Event type-aware timing interaction intensity adaptive decay method
By assigning independent decay rates and incentive weights to different types of interactive events, this method solves the problem of insufficient differentiation of event type influence in existing technologies, realizes refined modeling of user behavior, and improves the prediction accuracy and adaptability of recommendation systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHAOXING UNIVERSITY
- Filing Date
- 2026-04-24
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies cannot effectively distinguish the duration and intensity of the impact of different types of events, resulting in an inability to accurately depict the temporal dynamics of user behavior in information retrieval and recommendation systems, which affects the prediction accuracy of models and the effectiveness of personalized recommendations.
An event type-aware, time-series interaction intensity adaptive decay method is adopted. By assigning independent and learnable decay rate parameters and inter-type incentive weights to different types of interaction events, a conditional intensity function is constructed to quantify the historical interaction impact of different event types.
It enables refined modeling of different event types, improves the prediction accuracy of the model in personalized recommendation and user lifetime value prediction, and enhances the model's generalization ability and practical application value.
Smart Images
Figure CN122432602A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of time series data modeling technology, and in particular to an event type-aware adaptive decay method for time series interaction intensity. Background Technology
[0002] In applications such as information retrieval, recommendation systems, and social network analysis, user interactions with items (e.g., products, documents, content) exhibit significant temporal dynamics. User interests drift over time, item popularity fluctuates cyclically, and short-term continuous interactions often reflect a user's immediate intent. For example, in e-commerce platforms, user clicks, adding items to cart, and purchases constitute a continuous sequence of events with strong temporal dependencies between adjacent events; in conversational search, a user's current query is often closely related to their previous query sequence. Accurately capturing these temporal dynamics is crucial for improving the precision and personalization of search results.
[0003] In existing technologies, information retrieval systems typically model user-item interactions as static graph structures, using methods such as graph neural networks to learn node embedding representations. These static modeling methods ignore the temporal information of interaction events, treating interactions occurring at different times as equally important, failing to distinguish between long-term user preferences and short-term intentions, and struggling to capture the changing patterns of interaction density. To address this issue, some researchers have proposed dynamic graph representation learning methods, discretizing time into fixed-length snapshot windows and independently analyzing graph structure changes within each window. However, the discrete snapshot method has inherent drawbacks: a fixed window size makes it difficult to simultaneously capture both fine-grained interaction temporal details and macroscopic trends; interaction events at window boundaries may be incorrectly assigned to adjacent windows, distorting temporal information; more importantly, the snapshot method treats all interactions within a window as occurring simultaneously, losing precise time interval information between events and failing to model the dynamic evolution of interaction intensity.
[0004] Hawkes dot processes have attracted attention for modeling temporal events due to their ability to describe the "self-excitation" property of event sequences. In a Hawkes process, the occurrence of a historical event increases the probability of a future event, and this self-excitation effect decays and propagates through a kernel function. Researchers have applied Hawkes processes to dynamic network representation learning, generating time-aware node embeddings by modeling the excitation effect of historical interactions between nodes on current interactions. However, existing Hawkes process-based dynamic graph models share a common drawback: they typically employ a globally uniform decay kernel function, such as setting the same exponential decay rate for all types of historical events. This approach ignores the differences in the duration of influence of different semantic events.
[0005] In real-world information retrieval and recommendation systems, the intensity and duration of the impact of different types of events on subsequent behaviors vary significantly. For example, in e-commerce, the intensity of user preference conveyed by a single "purchase" and its influence period are typically much longer than that of a single "browse / click." In social media, the decay rates of the impact of "comments" and "likes" are also drastically different. A uniform decay assumption forces models to treat these fundamentally different temporal dynamics in the same way, resulting in models being unable to finely characterize long-term dependencies and short-term interests in complex behavioral sequences, thus limiting their performance ceiling in terms of accurate prediction and deep understanding of user intent. Furthermore, existing methods typically assume that all event types have the same incentive intensity for the target event, or simply use fixed parameters for weighting, failing to learn the complex incentive relationships between different types of events. For example, the incentive intensity of "purchase" for "purchase" should be much higher than that of "click" for "purchase." Therefore, how to achieve refined modeling of the influence patterns of heterogeneous events has become a key issue in improving the accuracy of temporal interaction prediction. Summary of the Invention
[0006] To address the technical problem in existing technologies that cannot assign independent attenuation parameters to different event types in order to finely model the influence patterns of heterogeneous events, this invention provides an event type-aware adaptive attenuation method for temporal interaction intensity.
[0007] The technical solution provided by this invention is as follows:
[0008] The event type-aware, time-series interaction intensity adaptive attenuation method provided by this invention includes:
[0009] S1. Obtain a dynamic interactive event sequence, wherein each event in the dynamic interactive event sequence includes a first node, a second node, an event type, and a timestamp;
[0010] S2. Generate node embedding vectors that evolve over time for the nodes;
[0011] S3. For the target event in the dynamic interaction event sequence, construct a conditional strength function to predict the probability of the target event occurring. The conditional strength function includes a baseline strength term and a direct historical interaction stimulus term.
[0012] The direct historical interaction stimulus term quantifies the impact of historical interaction events of different event types on future events as they decay over time by assigning an independent and learnable decay rate parameter to each event type in the dynamic interaction event sequence.
[0013] Furthermore, the event type described in S1 is a mutually exclusive interactive behavior category with specific semantics, including at least one of click, purchase, favorite, or comment.
[0014] Furthermore, the conditional strength function described in S3 is expressed as:
[0015] S301, Define the target event Where u is the first node and v is the second node. The event type of the target event is t, and the timestamp of the target event is t.
[0016] S302, The conditional strength function is constructed as follows:
[0017]
[0018] in, Let e be the conditional intensity of the target event e at time t. For the baseline intensity term, This refers to the direct historical interaction incentive term.
[0019] Furthermore, the direct historical interaction stimulus term described in S302 is calculated through the following steps:
[0020] S3021. Obtain the set of historical interaction events between the first node u and the second node v. Each historical interaction event in the set of historical interaction events is represented as: This indicates that the first node u and the second node v were in historical time. The event type is Interaction;
[0021] S3022, Traverse the set of historical interaction events that satisfy... All historical interaction events are summed according to the following formula:
[0022]
[0023] in, For the direct historical interaction stimulus term, This is the set of historical interaction events between the first node u and the second node v. "A historical interaction event is one of the historical interaction events in the set of historical interaction events, indicating that the first node u and the second node v had an event of type '' at historical time t'." Interaction, These are learnable inter-type incentive weights used to quantify event types. Historical interactive events for event types The incentive effect of the target event, The learnable decay rate parameter is used to control the event type. The duration of the influence of historical interaction events on the target event, where t is the timestamp of the target event and t' is the timestamp of the historical interaction events.
[0024] Furthermore, the baseline intensity term described in S302 is calculated through the following steps:
[0025] S3023, Obtain the node embedding vectors of the first node u and the second node v at the target timestamp t. and The dimension of the node embedding vector is d;
[0026] S3024. Obtain the event type of the target event. Associated learnable type-specific transformation matrix ;
[0027] S3025. Calculate the baseline strength term according to the following formula:
[0028]
[0029] in, For the baseline intensity term, and These are the node embedding vectors of the first node u and the second node v at the target timestamp t, respectively. For the event type of the target event The associated learnable type-specific transformation matrix, where d is the dimension of the node embedding vector. It is an exponential function. This represents the transpose of a vector.
[0030] Furthermore, S3 also includes a step of optimizing the parameters in the conditional intensity function:
[0031] S311. With the goal of maximizing the log-likelihood function of the dynamic interaction event sequence, jointly optimize the parameters in the node embedding vector and the conditional strength function;
[0032] S312. Output the optimized node embedding vector.
[0033] Furthermore, S601 specifically includes the following sub-steps:
[0034] S6011. For each positive event in the dynamic interactive event sequence, sample multiple negative events from the node set;
[0035] S6012. Construct a loss function based on the conditional strength function value of the positive event and the conditional strength function value of the negative event;
[0036] S6013. Update the node embedding vector and the parameters in the conditional strength function by minimizing the loss function.
[0037] Furthermore, the loss function described in S6012 is an approximate log-likelihood loss function based on negative sampling, calculated by the following formula:
[0038]
[0039] in, This is an approximate log-likelihood loss function based on negative sampling. The set of events in the dynamic interactive event sequence. For positive events, Let e be the conditional strength function value of the positive event e. Here, log is the sigmoid function, log is the natural logarithm, and K is the number of negative samples. For the kth sampled negative node, negative event The conditional strength function value.
[0040] Furthermore, generating the node embedding vector that evolves over time for the node in S2 also includes the following steps:
[0041] S201. When a new node appears, its node embedding vector is initialized according to the available feature information of the new node, wherein the available feature information includes user demographic features or project content features.
[0042] Furthermore, generating the node embedding vector that evolves over time for the node in S2 also includes the following steps:
[0043] S211. In response to receiving an interaction event, update the embedding vectors of the two interaction nodes based on the pre-interaction embedding vectors of the interaction nodes of the interaction event, the event type of the interaction event, and the time interval between the interaction event and the most recent interaction between each of the two interaction nodes.
[0044] The beneficial effects of the technical solution provided by this invention include at least the following:
[0045] (1) In this invention, an event type-aware direct historical interaction stimulus term is constructed by assigning an independent and learnable decay rate parameter to each event type. Specifically, in the conditional strength function of the Hawkes process, different types of interactive events (such as clicks, purchases, and favorites) each have their own exclusive decay rate parameter, enabling the model to automatically learn the "memory half-life" of different behaviors from the data. For example, the impact of a "purchase" behavior lasts for a long time and decays slowly, while the impact of a "click" behavior lasts for a short time and decays rapidly. This technique solves the problem that the uniform decay assumption in the prior art cannot distinguish the heterogeneity of event types, and realizes refined modeling of the duration of the impact of different interactive behaviors such as clicks, purchases, and favorites, significantly improving the modeling accuracy of complex behavior sequences.
[0046] (2) In this invention, the incentive relationship between different event types is quantified by setting a learnable inter-type incentive weight matrix. Specifically, inter-type incentive weights are introduced into the direct historical interaction incentive item to describe the incentive intensity of historical interaction event types on the target event type. This weight matrix can automatically learn the incentive patterns between different types of events. For example, the incentive intensity of "buy" on "buy" is higher than that of "click" on "buy", and the incentive intensity of "comment" on "comment" is higher than that of "like" on "comment". This technical means solves the problem of ignoring the incentive differences between event types in the prior art, enabling the model to more realistically reflect the logic of user behavior and improve the prediction accuracy in scenarios such as personalized recommendation, user lifetime value prediction, and precise marketing timing selection.
[0047] (3) In this invention, by incorporating the decay rate parameter and inter-type incentive weights into an end-to-end optimization framework, automatic learning of the influence patterns of event types is achieved. Specifically, all parameters (including node embedding vectors, type-specific transformation matrices, inter-type incentive weights, and decay rate parameters) are jointly optimized with the goal of maximizing the log-likelihood of the dynamic interaction event sequence. The model does not require manual pre-setting of decay parameters or incentive weights, but directly learns the optimal decay rate and incentive pattern for each event type from the data, and can adaptively adjust the degree of response to different interaction densities. This technique solves the problem that parameters in existing methods are difficult to adapt to changes in data distribution, enabling the model to flexibly cope with differences in interaction patterns under different scenarios, and improving the model's generalization ability and practical application value. Attached Figure Description
[0048] Figure 1 A flowchart illustrating the event type-aware, time-series interaction intensity adaptive attenuation method provided in this embodiment of the invention;
[0049] Figure 2This is a schematic diagram illustrating the principle of the event type-aware, time-series interaction intensity adaptive attenuation method provided in an embodiment of the present invention. Detailed Implementation
[0050] Reference manual attached Figure 1 The diagram illustrates a flowchart of the event type-aware, time-series interaction intensity adaptive attenuation method provided in an embodiment of the present invention.
[0051] This invention provides an event type-aware adaptive attenuation method for temporal interaction intensity, the processing flow of which may include the following steps:
[0052] S1. Obtain a dynamic interactive event sequence, wherein each event in the dynamic interactive event sequence includes a first node, a second node, an event type, and a timestamp.
[0053] In this embodiment, a time-series network is considered. ,in For a set of nodes, This is a collection of interactive events. Each event... Associate a timestamp and an event type Event type collection Used to capture the semantic heterogeneity of interactions, such as clicks, purchases, favorites, and comments in information retrieval or recommendation system scenarios. An event can be represented as... This indicates that node i and node j cross the data at time t via type . An interaction occurs.
[0054] S2 generates node embedding vectors that evolve over time for each node.
[0055] The goal of dynamic network representation learning is to learn a mapping, given an initial feature matrix and a sequence of events, to generate a time-dependent embedding vector for any node v at any time t. These embedding vectors should simultaneously capture the temporal dynamics of a node's structural role and its interaction history.
[0056] In one implementation, when a new interaction event Upon arrival, the embedding updates for nodes u and v are triggered. It should be noted that the specific evolution mechanism of the node embedding vector (such as time message construction, neighbor aggregation, gated recurrent unit update, etc.) can be implemented using dynamic graph representation learning methods known in the art, and this invention does not impose specific limitations on it.
[0057] S3. For the target event in the dynamic interaction event sequence, construct a conditional strength function to predict the probability of the target event occurring. The conditional strength function includes a baseline strength term and a direct historical interaction stimulus term.
[0058] In this embodiment, the Hawkes point process is used to model the temporal dynamics of interactions and capture the "self-exciting" characteristics of event sequences. The core of the Hawkes process is the conditional intensity function, which is used to predict the instantaneous probability of an event occurring at time t. Among them, the direct historical interaction excitation term quantifies the impact of historical interaction events of different event types on future events after decay over time by assigning independent and learnable decay rate parameters to each event type in the dynamic interaction event sequence.
[0059] S301. Define the target event , where u is the first node, v is the second node, is the event type of the target event, and t is the timestamp of the target event.
[0060] S302. Construct the conditional intensity function as:
[0061]
[0062] Among them, is the conditional intensity of the target event e at time t, is the baseline intensity term, is the direct historical interaction excitation term.
[0063] S3021. Obtain the set of historical interaction events between the first node u and the second node v , and each historical interaction event in the set of historical interaction events is represented as , indicating that the first node u and the second node v had an interaction of event type at historical time t'.
[0064] S3022. Traverse all historical interaction events in the set of historical interaction events that satisfy t' < t, and calculate the cumulative sum according to the following formula:
[0065]
[0066] Among them, is the direct historical interaction excitation term, is the set of historical interaction events between the first node u and the second node v, is a historical interaction event in the set of historical interaction events, indicating that the first node u and the second node v had an interaction of event type at historical time t', is the learnable inter-type excitation weight, which is used to quantify the impact of historical interaction events of event type on event type The incentive effect of the target event, The learnable decay rate parameter is used to control the event type. The duration of the influence of historical interaction events on the target event, where t is the timestamp of the target event. This is the timestamp of the historical interaction event.
[0067] like Figure 2 As shown, the dynamic interactive event sequence contains multiple event types (such as clicks and purchases), each with different incentive patterns and decay rules. This invention addresses this by assigning an independent decay rate parameter to each event type. Inter-type incentive weights This allows the model to automatically learn the "memory half-life" of different behaviors from the data. For example, the decay rate of a single "purchase" behavior. Smaller (longer duration of impact), while the decay rate of a single "click" action. The impact is relatively large (with a short duration), thus enabling refined modeling of the impact patterns of heterogeneous events.
[0068] S3023, Obtain the node embedding vectors of the first node u and the second node v at the target timestamp t. and The dimension of the node embedding vector is d.
[0069] S3024. Obtain the event type of the target event. Associated learnable type-specific transformation matrix .
[0070] S3025. Calculate the baseline strength term according to the following formula:
[0071]
[0072] in, For the baseline intensity term, and These are the node embedding vectors of the first node u and the second node v at the target timestamp t, respectively. For the event type of the target event The associated learnable type-specific transformation matrix, where d is the dimension of the node embedding vector. It is an exponential function. This represents the transpose of a vector.
[0073] S311. With the goal of maximizing the log-likelihood function of the dynamic interactive event sequence, jointly optimize the parameters in the node embedding vector and the conditional strength function.
[0074] S312. Output the optimized node embedding vector.
[0075] In one implementation, S311 specifically includes the following sub-steps:
[0076] S6011. For each positive event in the dynamic interactive event sequence, sample multiple negative events from the node set.
[0077] In one implementation, the negative events are sampled according to the degree distribution of the nodes, that is, the probability of a node being sampled as a negative node is proportional to the frequency of its occurrence in historical interactions.
[0078] S6012. Construct a loss function based on the conditional strength function value of the positive event and the conditional strength function value of the negative event.
[0079] Since the compensation term of the dynamic interactive event sequence is computationally complex on large-scale graphs, this embodiment uses a negative sampling strategy to approximate the log-likelihood function.
[0080] In one implementation, the loss function in S6012 is an approximate log-likelihood loss function based on negative sampling, calculated by the following formula:
[0081]
[0082] in, This is an approximate log-likelihood loss function based on negative sampling. The set of events in the dynamic interactive event sequence. For positive events, Let e be the conditional strength function value of the positive event e. Here, log is the sigmoid function, log is the natural logarithm, and K is the number of negative samples. For the kth sampled negative node, negative event The conditional strength function value.
[0083] S6013. Update the node embedding vector and the parameters in the conditional strength function by minimizing the loss function.
[0084] In one implementation, the Adam optimizer is used to minimize the loss function.
[0085] All parameters (including initial embedding, type matrix) and Hawkes coefficient Learning is achieved through end-to-end optimization. In one implementation, the Adam optimizer is used for parameter updates.
[0086] In one implementation, generating a node embedding vector that evolves over time for the node in S2 further includes the following steps:
[0087] S201. When a new node appears, its node embedding vector is initialized according to the available feature information of the new node, wherein the available feature information includes user demographic features or project content features.
[0088] If the new node has no available feature information, it is initialized using a learnable type-specific prior embedding.
[0089] In one implementation, generating a node embedding vector that evolves over time for the node in S2 further includes the following steps:
[0090] S211. In response to receiving an interaction event, update the embedding vectors of the two interaction nodes based on the pre-interaction embedding vectors of the interaction nodes of the interaction event, the event type of the interaction event, and the time interval between the interaction event and the most recent interaction between each of the two interaction nodes.
[0091] In one implementation, the update step aggregates only the most recent L interactions between the two interacting nodes, where L is a preset positive integer.
[0092] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:
[0093] (1) In this invention, an event type-aware direct historical interaction stimulus term is constructed by assigning an independent and learnable decay rate parameter to each event type. Specifically, in the conditional strength function of the Hawkes process, different types of interactive events (such as clicks, purchases, and favorites) each have their own exclusive decay rate parameter, enabling the model to automatically learn the "memory half-life" of different behaviors from the data. For example, the impact of a "purchase" behavior lasts for a long time and decays slowly, while the impact of a "click" behavior lasts for a short time and decays rapidly. This technique solves the problem that the uniform decay assumption in the prior art cannot distinguish the heterogeneity of event types, and realizes refined modeling of the duration of the impact of different interactive behaviors such as clicks, purchases, and favorites, significantly improving the modeling accuracy of complex behavior sequences.
[0094] (2) In this invention, the incentive relationship between different event types is quantified by setting a learnable inter-type incentive weight matrix. Specifically, inter-type incentive weights are introduced into the direct historical interaction incentive item to describe the incentive intensity of historical interaction event types on the target event type. This weight matrix can automatically learn the incentive patterns between different types of events. For example, the incentive intensity of "buy" on "buy" is higher than that of "click" on "buy", and the incentive intensity of "comment" on "comment" is higher than that of "like" on "comment". This technical means solves the problem of ignoring the incentive differences between event types in the prior art, enabling the model to more realistically reflect the logic of user behavior and improve the prediction accuracy in scenarios such as personalized recommendation, user lifetime value prediction, and precise marketing timing selection.
[0095] (3) In this invention, by incorporating the decay rate parameter and inter-type incentive weights into an end-to-end optimization framework, automatic learning of the influence patterns of event types is achieved. Specifically, all parameters (including node embedding vectors, type-specific transformation matrices, inter-type incentive weights, and decay rate parameters) are jointly optimized with the goal of maximizing the log-likelihood of the dynamic interaction event sequence. The model does not require manual pre-setting of decay parameters or incentive weights, but directly learns the optimal decay rate and incentive pattern for each event type from the data, and can adaptively adjust the degree of response to different interaction densities. This technique solves the problem that parameters in existing methods are difficult to adapt to changes in data distribution, enabling the model to flexibly cope with differences in interaction patterns under different scenarios, and improving the model's generalization ability and practical application value.
[0096] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. An event-type-aware, time-series interaction intensity adaptive attenuation method, characterized in that, include: S1. Obtain a dynamic interactive event sequence, wherein each event in the dynamic interactive event sequence includes a first node, a second node, an event type, and a timestamp; S2. Generate node embedding vectors that evolve over time for the nodes; S3. For the target event in the dynamic interaction event sequence, construct a conditional strength function to predict the probability of the target event occurring. The conditional strength function includes a baseline strength term and a direct historical interaction stimulus term. The direct historical interaction stimulus term quantifies the impact of historical interaction events of different event types on future events as they decay over time by assigning an independent and learnable decay rate parameter to each event type in the dynamic interaction event sequence.
2. The event type-aware temporal interaction intensity adaptive attenuation method according to claim 1, characterized in that, include: The event type described in S1 is a category of mutually exclusive interactive behaviors with specific semantics, including at least one of click, purchase, favorite, or comment.
3. The event type-aware temporal interaction intensity adaptive attenuation method according to claim 1, characterized in that, The conditional strength function described in S3 is expressed as follows: S301, Define the target event Where u is the first node and v is the second node. The event type of the target event is t, and the timestamp of the target event is t. S302, The conditional strength function is constructed as follows: ； in, Let e be the conditional intensity of the target event e at time t. For the baseline intensity term, This refers to the direct historical interaction incentive term.
4. The event type-aware temporal interaction intensity adaptive attenuation method according to claim 3, characterized in that, The direct historical interaction stimulus term described in S302 is calculated through the following steps: S3021. Obtain the set of historical interaction events between the first node u and the second node v. Each historical interaction event in the set of historical interaction events is represented as: This indicates that the first node u and the second node v were in historical time. The event type is Interaction; S3022, Traverse the set of historical interaction events that satisfy... All historical interaction events are summed according to the following formula: ； in, For the direct historical interaction stimulus term, This is the set of historical interaction events between the first node u and the second node v. "A historical interaction event is one of the historical interaction events in the set of historical interaction events, indicating that the first node u and the second node v had an event of type '' at historical time t'." Interaction, These are learnable inter-type incentive weights used to quantify event types. Historical interactive events for event types The incentive effect of the target event, The learnable decay rate parameter is used to control the event type. The duration of the influence of historical interaction events on the target event, where t is the timestamp of the target event and t' is the timestamp of the historical interaction events.
5. The event type-aware temporal interaction intensity adaptive attenuation method according to claim 3, characterized in that, The baseline strength term described in S302 is calculated through the following steps: S3023, Obtain the node embedding vectors of the first node u and the second node v at the target timestamp t. and The dimension of the node embedding vector is d; S3024. Obtain the event type of the target event. Associated learnable type-specific transformation matrix ; S3025. Calculate the baseline strength term according to the following formula: ； in, For the baseline intensity term, and These are the node embedding vectors of the first node u and the second node v at the target timestamp t, respectively. For the event type of the target event The associated learnable type-specific transformation matrix, where d is the dimension of the node embedding vector. It is an exponential function. This represents the transpose of a vector.
6. The event type-aware temporal interaction intensity adaptive attenuation method according to claim 1, characterized in that, S3 also includes a step of optimizing the parameters in the conditional strength function: S311. With the goal of maximizing the log-likelihood function of the dynamic interaction event sequence, jointly optimize the parameters in the node embedding vector and the conditional strength function; S312. Output the optimized node embedding vector.
7. The event type-aware temporal interaction intensity adaptive attenuation method according to claim 6, characterized in that, S601 specifically includes the following sub-steps: S6011. For each positive event in the dynamic interactive event sequence, sample multiple negative events from the node set; S6012. Construct a loss function based on the conditional strength function value of the positive event and the conditional strength function value of the negative event; S6013. Update the node embedding vector and the parameters in the conditional strength function by minimizing the loss function.
8. The event type-aware temporal interaction intensity adaptive attenuation method according to claim 7, characterized in that, The loss function described in S6012 is an approximate log-likelihood loss function based on negative sampling, calculated by the following formula: ； in, This is an approximate log-likelihood loss function based on negative sampling. The set of events in the dynamic interactive event sequence. For positive events, Let e be the conditional strength function value of the positive event e. Here, log is the sigmoid function, log is the natural logarithm, and K is the number of negative samples. For the kth sampled negative node, negative event The conditional strength function value.
9. The event type-aware temporal interaction intensity adaptive attenuation method according to claim 2, characterized in that, S2 also includes the following steps to generate node embedding vectors that evolve over time: S201. When a new node appears, its node embedding vector is initialized according to the available feature information of the new node, wherein the available feature information includes user demographic features or project content features.
10. The event type-aware, time-series interaction intensity adaptive attenuation method according to claim 1, characterized in that, S2 also includes the following steps to generate node embedding vectors that evolve over time: S211. In response to receiving an interaction event, update the embedding vectors of the two interaction nodes based on the pre-interaction embedding vectors of the interaction nodes of the interaction event, the event type of the interaction event, and the time interval between the interaction event and the most recent interaction between each of the two interaction nodes.