Early information propagation prediction method and system based on diffusion model

Abstract

The invention provides an early information propagation prediction method and system based on a diffusion model. The method comprises the following steps: acquiring a statistical user social graph, a user currently participating in propagation of information to be predicted and a time point when the user participates in propagation; determining a user code of a user participating in propagation of the to-be-predicted information based on the user social graph, and determining a user sequence group based on the user code and a time point at which the user participates in propagation; constructing a historical propagation sequence based on the user sequence group sequence; inputting the historical propagation sequence into a cascade representation learning module to obtain a first cascade representation sequence group; the first cascade representation sequence group is input into a pre-trained cascade generation model, the cascade generation model outputs a second cascade representation sequence group, and the cascade generation model is of a diffusion model structure; and inputting the second cascade representation sequence group into a pre-trained popularity prediction module, and outputting a predicted popularity value.

Description

Technical Field

[0001] The present invention relates to the field of information dissemination technology, and in particular to an early information dissemination prediction method and system based on a diffusion model. Background Art

[0002] The popularity prediction of information dissemination is an important research direction in the current field of social network analysis. It has wide application value in recommendation systems, advertising and social media trend analysis. The process of information dissemination is usually described as information cascade, which refers to the process of information spreading dynamically through user networks. The goal of popularity prediction is to quantify the attention a message can receive, such as the number of forwardings or shares. Accurately predicting the popularity of information dissemination can provide a scientific basis for optimizing dissemination strategies, improving user engagement, and formulating marketing plans.

[0003] Existing popularity prediction methods can be divided into two categories: feature-based methods and deep learning-based methods. Feature-based methods make predictions by manually extracting key features such as dissemination content, publisher attributes, and publishing time. Although such methods can provide certain predictive capabilities in the early stages, they are often limited in accuracy due to the complexity and inadequacy of feature design. Deep learning methods have made significant progress in information dissemination modeling in recent years, especially the application of recurrent neural networks and graph neural networks, which provide powerful tools for capturing the time series characteristics and structural characteristics of dissemination. However, these methods rely on a long observation window, and their predictive capabilities usually decline in the early dissemination stage, making it difficult to meet real-time requirements. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides an early information propagation prediction method based on a diffusion model to eliminate or improve one or more defects existing in the prior art.

[0005] One aspect of the present invention provides a method for predicting early information propagation based on a diffusion model, the method comprising the steps of: Obtain the statistical user social graph and the users currently participating in the propagation of the information to be predicted and the time point when the users participated in the propagation; Determine user codes of users who participate in the propagation of the information to be predicted based on the user social graph, and determine a user sequence group based on the user codes and the time point when the user participates in the propagation; Based on each user who participates in the propagation of the information to be predicted and the time point when the user participates in the propagation, a historical propagation sequence is sequentially constructed; Inputting the historical propagation sequence into a pre-trained cascade representation learning module, the cascade representation learning module outputting a first cascade representation sequence group; Inputting the first cascade representation sequence group into a pre-trained cascade generation model, the cascade generation model outputs a second cascade representation sequence group, and the cascade generation model is a diffusion model structure; The second cascade representation sequence group is input into a pre-trained popularity prediction module, and a predicted popularity value is output.

[0006] Using the above scheme, traditional deep learning methods rely on a large amount of historical data and a long observation window, which makes the prediction effect unsatisfactory in the early stage of information dissemination; this scheme constructs a cascade generation model based on the diffusion process, adopts a time interpolator and an inverse denoising mechanism, and can generate high-quality cascade representations. Even when there is insufficient data in the early stage, this scheme can still capture the temporal dynamics and structural characteristics of the cascade, thereby providing more accurate input features for popularity prediction, significantly improving the prediction effect and robustness of the model in the early stage of information dissemination.

[0007] In some embodiments of the present invention, the user social graph is a user relationship represented by an adjacency matrix, the adjacency matrix includes rows and columns corresponding to each user, user codes of users participating in the dissemination of information to be predicted are determined based on the user social graph, and the rows of users participating in the dissemination of information to be predicted in the adjacency matrix are used as user codes.

[0008] In some embodiments of the present invention, in the step of determining a user sequence group based on the user code and the time point at which the user participates in the dissemination, the time point is encoded to obtain a timestamp code, and the user code and the timestamp code are concatenated to obtain a user sequence group.

[0009] In some embodiments of the present invention, the structure of the cascade representation learning module is a series architecture of Transformer and graph attention network.

[0010] In some embodiments of the present invention, the cascade generation model includes a sequentially connected time interpolation denoising module and a reverse denoising module, the time interpolation denoising module includes multiple denoising sub-modules, the reverse denoising module includes multiple denoising sub-modules, the time interpolation denoising module outputs a user sequence group at a target prediction time point, the reverse denoising module is used to update the user sequence group at the target prediction time point, and use the finally updated user sequence group at the target prediction time point as the second cascade representation sequence group.

[0011] In some embodiments of the present invention, a plurality of denoising submodules in the time interpolation denoising module are sequentially connected in series, and each denoising submodule is used to predict a user sequence group at a next time point.

[0012] In some embodiments of the present invention, multiple denoising sub-modules in the reverse denoising module are connected in series sequentially, and each of the denoising sub-modules includes an interpolation network and a prediction network. The interpolation network is used to interpolate the user sequence group between the first cascade representation sequence group and the user sequence group of the target prediction time point to obtain an interpolated user sequence group, and the interpolated user sequence group is input into the prediction network. The user sequence group of the target prediction time point is updated based on the interpolated user sequence group through the prediction network.

[0013] In some embodiments of the present invention, the interpolation network and the noise adding submodule have the same network structure, which are sequentially connected convolutional layers, attention layers, embedding layers, attention layers, embedding layers, attention layers and convolutional layers.

[0014] In some embodiments of the present invention, in the step of inputting the second cascade representation sequence group into a pre-trained popularity prediction module and outputting a predicted popularity value, the popularity prediction module adopts an MLP network.

[0015] The second aspect of the present invention also provides an early information propagation prediction system based on a diffusion model, the system comprising a computer device, the computer device comprising a processor and a memory, the memory storing computer instructions, the processor being used to execute the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the system implements the steps implemented by the method described above.

[0016] The third aspect of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps implemented by the aforementioned early information propagation prediction method based on a diffusion model.

[0017] Additional advantages, purposes, and features of the present invention will be described in part in the following description, and will become apparent to those skilled in the art after studying the following, or may be learned from the practice of the present invention. The purposes and other advantages of the present invention may be specifically pointed out and obtained in the specification and the accompanying drawings.

[0018] Those skilled in the art will appreciate that the objectives and advantages that can be achieved with the present invention are not limited to the above specific description, and the above and other objectives that can be achieved by the present invention will be more clearly understood from the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings described herein are used to provide a further understanding of the present invention, constitute a part of this application, and do not constitute a limitation of the present invention.

[0020] Figure 1A schematic diagram of an implementation of an early information propagation prediction method based on a diffusion model of the present invention; Figure 2 A schematic diagram of a processing architecture of an early information propagation prediction method based on a diffusion model of the present invention; Figure 3 It is another processing architecture schematic diagram of the early information propagation prediction method based on the diffusion model of the present invention; Figure 4 A schematic diagram for comparing the present invention with the prior art; Figure 5 Performance curves of the prior art and this solution in the analysis of sensitivity to observation time; Figure 6 It is the curve of hyper-parameter sensitivity analysis of the existing technology and this solution; Figure 7 Schematic diagram of the ablation results of the existing technology and this solution on the prediction task. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0022] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, only structures and / or processing steps closely related to the solutions according to the present invention are shown in the accompanying drawings, while other details that are not closely related to the present invention are omitted.

[0023] Introduction to existing technology: In the field of information dissemination popularity prediction, existing technologies can be mainly divided into feature-based methods and deep learning-based methods.

[0024] Feature-based methods extract key features in information dissemination (such as content quality, publisher influence, dissemination time, etc.) and combine them with traditional machine learning algorithms for prediction. This type of method can quickly give popularity predictions in the early stages of information dissemination. Because it does not rely on a large amount of historical data support, it has high real-time performance. For example, a feature-based popularity prediction method can predict popularity in the early stages of information dissemination by analyzing the content and user characteristics of the information, and is suitable for scenarios with less data. Nevertheless, these methods rely heavily on feature selection, require manual extraction and optimization of features, and because their modeling is relatively simple, they cannot fully capture complex dissemination patterns, so their accuracy is low.

[0025] Deep learning-based methods can capture the temporal dependencies and structural relationships between users in the process of information dissemination by automatically learning the potential patterns in time series data and graph structure data, thereby greatly improving the prediction accuracy. For example, the DeepHawkes model uses GRU to model the temporal dynamics of information dissemination, and predicts the dissemination trend of information by effectively capturing the dissemination path; and the CasFlow model, which combines the GNN and GRU network structures, can simultaneously consider the structural characteristics and temporal evolution of the dissemination graph, significantly improving the accuracy of popularity prediction.

[0026] Disadvantages of existing technology like Figure 4 As shown in the figure, although existing feature-based methods can make real-time predictions when data is scarce, they face the problems of difficulty in feature selection and low accuracy. Feature selection is crucial for feature-based methods, but this process usually requires manual design and optimization by domain experts and cannot fully capture the complex relationships and deep-level features in the information propagation process. For example, although feature-based popularity prediction methods can handle short-term propagation well, the limitations of feature selection make the prediction accuracy low in complex and long-term propagation processes. In addition, feature-based methods lack deep modeling of the propagation process and cannot effectively capture the temporal dynamics of the information propagation process and the complex interactive relationships between users. Therefore, they have poor accuracy in propagation predictions over long time spans.

[0027] Although deep learning-based methods can automatically extract features and perform complex modeling, they rely on a large amount of historical data for training and require a long observation window to obtain relatively accurate predictions. In the early stages of information dissemination, due to the lack of sufficient historical data, deep learning-based models often cannot provide efficient real-time predictions. Especially in rapidly changing environments such as social media, although the CasFlow model takes into account the graph structure characteristics and dissemination dynamics, it also requires a large amount of historical data for training and cannot provide accurate predictions when the amount of data is small. In addition, the training process of these deep learning models requires a lot of computing resources and a long training time. Especially when processing large-scale social network data, the computational complexity is high, which limits its feasibility in real-time application scenarios.

[0028] Therefore, it is necessary to propose a new method that can generate more accurate and reliable prediction results under the conditions of data scarcity and short observation windows.

[0029] like Figure 1 , 2 As shown in Figure 3, the present invention proposes an early information propagation prediction method based on a diffusion model, the steps of the method include: Step S100, obtaining the statistical user social graph and the users currently participating in the propagation of the information to be predicted and the time point when the users participated in the propagation; In the specific implementation process, the user social graph user indicates whether there is a social relationship between users. Specifically, the social relationship graph can be represented by an adjacency matrix. The rows and columns in the adjacency matrix correspond to users. If two users have a social relationship, the value at the intersection of the row and the column is set to 1, otherwise it is set to 0.

[0030] In the specific implementation process, the user social graph is obtained based on pre-statisticed data. Specifically, it can be constructed for staff to evaluate based on processing experience; it can also be constructed to count the number of interactions between users. The number of interactions between each user is constructed as an input vector, and the adjacency matrix is ​​output through a neural network model.

[0031] Step S200, determining user codes of users who participate in the propagation of the information to be predicted based on the user social graph, and determining a user sequence group based on the user codes and the time point when the user participates in the propagation; In a specific implementation process, in the step of determining the user codes of users participating in the propagation of the information to be predicted based on the user social graph, the values ​​of the rows corresponding to the users in the adjacency matrix are constructed as the user codes.

[0032] Step S300, sequentially constructing a historical transmission sequence based on each user who participated in the transmission of the information to be predicted and the time point when the user participated in the transmission; In a specific implementation process, the users who participate in the propagation of the information to be predicted and the time points at which the users participate in the propagation are regarded as a user group, and each user group is sorted based on the order of the time points to obtain the historical propagation sequence.

[0033] Adopting the above scheme, this scheme captures the temporal and structural characteristics of users in the process of information propagation, and then generates a propagation representation of the information cascade.

[0034] Step S400, inputting the historical propagation sequence into a pre-trained cascade representation learning module, and the cascade representation learning module outputs a first cascade representation sequence group; In a specific implementation, the first cascade representation sequence group corresponds to the last user group in the historical propagation sequence, and the last user group corresponds to the last time point; Specifically, assuming that a given user set and the observed historical cascade set The historical propagation sequence can be expressed as , records the information items in ascending time order The propagation process of each user At the point in time Forwarding information item.

[0035] Step S500, inputting the first cascade representation sequence group into a pre-trained cascade generation model, the cascade generation model outputting a second cascade representation sequence group, the cascade generation model being a diffusion model structure; Step S600: input the second cascade representation sequence group into a pre-trained popularity prediction module, and output a predicted popularity value.

[0036] Using the above scheme, traditional deep learning methods rely on a large amount of historical data and a long observation window, which makes the prediction effect unsatisfactory in the early stage of information dissemination; this scheme constructs a cascade generation model based on the diffusion process, adopts a time interpolator and an inverse denoising mechanism, and can generate high-quality cascade representations. Even when there is insufficient data in the early stage, this scheme can still capture the temporal dynamics and structural characteristics of the cascade, thereby providing more accurate input features for popularity prediction, significantly improving the prediction effect and robustness of the model in the early stage of information dissemination.

[0037] In addition, the present invention has also achieved significant improvements in computational efficiency. By introducing an efficient time-dynamic interpolation mechanism, the model can quickly generate future cascade representations in a shorter time, reducing the reliance on long observation windows and effectively reducing the consumption of computing resources. At the same time, the generation process of the present invention adopts a strategy of step-by-step prediction and interpolation sampling, which greatly improves the real-time adaptability of the model, can quickly respond to new information propagation cascade data changes, and provide efficient and accurate popularity prediction.

[0038] In some embodiments of the present invention, the user social graph is a user relationship represented by an adjacency matrix, the adjacency matrix includes rows and columns corresponding to each user, user codes of users participating in the dissemination of information to be predicted are determined based on the user social graph, and the rows of users participating in the dissemination of information to be predicted in the adjacency matrix are used as user codes.

[0039] In some embodiments of the present invention, in the step of determining a user sequence group based on the user code and the time point at which the user participates in the dissemination, the time point is encoded to obtain a timestamp code, and the user code and the timestamp code are concatenated to obtain a user sequence group.

[0040] In some embodiments of the present invention, the structure of the cascade representation learning module is a series architecture of Transformer and graph attention network.

[0041] In the specific implementation process, Transformer-based and Graph Attention Network (GAT)-based architectures are used to model the contextual dependencies of users in the cascade and the structural information in the social graph, respectively; In some embodiments of the present invention, the cascade representation learning module aims to learn the representation of the input cascade by leveraging the social graph and the user interactions in the diffusion cascade to capture the temporal nature of the information propagation process and the interactive relationship between users. To achieve this goal, a Transformer-based and Graph Attention Network (GAT) architecture is used to model the contextual dependencies of users in the cascade and the structural information in the social graph, respectively.

[0042] In some embodiments of the present invention, the cascade generation model includes a sequentially connected time interpolation denoising module and a reverse denoising module, the time interpolation denoising module includes multiple denoising sub-modules, the reverse denoising module includes multiple denoising sub-modules, the time interpolation denoising module outputs a user sequence group at a target prediction time point, the reverse denoising module is used to update the user sequence group at the target prediction time point, and use the finally updated user sequence group at the target prediction time point as the second cascade representation sequence group.

[0043] In some embodiments of the present invention, a plurality of denoising submodules in the time interpolation denoising module are sequentially connected in series, and each denoising submodule is used to predict a user sequence group at a next time point.

[0044] Adopting the above scheme, this scheme focuses on the macro level of information dissemination through the time interpolation and noise module, and estimates the overall scale of information dissemination by generating cascaded future states; this module simulates the noise process in the diffusion process. First, the time dynamics in the historical data are smoothed by time conditional interpolation to reduce the discontinuity in the propagation process; second, appropriate randomness is introduced through the noise process to simulate the uncertainty and complexity in information dissemination, so as to enhance the robustness of the model and its adaptability to the dissemination trend.

[0045] In some embodiments of the present invention, multiple denoising sub-modules in the reverse denoising module are connected in series sequentially, and each of the denoising sub-modules includes an interpolation network and a prediction network. The interpolation network is used to interpolate the user sequence group between the first cascade representation sequence group and the user sequence group of the target prediction time point to obtain an interpolated user sequence group, and the interpolated user sequence group is input into the prediction network. The user sequence group of the target prediction time point is updated based on the interpolated user sequence group through the prediction network.

[0046] In a specific implementation, in the diffusion process modeling, the present invention uses the Transformer structure to model the interaction of users in each cascade.

[0047] In the specific implementation process, at each time step, the interpolation network is used to smooth the current prediction result and generate an intermediate representation to ensure the continuity between time steps. The specific process is iteratively executed to gradually advance the time evolution until the cascade representation of the target time step is generated.

[0048] The above scheme is adopted, in which the interpolation network introduces appropriate randomness to simulate the uncertainty and complexity in information propagation to enhance the robustness of the model and its adaptability to propagation trends; the prediction network simulates the denoising process in the diffusion process. Based on the future propagation state generated by time interpolation, a popularity increment predictor for information propagation is trained.

[0049] In the specific implementation process, this solution also includes pre-training the cascade generation model, using the mean square error as the loss function to train the cascade generation model.

[0050] In some embodiments of the present invention, the interpolation network and the noise adding submodule have the same network structure, which are sequentially connected convolutional layers, attention layers, embedding layers, attention layers, embedding layers, attention layers and convolutional layers.

[0051] Specifically, the input information is encoded step by step through convolutional layers, temporal embedding layers, and residual cross-attention mechanisms.

[0052] In some embodiments of the present invention, in the step of inputting the second cascade representation sequence group into a pre-trained popularity prediction module and outputting a predicted popularity value, the popularity prediction module adopts an MLP network.

[0053] In the specific implementation process, a multi-layer perceptron (MLP) is used as the prediction network, and the prediction result is finally obtained by learning the relationship between the cascade representation and the popularity increment.

[0054] During the specific implementation process, this solution also includes pre-training the popularity prediction module and using mean square logarithmic error (MSLE) as the loss function. Specifically, by minimizing the loss function, the popularity prediction module can accurately learn the mapping relationship between cascade evolution and final popularity, and realize the prediction of future popularity.

[0055] In summary, the present invention solves the limitations of the information diffusion popularity prediction task in the prior art when faced with problems such as data scarcity, strong temporal dynamics, and high real-time requirements. Although existing deep learning-based methods can model complex time series and graph structure data well, their reliance on large amounts of historical data and long observation windows makes these methods poor in data-scarce scenarios, especially at the beginning of information diffusion. Therefore, the present invention proposes a new generative information diffusion prediction method that overcomes the challenges of insufficient data and early prediction by generating cascade representations, thereby providing more accurate popularity predictions; The deep learning methods of the prior art usually require a long training time and a large amount of computing resources when predicting popularity, which makes them unable to perform at their best in scenarios with high real-time requirements. In the early stages of information dissemination, since the dissemination information has not yet been fully developed, it is difficult for traditional models to quickly adapt and provide effective predictions. Therefore, the present invention reduces the dependence on a large amount of historical data by designing an efficient generation model, and at the same time, by introducing time dynamic interpolation based on the diffusion process, the model can quickly adapt to new dissemination cascade data, thereby improving the accuracy and efficiency of real-time predictions.

[0056] Experimental example: 1. Experimental Setup 1. Dataset This experiment evaluates the proposed framework on three public datasets: Twitter, Weibo, and APS.

[0057] In this experiment, 70% of the cascades are randomly selected as training sets, 15% as validation sets, and the remaining 10% as test sets. The detailed information of the data set is shown in Table 1: Table 1 2. Comparison Method This experimental example compares this solution with the following prior arts, which include: DeepHawkes: Models each cascade as a collection of diffusion paths between users, using gated recurrent units (GRUs) to capture the sequential progression of the cascade.

[0058] MS-HGAT: Constructs a series of temporally sampled hypergraphs that encapsulate multiple cascades and users, and uses hypergraph learning to compute cascade representations.

[0059] CasCN: We model each cascade as a temporal graph sequence and utilize a combination of graph neural networks (GNNs) and long short-term memory (LSTM) networks to learn robust cascade representations.

[0060] TempCas: Integrates specialized sequence modeling techniques aimed at capturing overall temporal patterns, complementing its learning on cascade graphs.

[0061] CasFlow: We first derive user representations from social networks and cascade graphs, and then use GRU combined with variational autoencoders (VAE) to encode the cascade representations.

[0062] CTCP: A state-of-the-art cascade prediction method that groups multiple cascades by sharing propagating users, enabling a unified temporal and structural learning process across cascades.

[0063] 3. Evaluation Metrics This experimental example uses the following four widely recognized metrics to evaluate the performance of the compared methods: Mean Squared Logarithmic Error (MSLE): quantifies the prediction error by calculating the mean of the squared difference between the logarithm of the predicted value and the logarithm of the true value. A lower MSLE value indicates a smaller logarithmic difference between the model prediction and the true value, and a higher prediction accuracy.

[0064] Mean absolute logarithmic error (MALE): The average of the absolute differences between the logarithms of the predicted values ​​and the logarithms of the true values. A lower MALE value means that the predicted values ​​differ less on average from the true values ​​on a logarithmic scale.

[0065] Mean Absolute Percent Error (MAPE): Measures the percentage of forecast error by calculating the average of the absolute value of the difference between the forecast and the true value and the true value. A lower MAPE value indicates a smaller average percentage difference between the forecast and the true value and a higher forecast accuracy.

[0066] Pearson Correlation Coefficient (PCC): Measures the linear correlation between the predicted value and the true value. A higher PCC value (close to 1) indicates a stronger positive correlation between the predicted value and the true value, indicating that the model is better able to capture trends and patterns in the data.

[0067] 2. Performance Evaluation 1. Model Performance In order to compare the performance of this solution under different data sets, the experimental examples were analyzed on Twitter, Weibo, and APS data. The results are shown in Table 2: Table 2 The following important conclusions can be drawn from the results in Table 2: Feature-based models, such as DeepHawkes, lag behind other methods on all evaluation metrics. This can be attributed to the inherent limitations of feature-based models in capturing the complex nonlinear evolution patterns of cascade size, especially in the early stages of information propagation.

[0068] Graph-based models generally outperform sequence-based models, which illustrates the importance of integrating structural and temporal information in cascade graphs. For example, CasFlow and CTCP show good performance on the Weibo dataset, indicating their effectiveness in modeling the temporal and structural dynamics of information propagation. However, their performance is not outstanding over shorter observation periods.

[0069] Our scheme outperforms other models on all metrics, especially on the Twitter and Weibo datasets with shorter observation windows, which highlights its ability to generate accurate predictions with less initial data.

[0070] 2. Observation time sensitivity analysis To further evaluate the performance of this solution under different observation time windows, this experimental example conducted observation time sensitivity analysis on the Twitter and APS datasets and plotted the performance curves of the Top3 models in these intervals.

[0071] like Figure 5 As shown in Figure 2, the PCC values ​​of all models increase with the extension of the observation time window. This indicates that longer observation time windows help the model better capture the underlying dynamics of information propagation. However, our scheme maintains a high PCC value in all observation time windows. In particular, our scheme significantly outperforms other models in the early stages of the information cascade (i.e., in shorter observation time windows), which indicates that our scheme can effectively capture the initial dynamics of information propagation and better model the cascade dynamics when data is sparse.

[0072] 3. Hyperparameter sensitivity analysis In order to evaluate the sensitivity of the model to hyperparameters, this experiment performed a sensitivity analysis on two key hyperparameters: time step and number of iterations in the sampling process. from Figure 6 It can be seen that if the time step is too large, the generated embedding may not be consistent with the true cascade evolution distribution. In addition, too many iterations may introduce noise, thereby reducing the effectiveness of the model.

[0073] 3. Ablation Experiment In order to evaluate the importance of each module in this scheme, this experimental example conducted an ablation experiment on the Twitter and APS datasets. This experimental example removed the temporal learning module (TL), the structural representation module (SL), and the cascade generation module (GM) respectively, and evaluated the performance of these variants. The relevant results are as follows Figure 7 shown.

[0074] Specifically, the time learning module is removed, that is, there is no time point in the user sequence group that is propagated through user participation; the encoding of the obtained time; Removing the structural representation module means that the user encoding obtained through the user social graph does not exist in the user sequence group; Removing the cascade generation module (GM) is the removed cascade generation model.

[0075] like Figure 7 As shown in Figure 2, removing the temporal learning module (TL): Removing the TL module leads to significant drops in MSLE and PCC, especially on the Twitter dataset. This suggests that temporal dynamics play a key role in the early stages of information diffusion, where the timing and rhythm of retweets or citations are key indicators of future cascade popularity. The temporal features captured by this module are particularly important when the observation period is short, such as in the social media environment.

[0076] Removing the Structural Representation Module (SL): Removing the SL module also leads to a drop in performance, but its impact is smaller than that of the TL module, especially on the APS dataset. In networks like APS, where the structure changes slowly, temporal evolution may play a more important role in predicting future cascade growth.

[0077] Removing the Cascade Generation Module (GM): Removing the GM module leads to a significant drop in performance, which shows that the cascade generation module is crucial in this application. The ability of this module to simulate future cascade propagation based on limited initial data is crucial for long-term forecasting. Its significant contribution to the MSLE and PCC metrics highlights the importance of generating synthetic data.

[0078] In summary, this scheme proposes a novel generative approach for the task of information diffusion prediction, aiming to address the challenge of early information popularity prediction. This scheme draws on the gradual denoising and denoising process of the diffusion model, uses a temporal conditional interpolator as a denoising process, and a denoising process for prediction, iteratively generating cascade representations that can effectively capture the rich temporal dynamics of information diffusion. Subsequently, the sampling stage produces cascade representations, which are input into the prediction module to generate popularity predictions. The comprehensive evaluation of this scheme on different datasets such as Twitter, Weibo, and APS highlights the effectiveness of this scheme, demonstrating that it outperforms the state-of-the-art methods in early prediction tasks. This work not only improves the accuracy and timeliness of information popularity prediction, but also has important potential value for applications such as recommendation systems, targeted advertising, and social media trend analysis.

[0079] The beneficial effects of the present invention include: 1. The present invention models the information propagation process as a diffusion generation process by combining a time interpolator and an inverse denoiser. In the forward diffusion, the time interpolator is used to achieve smooth generation of cascade representations at different time steps, thereby capturing temporal dynamics and early propagation characteristics. In the inverse denoising process, the future cascade representations are generated by gradual denoising, effectively solving the problems of data scarcity and long-term observation window dependence; 2. The present invention proposes a generation and sampling mechanism that combines interpolation and denoising. Through step-by-step prediction and interpolation correction, the model can generate future cascade representations in a short time, reducing the reliance on long-term historical data. At the same time, the mechanism ensures that the generation process has temporal smoothness and continuity through continuous modeling of time steps, thereby improving the prediction efficiency and real-time performance of the model.

[0080] 3. The present invention does not rely on a large amount of historical data, and can achieve more accurate cascade feature learning when the data is sparse or the observation window is small, and can meet the timeliness requirements of computing efficiency.

[0081] An embodiment of the present invention also provides an early information propagation prediction system based on a diffusion model, the system comprising a computer device, the computer device comprising a processor and a memory, the memory storing computer instructions, the processor being used to execute the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the system implements the steps implemented by the method described above.

[0082] The embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps implemented by the above-mentioned early information propagation prediction method based on the diffusion model are implemented. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the technical field.

[0083] It should be understood by those skilled in the art that the exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application-specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier.

[0084] It should be clear that the present invention is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present invention.

[0085] In the present invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with features of other embodiments or replace features of other embodiments.

[0086] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the embodiments of the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An early information propagation prediction method based on a diffusion model, characterized in that: The steps of the method include: Obtain the statistical user social graph and the users currently participating in the dissemination of the information to be predicted and the time point when the users participated in the dissemination; Determine user codes of users who participate in the propagation of the information to be predicted based on the user social graph, and determine a user sequence group based on the user codes and the time point when the user participates in the propagation; Based on each user who participates in the propagation of the information to be predicted and the time point when the user participates in the propagation, a historical propagation sequence is sequentially constructed; Inputting the historical propagation sequence into a pre-trained cascade representation learning module, the cascade representation learning module outputting a first cascade representation sequence group; Inputting the first cascade representation sequence group into a pre-trained cascade generation model, the cascade generation model outputting a second cascade representation sequence group, the cascade generation model being a diffusion model structure; The second cascade representation sequence group is input into a pre-trained popularity prediction module, and a predicted popularity value is output.

2. The early information propagation prediction method based on diffusion model according to claim 1 is characterized in that: The user social graph is a user relationship represented by an adjacency matrix, wherein the adjacency matrix includes rows and columns corresponding to each user. User codes of users participating in the propagation of information to be predicted are determined based on the user social graph, and the rows of users participating in the propagation of information to be predicted in the adjacency matrix are used as user codes.

3. The early information propagation prediction method based on diffusion model according to claim 1 is characterized in that: In the step of determining a user sequence group based on the user code and the time point at which the user participates in the dissemination, the time point is encoded to obtain a timestamp code, and the user code and the timestamp code are concatenated to obtain a user sequence group.

4. The early information propagation prediction method based on diffusion model according to claim 1 is characterized in that: The structure of the cascade representation learning module is a series architecture of Transformer and graph attention network.

5. The early information propagation prediction method based on the diffusion model according to any one of claims 1 to 4, characterized in that: The cascade generation model includes a sequentially connected time interpolation denoising module and a reverse denoising module, the time interpolation denoising module includes a plurality of denoising submodules, the reverse denoising module includes a plurality of denoising submodules, the time interpolation denoising module outputs a user sequence group at a target prediction time point, the reverse denoising module is used to update the user sequence group at the target prediction time point, and use the finally updated user sequence group at the target prediction time point as the second cascade representation sequence group.

6. The early information propagation prediction method based on diffusion model according to claim 5 is characterized in that: The multiple noise adding submodules in the time interpolation noise adding module are connected in series in sequence, and each noise adding submodule is used to predict the user sequence group at the next time point.

7. The early information propagation prediction method based on diffusion model according to claim 5 is characterized in that: The multiple denoising submodules in the reverse denoising module are connected in series in sequence, and each of the denoising submodules includes an interpolation network and a prediction network. The interpolation network is used to interpolate the user sequence group between the first cascade representation sequence group and the user sequence group at the target prediction time point to obtain an interpolated user sequence group, and the interpolated user sequence group is input into the prediction network. The user sequence group at the target prediction time point is updated based on the interpolated user sequence group through the prediction network.

8. The early information propagation prediction method based on diffusion model according to claim 7 is characterized in that: The interpolation network and the noise adding submodule have the same network structure, which are sequentially connected convolutional layer, attention layer, embedding layer, attention layer, embedding layer, attention layer and convolutional layer.

9. The early information propagation prediction method based on diffusion model according to claim 1 is characterized in that: In the step of inputting the second cascade representation sequence group into a pre-trained popularity prediction module and outputting a predicted popularity value, the popularity prediction module adopts an MLP network.

10. An early information propagation prediction system based on a diffusion model, characterized in that: The system includes a computer device, which includes a processor and a memory, wherein the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps implemented by the method as described in any one of claims 1 to 9.

Images