Event prediction method based on ordinary differential equations and supernetworks
By using methods based on ordinary differential equations and hypernetworks, the problems of information transmission between users and capture of individual event characteristics were solved, achieving more accurate event prediction.
Patent Information
- Application Number
- CN202310792186.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-30
- Publication Date
- 2025-12-30
- Estimated Expiration
- 2043-06-30
AI Technical Summary
Existing technologies neglect the continuous information transmission between users and the interaction between user embedded channels in event prediction, and lack the ability to capture individual event characteristics.
We employ a method based on ordinary differential equations and hypernetworks, using graph neural networks to model information transmission between users and interactions between user embedding channels. We utilize hypernetworks to capture individual characteristics of events and model the instantaneous changes in user embeddings when the network structure changes through message passing mechanisms.
It improves the accuracy of event prediction by accurately modeling the continuous changes in user embeddings and individual event characteristics, thereby enhancing the accuracy of prediction.
Smart Images
Figure CN116796173B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an event prediction method in the field of link prediction, and more specifically, to an event prediction method based on ordinary differential equations and hypernetworks. Background Technology
[0002] The rapid development of social networks has facilitated information dissemination and interaction among users. Analyzing user behavior and uncovering its evolutionary patterns can provide more personalized services and guide user behavior in a targeted manner, thereby improving the online environment. A network event can be viewed as an interaction between network users. Network events contain rich information, and event prediction tasks utilize historical events to predict potential future events, thereby analyzing user behavior.
[0003] Currently, there are two main methods for event prediction based on dynamic graphs: one is to apply static graph representation learning methods to multiple snapshots of the dynamic graph and then process the output of the snapshots using time series units; the other is to continuously model the dynamic graph, using methods such as RNNs and time series points to model the continuously changing trajectory embedded by the user, thereby completing the event prediction analysis.
[0004] Although some research has been done on event prediction, there are still many aspects worth exploring, mainly in the following areas:
[0005] 1. The continuous modeling of dynamic graphs is insufficient, failing to fully explore the continuous information transmission between users and the interaction between user embedded channels.
[0006] Information exchange between users is continuous. However, most current methods only focus on information exchange between users when an event occurs, while ignoring the interaction between user-embedded channels.
[0007] 2. Lack of capturing individual characteristics of events
[0008] Every event has unique contextual information, meaning it possesses individual characteristics, thus requiring attention to the differences between events. However, existing methods use fixed patterns to handle different events, which lacks focus on the individual characteristics of each event. Summary of the Invention
[0009] The technical problem to be solved by this invention is to overcome the problems of neglecting the continuous information transmission between users and the interaction between user embedded channels, and the lack of capturing individual event characteristics in the existing technology. This invention proposes an event prediction method based on ordinary differential equations and hypernetworks.
[0010] To achieve the above objectives, the present invention employs the following technical solution:
[0011] 1. The event prediction method based on ordinary differential equations and hypernetworks includes the following steps:
[0012] (1) The data processing module processes all event data in the dataset:
[0013] 1) Sort all event data in the dataset according to the chronological order of their occurrence;
[0014] 2) Normalize the time items in all event data in the dataset to the interval [0, 1];
[0015] 3) Based on steps 1) and 2), the dataset is represented as:
[0016]
[0017] Among them, the triple (a i b i , t i ) represents user a i and user b i At time t i The event that occurs at time t i ∈[0, 1]. An EventSet is an ordered set in which all events are sorted in chronological order. Concurrent events that occur at the same time can be sorted in any order within the EventSet.
[0018] 4) Initialize and record the user embedding;
[0019] (2) Acquire and process relevant data for all events occurring at the same time, including the following steps:
[0020] 1) The data loading module obtains the triples of all events that occur at the same time, namely the user sequence number involved in the event and the time when the event occurred;
[0021] 2) Calculate the difference between the time of the event in 1) and the previous different time in the dataset. If the obtained time of the event is 0, then the difference with the previous different time is set to 0.
[0022] 3) Construct the adjacency matrix of the observation graph based on the event triples obtained in 1).
[0023] (3) Modeling information transmission between users and interaction between user embedded channels based on ordinary differential equations using graph neural networks:
[0024] 1) Concatenate the latest updated user embeddings;
[0025] 2) Concatenate the initial user embeddings;
[0026] 3) Model information transmission between users and the interaction between user embedded channels using ordinary differential equations based on graph neural networks;
[0027] (4) Using hypernetworks to capture individual characteristics of events:
[0028] 1) Obtain positive and negative samples of events based on timestamps and event triples;
[0029] 2) Based on the user pair sequence numbers of the positive and negative samples in 1), the user embeddings corresponding to the user sequence numbers are concatenated to form an event representation;
[0030] 3) Use hypernetworks to fine-tune the event representation and capture the individual characteristics of events;
[0031] 4) Calculate the scores of positive and negative samples using the fine-tuned event representation;
[0032] 5) Calculate the event loss using the scores of positive and negative samples from 4).
[0033] (5) Modeling the instantaneous changes in user embedding when the network structure changes using message passing mechanisms:
[0034] 1) Obtain changes in user-embedded data based on message passing mechanisms;
[0035]
[0036] Where Δx(a,t) is the change in user a's embedding over time t, x(a,t) - Let σ be the embedding of user a before the event occurs, σ be the activation function, and W1 and W2 be weight matrices. N(a, t) represents the set of users who had an event with user a at time t. a，t C represents the number of items that interacted with user a at time t. a，t =|N(a,t)|.
[0037] 2) Obtain a new user embedding based on the changes in the user embedding and the user embedding in 1).
[0038] x(a,t) + )=x(a,t - )+Δx(a,t)
[0039] User-embedded updates can be organized into a matrix form, as follows:
[0040]
[0041] Here, represents the structural change of the observed graph over time compared to the previous time step. is a diagonal matrix constructed by summing each row. and are learnable parameter matrices;
[0042] X A (t + ) = X A (t - )+M A (t)⊙ΔX A (t)
[0043] This is a mask matrix consisting of 0s and 1s. The row containing the user whose event occurred has all values of 1, and the remaining rows have all values of 0.
[0044] 3) Calculate the jump loss based on the user-embedded update;
[0045] (6) Process (2)-(4) for each time point in the dataset until all data in the training set is trained. The trained model is then applied to the event detection task.
[0046] The technical solution describes the use of ordinary differential equations based on graph neural networks to model information transmission between users and the interaction between user embedded channels. The steps are as follows:
[0047] 1) Construct an aggregation matrix using the adjacency matrix of the observation graph. The aggregation matrix has the following form:
[0048]
[0049] Where A k To observe the adjacency matrix of the graph, I is the identity matrix, and D is... k This is the degree matrix. α is a hyperparameter that controls the speed at which the central node influences its neighboring nodes; the value of this parameter can be learned from the data.
[0050] 2) Use the latest updated user embedding as input to the hypernetwork to generate the scaling operation amount μ. k and displacement operation amount
[0051]
[0052] 3) Use the scaling operation amount μ from step 2). k and displacement operation amount Scaling and shifting operations are performed on parameter α:
[0053]
[0054] This represents element-wise multiplication, where different values of α can be generated for different time periods by scaling and shifting α.
[0055] 4) The latest updated user embedding after concatenation, the initial user embedding after concatenation, and the aggregation matrix obtained in 1) are used as inputs to the ordinary differential equation, which has the following form:
[0056]
[0057] Among them, L k It is an aggregate matrix, W k This is the weight matrix. The first term on the right-hand side of the equation models the information transmission between users in the network, the second term models the interaction between user embedding channels, and the third term is the initial embedding of the node. Preserving the initial embedding of the node can prevent the model from becoming oversmooth.
[0058] 5) Obtain and record the new user embedding;
[0059] The steps described in the technical solution for obtaining positive and negative event samples based on timestamps and event triples are as follows:
[0060] 1) Concatenate the user embedding with the initial user embedding;
[0061] 2) Locate the event triples involved based on the time, and record the user pair sequence numbers involved in the event triples;
[0062] 3) Based on the user pair sequence number in 2), extract the corresponding user embedding from all user embeddings to form positive samples;
[0063] 4) Randomly generate 8 user pair numbers, extract the corresponding user embeddings from all user embeddings, and form negative samples;
[0064] The technical solution describes the fine-tuning of event representation using hypernetworks, and the steps are as follows:
[0065] 1) Use the initialization parameters of the fully connected layer as the event prior θ e ;
[0066] 2) Using the event representation as input to the fully connected layer, the scaling offset α is obtained. (a，b，t) and displacement offset β (a，b，t) ;
[0067] 3) Use the scaling offset α from step 2). (a，b，t) and displacement offset β (a，b，t) Scaling and shifting operations on prior events:
[0068]
[0069] 4) Using the steps in step 3) Scaling and shifting operations are performed on the event representation;
[0070] Compared with the prior art, the beneficial effects of the present invention are:
[0071] 1. The event prediction method based on ordinary differential equations and hypernetworks described in this invention utilizes ordinary differential equations to model the continuous changes in user embeddings. Simultaneously, by using graph neural networks to define ordinary differential equations, it can model the continuous information transmission between users and the interaction between user embedding channels. The resulting user embeddings ultimately complete the event prediction task. Traditional event prediction methods lack consideration for continuous information transmission, focusing only on information transmission between users when an event occurs. This fails to model the continuous changes in user embeddings, affecting the accuracy of the event prediction task. This invention, however, provides a more accurate and complete model of the dynamic changes in user embeddings, improving the accuracy of event prediction.
[0072] 2. The event prediction method based on ordinary differential equations and hypernetworks described in this invention utilizes hypernetworks to capture individual event features. By extracting regulatory information from event representations, the event representations are fine-tuned to capture the unique context of different events, thus modeling the individual features of events. Traditional methods use fixed patterns to process different events, causing the representations of different events to tend to vary within the same pattern, failing to capture the individual features of events. This invention, however, fully considers the differences between events, thereby improving prediction accuracy.
[0073] In summary, this invention models the continuous changes in user embeddings by addressing the continuous information transmission between users and the interaction between user embedding channels. It also fine-tunes the event representation based on individual event characteristics, thereby using the obtained user embeddings to complete the event prediction task. Attached Figure Description
[0074] Figure 1 This is a schematic diagram showing the functions and connections of the various modules of the computer program that implements the event prediction method based on ordinary differential equations and hypernetworks described in this invention.
[0075] Figure 2 This is a flowchart of the event prediction method based on ordinary differential equations and hypernetworks described in this invention.
[0076] Figure 3 This is a flowchart illustrating the process of modeling information transmission between users and the interaction between user embedded channels using ordinary differential equations based on graph neural networks in the event prediction method based on ordinary differential equations and hypernetworks described in this invention.
[0077] Figure 4 This is a flowchart illustrating the process of obtaining positive and negative event samples based on timestamps and event triples in the event prediction method based on ordinary differential equations and hypernetworks described in this invention.
[0078] Figure 5 This invention relates to an event prediction method based on ordinary differential equations and hypernetworks, which utilizes hypernetworks to fine-tune event representations. Detailed Implementation
[0079] The invention will now be described in detail with reference to the accompanying drawings:
[0080] The technical problem to be solved by the event prediction method based on ordinary differential equations and hypernetworks described in this invention is to overcome the shortcomings of the prior art. It addresses key issues in event prediction tasks such as the continuous changes in user embedding, the continuous information transmission between users, the interaction between user embedding channels, and the individual characteristics of events. It proposes and implements new technologies and methods for event prediction tasks, improves the accuracy of prediction, and provides support for user behavior analysis.
[0081] See Figure 1 To achieve the goal of event prediction based on ordinary differential equations and hypernetworks, a computer program was developed. It comprises five functional modules: a data processing module, a data loading module, a module for modeling information transmission between nodes and interaction between node embedding channels using ordinary differential equations based on graph neural networks, a module for capturing individual event features using hypernetworks, and a module for modeling instantaneous changes in user embeddings when the network structure changes using message passing mechanisms. The functions of each module are as follows:
[0082] 1. Data Processing Module
[0083] The data processing module processes all event data in the dataset. This module mainly consists of three parts: sorting the events in the dataset, normalizing the event times, and initializing the user embedding.
[0084] (1) Sort the events in the dataset.
[0085] Because the event data in the dataset was not initially sorted in chronological order, and training on this dataset requires training at each time point in chronological order, it is necessary to sort all event data in chronological order based on the time component of the event data.
[0086] (2) Normalize the time of the event.
[0087] To facilitate calculation and observation, the time items of the event data need to be normalized so that the values of all time items are within the interval [0,1].
[0088] (3) Initialize user embedding
[0089] 2. Data Loading Module
[0090] The data loading module needs to load all events occurring at a certain point in time and generate relevant data before the start of each training round. The data loading module consists of two parts: obtaining event triples and constructing the adjacency matrix of the observation graph.
[0091] (1) Obtain the event triple
[0092] Because multiple events may occur at a single point in time, all events occurring at that point in time need to be loaded and processed before each training iteration. This is done by retrieving triplets of all events occurring at that point in time from the dataset, containing the user ID involved in the event and the time the event occurred.
[0093] (2) Construct the adjacency matrix of the observation graph
[0094] In an event triple, the user sequence number represents an edge in the graph, so the network structure of the observation graph changes over time. Each time an event triple is loaded, the network structure needs to be updated based on the user sequence numbers in the event triple, and the adjacency matrix of the observation graph needs to be reconstructed.
[0095] 3. Modeling information transmission between users and interaction between user-embedded channels using ordinary differential equations based on graph neural networks.
[0096] The method utilizes ordinary differential equations based on graph neural networks to model information transmission between users and interactions between user embedding channels. It aggregates neighborhood information by constructing an aggregation matrix. For the parameter α in the aggregation matrix that controls the influence of the central node on the velocity of its neighboring nodes, scaling and displacement operations generated by a hypernetwork are used to adjust it, generating different α values for each time period. Finally, the user embeddings are updated using ordinary differential equations. Compared to traditional methods, this method models a more complete dynamic change process of user embeddings, improving the accuracy of event prediction tasks.
[0097] 4. Utilizing hypernetworks to capture individual characteristics of events
[0098] The method described above utilizes hypernetworks to capture individual features of events, generates event priors, and generates scaling and shifting operations to regulate the event priors. The regulated event priors are then used to fine-tune the event representation and calculate the event loss, thereby achieving the goal of processing different events in different modes and modeling individual event features.
[0099] 5. Modeling the instantaneous changes in user embedding when network structure changes using message passing mechanisms.
[0100] The method described above for modeling the instantaneous changes in user embeddings when the network structure changes using a message passing mechanism first calculates the change in user embeddings caused by the change in the neighborhood of the network structure. Then, the updated user embeddings are obtained by adding the change to the latest updated user embeddings, and the jump loss is calculated based on the updated user embeddings.
[0101] See Figure 2 The steps of the event prediction method based on ordinary differential equations and hypernetworks described in this invention are as follows:
[0102] (1) The data processing module processes all event data in the dataset:
[0103] 1) Sort all event data in the dataset according to the chronological order of their occurrence, and normalize the time items in all event data to the interval [0, 1]. The final dataset is represented as follows:
[0104]
[0105] Among them, the triple (a i b i , t i ) represents user a i and user b i At time t i The event that occurs at time t i ∈[0, 1]. An EventSet is an ordered set in which all events are sorted in chronological order. Concurrent events that occur at the same time can be sorted in any order within the EventSet.
[0106] 2) Initialize user embedding and record;
[0107] (2) The acquisition and processing of relevant data for all events occurring at the same time includes the following steps:
[0108] 1) The data loading module obtains the triples of all events that occur at the same time, namely the user sequence number involved in the event and the time when the event occurred;
[0109] 2) Calculate the difference between the time of the event in 1) and the previous different time in the dataset. If the obtained time of the event is 0, then the difference with the previous different time is set to 0.
[0110] 3) Construct the adjacency matrix of the observation graph based on the event triples obtained in 1).
[0111] (3) Modeling information transmission between users and interaction between user embedded channels using ordinary differential equations based on graph neural networks:
[0112] 1) Concatenate the latest updated user embeddings;
[0113] 2) Concatenate the initial user embeddings;
[0114] 3) See Figure 3 This study utilizes ordinary differential equations based on graph neural networks to model information transfer between users and...
[0115] The user embeds the interaction between channels, and the steps are as follows:
[0116] ① Construct an aggregation matrix using the adjacency matrix of the observation graph. The aggregation matrix has the following form:
[0117]
[0118] Where A k To observe the adjacency matrix of the graph, I is the identity matrix, and D is... k This is the degree matrix. α is a hyperparameter that controls the speed at which the central node influences its neighboring nodes; the value of this parameter can be learned from the data.
[0119] ②Use the user embedding obtained in 1) as the input to the hypernetwork to generate the scaling operation quantity μ. k and displacement operation amount
[0120]
[0121] ③ Use the scaling operation amount μ from step ② k and displacement operation amount Scaling and shifting operations are performed on parameter α:
[0122]
[0123] This represents element-wise multiplication, where different values of α can be generated for different time periods by scaling and shifting α.
[0124] ④ Use the user embeddings obtained in 1), the initial user embeddings obtained in 2), and the aggregation matrix obtained in ① as inputs to the ordinary differential equation, which has the following form:
[0125]
[0126] Among them, L k It is an aggregate matrix, W k This is the weight matrix. The first term on the right-hand side of the equation models the information transmission between users in the network, the second term models the interaction between user embedding channels, and the third term is the initial embedding of the node. Preserving the initial embedding of the node can prevent the model from becoming oversmooth.
[0127] ⑤ Obtain new user embeddings and record them;
[0128] (3) The method of using hypernetworks to capture individual characteristics of events includes the following steps:
[0129] 1) See Figure 4 To obtain positive and negative samples of an event based on the timestamp and event triples, the steps are as follows:
[0130] ① Concatenate the user embedding with the initial user embedding;
[0131] ② Locate the event triples involved based on the time, and record the user pair sequence numbers involved in the event triples;
[0132] ③ Based on the user pair sequence number in ②, extract the corresponding user embedding from all user embeddings to form positive samples;
[0133] ④ Randomly generate 8 user pair numbers, extract the corresponding user embeddings from all user embeddings, and form negative samples;
[0134] 1) Based on the user pair sequence numbers of the positive and negative samples in 1), the user embeddings corresponding to the user sequence numbers are concatenated to form an event representation;
[0135] 2) See Figure 5 The event representation is fine-tuned using a hypernetic network, and the steps are as follows:
[0136] ① Use the initialization parameters of the fully connected layer as the event prior θ e ;
[0137] ② Using the event representation in 2) as the input to the fully connected layer, the scaling offset α is obtained. (a，b，t) and displacement offset β (a，b，t) ;
[0138] ③ Use the scaling offset α from step ② (a，b，t) and displacement offset β (a，b，t) Scaling and shifting operations on prior events:
[0139]
[0140] ④ Utilize the steps in ③ Perform scaling and displacement operations on the event representation in 2);
[0141] 3) Calculate the scores of positive and negative samples using the fine-tuned event representation;
[0142] 4) Calculate the event loss L using the scores of positive and negative samples from step 4). event ;
[0143] (4) The method of modeling the instantaneous changes in user embedding when the network structure changes using the message passing mechanism includes the following steps:
[0144] 1) Obtain changes embedded by the user based on the message passing mechanism:
[0145]
[0146] Where Δx(a,t) is the change in user a's embedding over time t, x(a,t) - Let σ be the embedding of user a before the event occurs, σ be the activation function, and W1 and W2 be weight matrices. N(a, t) represents the set of users who had an event with user a at time t. a，t C represents the number of items that interacted with user a at time t. a，t =|N(a,t)|.
[0147] 2) Obtain the new user embedding based on the changes in user embedding and user embedding in 1):
[0148] x(a,t) + )=x(a,t - )+Δx(a,t)
[0149] User-embedded updates can be organized into a matrix form, as follows:
[0150]
[0151] Where, ΔB k This indicates the observation plot at time t. k Compared to the previous time t k -1 structural change, i.e. ΔB k =B k -B k-1 D A It is a pair of ΔB k The diagonal matrix is constructed by summing each row. W1 and W2 are learnable parameter matrices;
[0152] X A (t + ) = X A (t - )+M A (t)⊙ΔX A (t)
[0153] Where M A (t) is a mask matrix consisting of 0s and 1s. The row containing all users whose events occurred has all values of 1, and the remaining rows have all values of 0;
[0154] 3) Calculate the jump loss L based on the user-embedded update. jump;
[0155] (5) Process (2)-(4) for each time point in the dataset until all data in the training set is trained. The trained model is then applied to the event detection task.
Claims
1. An ordinary differential equation and supernetwork-based event prediction method, characterized in that: (1) the ordinary differential equation and supernetwork-based event prediction method comprises the following steps: 1) processing all event data in the data set by a data processing module: ① sorting all event data in the data set according to the chronological order of occurrence; ② normalizing the time items in all event data in the data set to the interval [0, 1]; ③ based on steps ① and ②, representing the data set as: wherein the triple (a i ,b i ,t i ) represents an event that occurred at time t i between user a i and user b i , a i ∈ [0,1]; EventSet is an ordered set, and all events in the set are sorted in chronological order; for concurrent events that occur at the same time, they are sorted in arbitrary order in EventSet; 2) initializing user embedding and recording; (2) the step of obtaining and processing the related data of all events occurring at the same time comprises the following steps: 1) obtaining the triplets of all events occurring at the same time, i.e. the user sequence numbers involved in the events and the time of event occurrence, by a data loading module; 2) calculating the difference between the time of event occurrence in 1) and the previous different time in the data set, and if the obtained time of event occurrence is 0, the difference with the previous different time is set to 0; 3) constructing an observation graph adjacency matrix according to the event triplets obtained in 1); (3) modeling the information transmission between users and the interaction between user embedding channels using a graph neural network-based ordinary differential equation: 1) concatenating the latest updated user embedding; 2) concatenating the initial user embedding; 3) modeling the information transmission between users and the interaction between user embedding channels using a graph neural network-based ordinary differential equation: ① constructing an aggregation matrix using the observation graph adjacency matrix, and the form of the aggregation matrix is: where A k is the observation graph adjacency matrix, I is the identity matrix, D k is the degree matrix; a is a hyperparameter that controls the speed at which a central node influences its neighbor nodes, the value of which can be learned from data; ii) generating a scaling operation amount μ using the user embedding obtained in i) as input to the hypernetwork k and displacement operation amounts iii. using the scaling operation amount μ in step ii. k and the displacement operation amount scaling and displacement operations on the parameter α: where represents element-level multiplication, and different α can be generated for different time periods by scaling and shifting operations on α; ④ taking the user embedding obtained in 1), the initial user embedding obtained in 2), and the aggregation matrix obtained in ① as the input of the ordinary differential equation, and the form of the ordinary differential equation is: where L k is the aggregation matrix, W k is the weight matrix; the first term on the right side of the equation models the information transmission between users in the network, the second term models the interaction between user embedding channels, and the third term is the initial embedding of the node. ⑤ obtaining new user embedding and recording; (3) the step of using a supernetwork to capture individual characteristics of events comprises the following steps: 1) obtaining event positive and negative samples according to the timestamp and event triplets: ① concatenating the user embedding with the initial user embedding; ② finding the event triplets involved according to the time, and recording the user pair sequence numbers involved in the event triplets; ③ extracting the corresponding user embedding from all user embeddings according to the user pair sequence numbers in ② to form positive samples; ④ randomly generating 8 user pair sequence numbers, extracting the corresponding user embedding from all user embeddings to form negative samples; 2) concatenating the user embedding corresponding to the user sequence number to form event representation according to the user pair sequence numbers of the positive and negative samples in 1); 3) fine-tuning the event representation using a supernetwork:
1. Use the initialization parameters of the fully connected layer as the event prior θ e ; ② Take the event representation in 2) as the input of the full connection layer to obtain the scaling offset a (a,b,t) and the displacement offset b (a,b,t) ; iii. Use the scaling offset a and the shift offset b from step ii. (a,b,t) and the shift offset b (a,b,t) Scale and shift the event prior: • using the event representation in step iii) in step iv) scaling and shifting the event representation in 2); 4) calculating the scores of the positive and negative samples using the fine-tuned event representation; 5) Calculate event loss L using the fraction of positive and negative samples in 4) event ; (4) the step of modeling the instantaneous change of user embedding when the network structure changes using a message passing mechanism comprises the following steps: 1) obtaining the change amount of user embedding based on the message passing mechanism: 2) obtaining new user embedding according to the user embedding and the change amount of user embedding in 1): where Δx(a, t) is the change of embedding of user a at time t, x(a, t - ) is the embedding of user a before the event occurs, σ is an activation function, W1, W2 are weight matrices; N(a, t) represents the set of users who have events with user a at time t; C a,t represents the number of items that interact with user a at time t, that is, C a,t = |N(a, t)|; The update of user embedding can be organized in matrix form, which is as follows: x(a, t + ) = x(a, t - ) + Δx(a, t) Where, ΔB k This indicates the observation plot at time t. k Compared to the previous time t k-1 The structural change, i.e., ΔB k =B k -B k-1 ;D A It is a pair of ΔB k The diagonal matrix is constructed by summing each row; W1 and W2 are learnable parameter matrices; X A (t + )=X A (t - )+M A (t)⊙ΔX A (t) where M A (t) is a mask matrix consisting of 0, 1; the values of the rows in which all the users who have occurred are 1, and the values of the remaining rows are 0; 3) Compute the skip loss L according to the user-embedded updates jump ; (5) performing the processing of (2)-(4) on each time point in the data set until all data in the training set is trained, and applying the obtained model to the event detection task.