.

[0097] Subsequently, by combining the interactions between the head entity and the relation and the tail entity and the relation, the representation of the event ev is obtained, and the calculation formula is as follows:

[0098] ev = W s v s +W o v o (5)

[0099] Among them, W s and W o These are learnable parameters.

[0100] After obtaining the event representation, the node set of the event graph for timestamp t is represented as follows: use To represent an event The set of entities contained therein, i.e. in and These represent the head entity and the tail entity, respectively.

[0101] Since there are no predefined edges between events in the news event data, two types of heterogeneous edges are constructed to describe the relationships between events: shared entity edges and proximity edges, which are defined as follows:

[0102] Shared entity edge If two events with indices i and j share at least one entity, then there is a shared entity edge between these two events, defined as follows:

[0103]

[0104] Adjacent edge If two events with indices i and j, where one event represents a location within the k-nearest neighbor set of the other event, then there exists a neighbor edge between these two events, defined as follows:

[0105]

[0106] Where kNN(·) represents the k-nearest neighbor algorithm, ∨ indicates equivalence, and ∨ indicates or.

[0107] The two types of heterogeneous edges together form the edge set in the event graph, namely Subsequently, an event graph was constructed. It also utilizes a relation-aware graph convolutional neural network to aggregate information on the event graph.

[0108] Step 9: In the multi-step evolutionary clustering module for event perception, the soft membership matrix U of event pairs and the event clusters with timestamp t are obtained through fuzzy clustering algorithm.

[0109] In this step, the soft membership matrix U of the event to the event cluster is first obtained by optimizing the following objective function J, and the calculation formula is as follows:

[0110]

[0111] Where, N ev and N c Let represent the number of events and event clusters, respectively; ... The event represented by timestamp t. centroid of event cluster The membership degree, the superscript m represents the fuzzy smoothing hyperparameter, the value of m is set to be greater than 1, <·> represents the dot product operation, and ||·|| represents the magnitude of the vector.

[0112] The overlapping event clusters with timestamp t can then be obtained, calculated using the following formula:

[0113]

[0114]

[0115] in, The representation of event cluster j at timestamp t is obtained by weighting the event representation and the membership degree of the event to the event cluster. Indicates the cluster centroid The centroid of the cluster with index k other than k.

[0116] Step 10: Apply an event cluster-aware unsupervised alignment mechanism to align and merge event clusters with different historical timestamps, and construct an event cluster graph. in, The set of nodes representing timestamp t represents an event cluster; Let t represent the edge set of timestamp t, which represents the implicit correlation between event clusters, and construct a temporal smoothing loss based on the cosine distance between event clusters with different historical timestamps.

[0117] In this step, the entity distribution matrix CE of the event cluster is obtained by aggregating the entities contained in the events within each event cluster. t The calculation formula is as follows:

[0118] CE t =(U t ) T M t D -1 (11)

[0119]

[0120] Among them, M t Represents the event-entity mapping matrix. M represents t The i-th element, if the event If entity e is included, its value is 1; otherwise, it is 0. D represents a diagonal matrix, where each element represents the total number of times an entity appears in all events, used for normalization. t The soft membership matrix represents the number of events in a cluster of events. Indicates an event The set of entities contained therein, with the superscript T indicating transpose.

[0121] Subsequently, an event cluster-aware Hungarian matching algorithm is used to ensure smooth alignment of event clusters with different timestamps. It first constructs an affinity matrix A to quantify the similarity between event clusters with different timestamps. For example, to align event clusters with timestamps t-1 and t, the calculation formula is as follows:

[0122]

[0123] Where A is the element in the j-th row and k-th column of A. j,k The entity distribution matrix of event cluster j with timestamp t-1 The entity distribution matrix of event cluster k with timestamp t. The similarity between them is expressed by cos(·), which represents the cosine similarity function.

[0124] After the affinity matrix A is constructed, the optimal alignment is sought using the event cluster-aware Hungarian matching algorithm, which maximizes the sum of similarities between aligned event clusters. The process of seeking the optimal alignment can be viewed as an optimization problem, and the calculation formula is as follows:

[0125]

[0126] Where π represents the permutation function for achieving smooth alignment of event clusters with different timestamps. τ (j) represents the index of the event cluster with time stamp τ that is aligned with event cluster j of time stamp t, N c This represents the total number of event clusters.

[0127] After aligning the event clusters based on historical timestamps, the aligned event clusters are then average-pooled to obtain the fused event cluster representation. The calculation formula is as follows:

[0128]

[0129] Where, N t This represents the number of timestamps up to and including timestamp t. Represents the representation of event cluster j with timestamp t. Represents the event cluster with timestamp τ that is aligned with event cluster j of timestamp t.

[0130] Subsequently, an event cluster graph is constructed. in, The set of nodes representing timestamp t represents an event cluster; The edge set representing timestamp t represents the implicit correlation between event clusters.

[0131] Furthermore, the time smoothing loss is obtained based on the cosine distance between event clusters of historical timestamps, calculated as follows:

[0132]

[0133] in, This represents the time smoothing loss.

[0134] Step 11: Optimize node representations using a self-supervised optimization algorithm based on Euclidean distance and cosine similarity between nodes in the event cluster graph, and construct self-supervised smoothing constraints based on node similarity before and after optimization.

[0135] In this step, a pairwise distance matrix L is constructed based on the Euclidean distance between event clusters, and a distance weight matrix K is constructed, calculated as follows:

[0136]

[0137]

[0138] Wherein, the element L in the j-th row and k-th column of L j,k This represents the j-th event cluster with timestamp t in the event cluster graph after alignment and merging. and the k-th event cluster The Euclidean distance between them, K j,k This represents the element in the j-th row and k-th column of K. α is a hyperparameter with a value between 0 and 1, used to control the adjustment range of the distance weight.

[0139] Subsequently, an event cluster similarity matrix F is constructed based on the cosine similarity between event clusters, and a similarity adjustment matrix S is constructed. The calculation formula is as follows:

[0140]

[0141]

[0142] Among them, F j,k S represents the element in row j and column k of F. j,k Let S represent the element in the j-th row and k-th column of S. β is a hyperparameter with a value between 0 and 1, used to control the similarity adjustment range.

[0143] Self-supervised optimization of event clusters is achieved by integrating distance-weighted and similarity-adjusted methods, calculated as follows:

[0144]

[0145] in, N represents the self-supervised optimized event cluster. c This represents the total number of event clusters.

[0146] Furthermore, the self-supervised smoothing loss is obtained based on the cosine distance between event clusters before and after self-supervised optimization, and the calculation formula is as follows:

[0147]

[0148] in, This represents the self-supervised smoothing loss.

[0149] Step 12: In the event cluster graph information transmission module, an implicit relation encoder is used to capture the implicit correlations between event clusters, and a learnable sparse algorithm is used to sparsify the implicit correlations. Then, a graph neural network is used to aggregate information on the event cluster graph to update the representation of the event clusters, and the event representation is updated based on the soft membership matrix U of the event to the event cluster. Then update the entity representation Relationship representation

[0150] In this step, the implicit correlation between event clusters is captured using the implicit relation encoder in the event cluster graph information transfer module. The implicit correlation between event cluster pairs consisting of the i-th and j-th event clusters is defined by timestamp t. The calculation formula is as follows:

[0151]

[0152] in, ReLU represents the transformation function, implemented by a multilayer perceptron, and ReLU(·) represents the activation function.

[0153] Furthermore, the strength of implicit correlations is quantified to obtain... The calculation formula is as follows:

[0154]

[0155] Where σ represents the sigmoid function, which ensures that the obtained implicit correlation strength value is between 0 and 1, and Conv(·) represents the convolution operation.

[0156] Subsequently, the implicit correlation is sparsified using a learnable sparse algorithm to obtain... The calculation formula is as follows:

[0157]

[0158] Here, κ is a learnable parameter.

[0159] Then, information is aggregated on the event cluster graph using a graph neural network, and the calculation formula is as follows:

[0160]

[0161] in, This represents the implicit correlation after intensity quantization.

[0162] Then update the event cluster representation and the event representation, calculated as follows:

[0163]

[0164] in, This represents the updated representation of the i-th event cluster with timestamp t. This represents the element in the i-th row and j-th column of the soft membership matrix. Indicates that there are N timestamps in total. c The updated representation of the j-th event cluster in an event cluster. This represents the updated representation of the i-th event at timestamp t.

[0165] After updating the event representation, the entity representation is then updated. Relationship representation Use E ′ and R ′ They represent and The set. In the following content, for the sake of brevity, the subscripts i and j of the variables will be omitted.

[0166] Step 13: Combine the updated entity and relation representations output at timet t with the entity and relation representations input to the model at timet t using a time residual gate to generate the model's input at timet t+1.

[0167] In this step, the updated entity and relation representations output at timetamp t are combined with the representations input to the model at timetamp t using a time residual gate. The calculation formula is as follows:

[0168]

[0169] in, This represents the dot product operation. H is the updated entity or relation representation output by timestamp t. t-1It is the updated entity or relation representation output at timestamp t-1, that is, the representation of the input model at timestamp t, X t This is a time residual gate, which determines the proportion of information to be retained. The calculation formula is as follows:

[0170] X t =σ(W3H t-1 +b) (30)

[0171] Where σ is the sigmoid function, and W3 and b are learnable parameters.

[0172] Step 14: Input the entity and relation representations updated with different timestamps into an attention-based temporal dependency encoder to obtain a comprehensive entity and relation representation. and

[0173] In this step, an attention-based temporal dependency encoder is introduced to model the temporal dependencies between representations. The calculation formula for the position-embedded enhanced representation is as follows:

[0174] z t =[H t ;Φ(t)] (31)

[0175]

[0176] Where [;] denotes a concatenation operation, Φ(t) is a time-position encoder, d is the dimension represented, and {ω i} i＝1,2,…,d It is a learnable parameter, τ t It is a timestamp. Subsequently, given a representation sequence z... 1:T-1 ={z 1 ,z 2 ,…,z T-1 It utilizes a self-attention mechanism to capture the temporal dependencies between location-enhanced representations, outputting a comprehensive entity representation. The relationship between and synthesis

[0177] Step 15: Predict all possible relationships between entities at timestamp T using the convolution-based news event model ConvTransE, and calculate the prediction loss for a single sample using the cross-entropy function.

[0178] Given head entity Tail-end entity The news event prediction task aims to predict the relationship at the T-th timestamp based on news events in a historical window [1:T-1]. In this step, the convolutional news event model ConvTransE is used to predict the probability of different relationships occurring between entities, calculated as follows:

[0179]

[0180] in, It is a probability vector of the relationship. It is a collection of news events in the history window [1:T-1]. and It is a comprehensive entity representation. It is a comprehensive relational representation matrix, where each row corresponds to a comprehensive relational representation. ConvTransE(·) represents a one-dimensional convolutional layer and a fully connected layer.

[0181] Then, based on the prediction results of all relationships, the cross-entropy loss of a single sample is calculated as the prediction loss. The calculation formula is as follows:

[0182]

[0183] Where, N r y represents the number of relations. i,j The label y represents the entity in sample i that has a relationship with j. i,j ∈{0,1}, p i,j The predicted probability of entity j in sample i is calculated by formula (33).

[0184] Step 16: Calculate the prediction loss for each sample in the batch. Time smoothing loss and self-supervised smoothing loss The sum is the prediction loss for all samples in this batch. Adjust the learnable parameters throughout the model.

[0185] In this step, the total prediction loss for all samples in the batch is calculated. The calculation method is as follows:

[0186]

[0187] in, Let F represent the prediction loss for the v-th sample in the batch, and let F represent the number of samples in each batch. This represents the time smoothing loss for the v-th sample. Let represent the self-supervised smoothing loss for the v-th sample, and λ and γ represent the weights. Then, based on the total prediction loss... Adjust the learnable parameters throughout the model.

[0188] Step 17: Repeat steps 4-16 until all batches of the training dataset have been used in model training.

[0189] Step 18: Repeat steps 4-17 until the specified number of iterations is reached.

[0190] Step 19: Given the head entity and the tail entity, input the processed data sample into the trained model to obtain all possible future relationships between the given entities.

[0191] In this step, given the head and tail entities, the news events in the history window are processed and then input into the pre-trained model to obtain the representation of the given head entity. Representation of tail entities The probability of each relationship occurring between given entities at the next timestamp can be calculated using formula (33).

[0192] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A news event prediction method based on heterogeneous evolutionary event clustering, characterized in that, Includes the following steps: Entity graphs are constructed based on news events, and information is aggregated on the entity graphs using relation-aware graph convolutional neural networks to obtain preliminary updated entity and relation representations. In the event graph construction module, event representations are generated by the tensor-based relation-aware event encoding module based on the initially updated entity and relation representations. And based on event representation, an event graph is constructed by treating events as nodes and heterogeneous relationships between events as edges, including: Based on event representation ,Will The node set of the event graph of timestamps is represented as subscript For the index of the event representation, using To represent an event The set of entities contained therein, i.e. ,in and These represent the head entity and the tail entity, respectively. Two types of heterogeneous edges, shared entity edges and proximity edges, are constructed to describe the relationships between events, as defined below: shared entity edge : If two events with indexes and share at least one entity, there is a shared entity edge between the two events, denoted as: ， Adjacent edge If the index is and In two events, one event represents being in the state represented by the other event. In the nearest neighbor set, there is a nearest neighbor edge between these two events, represented as: ， wherein, represents a k-neighbor algorithm, represents equivalence, represents or; Two kinds of isomorphic edges jointly constitute the edge set in the event graph, that is , and then the event graph is constructed based on the node set and the edge set , and the information aggregation is performed on the event graph by using the relationship-aware graph convolutional neural network. In the multi-step evolutionary clustering module of event perception, the soft membership matrix of event to event cluster is obtained based on the event graph through fuzzy clustering algorithm. Event clusters with different historical timestamps are aligned and fused to construct an event cluster graph, and a time smoothing loss is constructed based on the distance between event clusters. A self-supervised optimization algorithm is used to optimize the event cluster representation based on the distance and similarity between event clusters on the event cluster graph, and a self-supervised smoothing loss is constructed based on the similarity between event clusters before and after optimization. In the event cluster graph information transmission module, the implicit correlation between event clusters is captured by the implicit relation encoder and then sparsified and aggregated to update the representation of the event clusters. Then, the representation of the events is updated based on the soft membership matrix of the events to the event clusters, and then the entity and relation representations are updated. The entity and relation representations updated with different timestamps are input into a time-dependent encoder based on an attention mechanism to obtain a comprehensive entity and relation representation. All possible future relationships between entities are predicted by a convolution-based news event model. A total prediction loss, including prediction loss, temporal smoothing loss, and self-supervised smoothing loss, is constructed to train the entire network. The trained news event model is then used to predict news events. 2.The news event prediction method based on heterogeneous evolutionary event clustering according to claim 1, characterized in that, The entity and relation representations based on the initial update are used to generate event representations through a tensor-based relation-aware event encoding module, including: First, using tensors The element-wise interaction between the encoded entity representation and the relational representation is calculated using the following formula: ， ， in, Represents head entity With Relationship Interaction elements, Represents tail entity With Relationship Interaction elements, Tensor The element, subscript , , express In tensor The index values ​​in the three directions, The first entity represented by the head entity One element, The tail entity represents the first One element, The first relation is represented by the first relation. One element; Subsequently, the head entity and the interaction of the relation are combined and the tail entity and the interaction of the relation , obtaining an event representation The calculation formula is as follows: ， wherein, and are learnable parameters. 3.The news event prediction method based on heterogeneous evolutionary event clustering according to claim 1, characterized in that, The process of aligning and fusing event clusters with different historical timestamps to construct an event cluster graph includes: By aggregating the entities contained in the events within each event cluster, we obtain... Entity distribution matrix of event clusters at different times The calculation formula is as follows: ， ， in, Represents the event-entity mapping matrix. express The Middle Each element, if the event Contains entities If the value is 1, then its value is 1; otherwise, it is 0. Represents a diagonal matrix. Each element in the table represents the total number of times an entity appears in all events, used for normalization. The soft membership matrix represents the number of events in a cluster of events. Indicates an event The set of entities included, with superscript Indicates transpose; Subsequently, an event cluster-aware Hungarian matching algorithm is used to ensure smooth alignment of event clusters with different timestamps. This first involves constructing an affinity matrix. Quantify the similarity between event clusters with different timestamps, for example, if for timestamp and The event clusters with timestamps are aligned using the following formula: ， in, The Line 1 Column elements express Timestamp event clusters Entity distribution matrix and Timestamp event clusters Entity distribution matrix Similarity between them Represents the cosine similarity function; affinity matrix After the construction, the Hungarian matching algorithm is used to find the optimal alignment, which maximizes the sum of similarities between aligned event clusters. The formula is as follows: ， in, This represents a permutation function for achieving smooth alignment of event clusters with different timestamps. Indicates and Timestamp event clusters Aligned Index of event clusters with timestamps Indicates the total number of event clusters; After aligning the event clusters of historical timestamps, the aligned event clusters are average-pooled to obtain a fused event cluster representation The calculation formula is as follows: ， wherein, represents a timestamp up to and including the timestamp number, represents a cluster of events of a timestamp, represents a cluster of events aligned with a representation of a cluster of events of a timestamp; Subsequently, the event cluster graph is constructed wherein, represents a set of nodes of timestamps, representing aligned fused event clusters, represents a set of edges of timestamps, representing implicit correlations between event clusters. 4.The news event prediction method based on heterogeneous evolutionary event clustering according to claim 3, characterized in that, The formula for calculating the time smoothing loss is as follows: ， in, This represents the time smoothing loss. 5.The news event prediction method based on heterogeneous evolutionary event clustering according to claim 1, characterized in that, The method of optimizing event cluster representations using a self-supervised optimization algorithm based on the distance and similarity between event clusters in the event cluster graph includes: Constructing a pair-wise distance matrix based on the Euclidean distance between event clusters and constructing a distance weight matrix The calculation formula is as follows: ， ， in, The Middle Line 1 Column elements This indicates that the alignment and merging process constructs an event cluster graph. The timestamp of the event clusters and the event clusters The Euclidean distance between them express The Middle Line 1 Column elements, It is a hyperparameter used to control the adjustment magnitude of the distance weight; Subsequently, the event cluster similarity matrix is constructed based on the cosine similarity between the event clusters and the similarity adjustment matrix is constructed The calculation formula is as follows: ， ， in, express The Middle Line 1 Column elements, express The Middle Line 1 Column elements, It is a hyperparameter used to control the magnitude of similarity adjustment; Self-supervised optimization of event clusters is achieved by integrating distance-weighted and similarity-adjusted methods, calculated as follows: ， wherein, denotes the event cluster after self-supervised optimization, denotes the total number of event clusters. 6.The news event prediction method based on heterogeneous evolutionary event clustering according to claim 5, characterized in that, The formula for calculating the self-supervised smoothing loss is as follows: ， wherein, denotes the self-supervised smooth loss. 7.The news event prediction method based on heterogeneous evolutionary event clustering according to claim 1, characterized in that, The convolution-based news event model uses the ConvTransE model.

8. The news event prediction method based on heterogeneous evolutionary event clustering according to claim 1, characterized in that, The formula for calculating the predicted loss is as follows: ， in, Indicates the predicted loss. Indicates the number of relations. Indicates sample Relationship between entities The tag, Indicates sample Relationship between entities The predicted probability. 9.The news event prediction method based on heterogeneous evolutionary event clustering according to claim 1, wherein, The total predicted loss is expressed as: ， in, This represents the total prediction loss in a training batch. Indicates the first batch The prediction loss for each sample. This indicates the number of samples in each batch. Indicates the first Temporal smoothing loss for each sample Indicates the first Self-supervised smoothing loss for each sample, and Indicates the weight.

Images