Information microcosmic diffusion prediction model fusing cascade frequency domain features

Abstract

The invention relates to the technical field of information propagation prediction, and discloses an information microcosmic diffusion prediction model fusing cascade frequency domain characteristics, which comprises the following steps: firstly, constructing a social graph and an information diffusion hypergraph, and capturing the influence of social homogeneity and social influence factors on information propagation by using the powerful capability of a graph convolutional neural network; then, Fourier transform is ingeniously applied, time domain information of the cascade is mapped to a frequency domain, and the deep matching degree between the user and the cascade is learned by combining high-frequency and low-frequency components. Finally, the model performance is further optimized through an attention fusion layer, and the purpose is to generate more expressive user representation, so that the problems of feature redundancy and information loss are effectively avoided. Experiments on four public datasets, such as Douban and Meteracker, show that compared with the existing most advanced baseline model, the method provided by the invention has the advantages that the two key evaluation indexes, namely Hits (at) K and Map (at) K, are obviously improved, and the highest values are up to 10.94% and 23.03%.

Description

Technical Field

[0001] The present invention relates to the technical field of information propagation prediction, and in particular to an information microscopic diffusion prediction model integrating cascade frequency domain features. Background Art

[0002] The rise of social networks has completely changed people's lives and communication patterns. There is a large amount of information on the network platform. The way and path of the dissemination of this information have undergone tremendous changes due to the emergence of online social media such as Weibo and Facebook. However, the development of social networks has a dual nature. From a positive perspective, the powerful dissemination ability enables social networks to play an indispensable role in the fields of user personalized analysis and social recommendation. Therefore, the research on information diffusion prediction has important practical significance and social value, and provides basic support for monitoring and controlling the dynamic dissemination of information.

[0003] The task of predicting the microscopic diffusion of information cascades aims to explore the next node to be infected during the process of information propagation. Internet users receive a large amount of information every day, but it is uncertain whether they will forward this information. That is, the spread of information is not a simple one-way flow, but is related to multiple factors, such as personal interest preferences, social homogeneity, and social influence. These factors work together to determine the path and scale of information propagation on the network. The present invention hopes to infer complex influencing factors and determine users who may participate in the cascade in the future.

[0004] The process of information diffusion is usually complex and dynamic, and traditional time domain analysis methods are difficult to fully capture the laws involved. In terms of information diffusion modeling, researchers have proposed a variety of methods to predict information diffusion over the past few decades. Traditional methods either rely on the linear relationship of predefined propagation mechanisms or focus on extracting manual features such as timing, structure, and users. However, the information diffusion patterns in the real world are not static, and these two types of methods are limited by limited assumptions or cumbersome feature extraction.

[0005] In recent years, with the emergence and development of deep neural networks, end-to-end prediction methods based on deep learning have been proposed to improve the accuracy and efficiency of prediction. For example, TopoLSTM expands and improves the standard LSTM model to simulate the influence between cascades; DeepDiffuse uses temporal information through RNN and attention mechanism; SNIDSA introduces structural attention module and gating strategy to integrate structural information into the sequential information of RNN; Inf-VAE and DyDiff_VAE learn social homogeneity and temporal information based on the variational autoencoder framework to better learn the representation of information cascades and achieve higher efficiency; DyHGCN and GEGCN construct heterogeneous graphs containing social graph nodes and propagation graph nodes, and at the same time Learn the social relations of users and the propagation structure of cascades; NDM uses attention mechanism and convolutional network to model cascades to improve the generalization ability of the model; FOREST proposes a multi-scale diffusion prediction model that unifies micro and macro scales based on reinforcement learning; MINDS introduces adversarial training and orthogonal constraints to enhance user representation on this basis; MSHGAT and HyperINF obtain the interaction preferences of users in social networks through a series of diffusion hypergraphs divided by timestamps; H-Diffu and RotDif propose a hyperbolic space representation method for jointly modeling diffusion cascades and social graphs in hyperbolic space.

[0006] In the frequency domain, Fourier analysis plays an important role in studying the general approximation properties of neural networks. Discrete Fourier transform, especially fast Fourier transform, has been used to solve signal processing problems, such as fitting neural networks to electrocardiogram signals or evolving solutions of partial differential equations. Fourier transform realizes the conversion from time domain to frequency domain, and many works apply Fourier transform to different scenarios, such as social recommendation, time series prediction and other fields. However, in the field of cascade prediction, there is almost no research related to Fourier transform.

[0007] Existing methods have solved the above problems to a certain extent and achieved good results, but there are still three significant limitations: First, the user and cascade representations are mainly learned by sampling the information cascade diffusion subgraph, which leads to the loss of social influence of related nodes and dynamic information between subgraphs; second, all users participating in the cascade are treated equally, and loyal users with stable attraction and users attracted for a short period of time are not distinguished, and different users contribute differently to the cascade. Although it is relatively easy to obtain information about users who participate stably, it is extremely challenging to capture information about users who burst out instantly; third, learning the embedded representation of users from the social graph and the diffusion graph independently, and then fusing them through a simple gating method, may not accurately describe the influence of users, and may retain redundant information or lose important information during the fusion process. These defects lead to unsatisfactory performance of existing methods. In addition, the information propagation process is usually complex and dynamic, and traditional time domain analysis methods are difficult to fully capture the laws therein.

[0008] Therefore, it is necessary to provide an information micro-diffusion prediction model that integrates cascaded frequency domain features to solve the above technical problems. Summary of the invention

[0009] In order to solve the above problems, the present invention proposes an Integrating Cascade Frequency Domain Features for Information Micro-diffusion Prediction (FFIDP) model that integrates cascade frequency domain features. The time domain signal is converted to the frequency domain by Fourier transform, and the frequency characteristics in information propagation are used to fully fuse low-frequency signals and high-frequency signals, so as to more accurately predict the information propagation path and behavior. In addition, in order to capture the correlation between social factors and frequency factors, FFIDP designs an additional fusion layer to aggregate user features. Compared with splicing operations and gated fusion, the introduced attention fusion strategy retains effective information by calculating the correlation score of the two vectors, further improving the prediction accuracy.

[0010] The information micro-diffusion prediction model integrating cascade frequency domain features provided by the present invention comprises: a user social homogeneity learning module, which is used to learn the social relationship of users through a graph convolutional network;

[0011] A frequency domain learning module is used to obtain interaction-based user embeddings via diffusion hypergraphs and then learn concatenated frequency features in the Fourier domain;

[0012] The decoupled information fusion module is used to learn the interactive features within the cascade, fuse the user embeddings at two levels, and retain the effective information to the maximum extent;

[0013] Cascade propagation prediction module is used to calculate the probability of candidate nodes being infected.

[0014] Preferably, the learning of the social homogeneity of users comprises the following steps:

[0015] Build a social relationship graph based on the user's social relationships;

[0016] A two-layer graph convolutional network is used to learn user social homogeneity on the social relationship graph and obtain the social homogeneity representation of all users.

[0017] Preferably, the frequency domain learning module includes:

[0018] Diffused hypergraph convolution, obtaining global information through the hypergraph attention network;

[0019] The matching degree between users and cascades is learned, and the user representation matrix is ​​mapped to the frequency domain through Fourier transform.

[0020] Preferably, the decoupling information fusion module includes:

[0021] Interactive learning within the cascade, using a multi-head self-attention module to explore local diffusion dependencies within the cascade;

[0022] The attention fusion module retains effective information by calculating the correlation score of two vectors.

[0023] Preferably, the cascade propagation prediction module includes:

[0024] Calculate the probability of the target user participating in the cascade;

[0025] The cross entropy loss function is used as a constraint.

[0026] Compared with related technologies, the information microscopic diffusion prediction model integrating cascaded frequency domain features provided by the present invention has the following beneficial effects:

[0027] 1) The present invention provides a novel perspective, which maps features from the time domain to the frequency domain through Fourier transform, identifies the frequency characteristics in the propagation process, and effectively uses high-frequency information and low-frequency information to diffuse prediction information.

[0028] 2) The present invention designs an attention fusion module to effectively integrate the different influencing factors of user forwarding behavior and solve the problems of feature redundancy and information loss.

[0029] 3) The model proposed in this paper shows strong competitiveness on four real datasets. Compared with the most advanced baseline work, the experimental results show that FFIDP improves the Hits@K score by 10.94% and the Map@K indicator by 23.03%. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1The overall framework diagram of an information microscopic diffusion prediction model integrating cascaded frequency domain features provided by the present invention;

[0031] Figure 2 is the influence diagram of high frequency and low frequency components in the present invention;

[0032] Figure 3 It is the influence diagram of the embedding dimension in the present invention;

[0033] Figure 4 is the influence diagram of the maximum cascade length in the present invention;

[0034] Figure 5 It is the influence diagram of the training ratio in the present invention. DETAILED DESCRIPTION

[0035] First of all, it should be noted that the forwarding behavior of users will be affected by two aspects. One is social homogeneity, which means that users with social connections tend to have similar preferences; the other is social influence, that is, the behavior of users will be influenced by their friends. In view of social homogeneity, the present invention constructs a social friendship graph G based on the social relationships of users. S =(U S ,E S ), where U S is the set of user nodes, E S is a set of social edges. If user u i ,u j ∈U S If there is a social relationship between them, then the edge e ij Save to E S . Constructing a propagation hypergraph HG for social influence D (U D ,E D ), similarly, U D ,E D are the set of user nodes and the set of edges through which information spreads among users. The cascade diffusion prediction task can be described as: predicting the participation cascade c based on the known social graph, diffusion graph, and historical cascades m ={(u1,t1),(u2,t2),…,(u k ,t k )},k=1,2,…,|c m |-1| The next user u k+1 .

[0036] The information microscopic diffusion prediction model integrating cascaded frequency domain features proposed in the present invention is further described below with reference to the accompanying drawings.

[0037] 1. Methods

[0038] This paper introduces an information microscopic diffusion prediction model that integrates cascaded frequency domain features. The overall architecture of FFIDP is as follows: Figure 1 As shown in the figure, it mainly consists of four parts: 1) User social homogeneity learning module, which learns the social relationship of users through graph convolutional network; 2) Frequency domain learning module, which obtains interaction-based user embedding through diffusion hypergraph and then learns the frequency features of cascade in Fourier domain; 3) Decoupled information fusion module, which learns the interaction features within the cascade, fuses the user embeddings of two levels, and retains the effective information to the maximum extent; 4) Cascade propagation prediction module, which calculates the probability of candidate nodes being infected. The following sections will introduce each part of the framework in detail.

[0039] 1.1. User social homogeneity learning module

[0040] Users tend to interact with users who have similar interests. This involves the principle of social homogeneity, that is, close friends usually have common interests in some aspects. Therefore, the social homogeneity of users can be learned through friend relationships, and for users who have not participated in any cascades, the characteristics of their neighbors can also be explored through the social relationship graph to understand their preferences. Therefore, social homogeneity learning can alleviate the cold start problem to a certain extent. A two-layer graph convolutional network is used to build a network on the social relationship graph G. S Learning user social homogeneity:

[0041]

[0042] The user's initial representation It is obtained by random initialization with normal distribution, N is the number of users, d is the dimension of embedding, and W S is a trainable parameter, and Self-loop G S The adjacency matrix and degree matrix of . After two layers of GCN convolution, the social homogeneity representation U of all users is obtained S .

[0043] 1.2 Frequency Domain Learning Module

[0044] Users' social homogeneity can reflect their social preferences, but the representation learned through social homogeneity cannot accurately reflect the preferences of users in participating in cascades. Therefore, the present invention constructs a propagation hypergraph based on historical cascades, learns user interactions at the information propagation level, and uses Fourier transform to capture the matching degree between users and cascades in the frequency domain.

[0045] 1.2.1 Diffused Hypergraph Convolution

[0046] The hypergraph contains the interaction information between users and cascades, so the hypergraph attention network is first used to obtain global information. The hypergraph convolution is divided into two stages: from supernode to hyperedge and from hyperedge to supernode. j The information of all super nodes in the update is used to update the characteristics of the hyperedge, which can be defined as:

[0047]

[0048] in is a trainable weight matrix, and d is the dimension of the embedding. In order to improve the expressiveness of the model and enable interactive learning, the hypergraph attention network uses the user representations learned from the friendship graph as the initial input instead of initializing the user representations from the normal distribution. ij Indicates u i In e j The attention score in is calculated as follows:

[0049]

[0050] where dis(·) is the Euclidean distance. j Node r j It can partially reflect the content of the cascade, so the convolution process retains the information of all such nodes and calculates the Euclidean distance between other nodes and this node as the attention score.

[0051] Then, u i The information of all participating hyperedges is aggregated to u i , using the obtained hyperedge features to update the supernode, the process can be defined as:

[0052]

[0053] Represents user u i The hyperedge collection. After the hypergraph attention convolutional network learns all user representations U D .

[0054] 1.2.2 User and Cascade Matching Learning

[0055] The user representation obtained from the diffusion hypergraph covers the coarse-grained global information, but is not enough to describe the fine-grained matching degree between users and cascades. Therefore, this paper proposes a new perspective to learn the details of user participation in a cascade by mapping values ​​to the frequency domain through Fourier transform. Fourier transform is the key to digital signal processing, which projects a series of values ​​into the frequency domain. Usually used Denotes the Fourier transform, and Denotes the inverse Fourier transform. Applying the Fourier transform to a feature can be achieved by matrix multiplication, where the matrix is ​​composed of the Fourier basis where i is the imaginary unit and j is the jth row. Given the target cascade c m , From U D The user representations found in are arranged in the order of participation in the cascade. To learn the similarity between users and cascades, we first perform a Fourier transform along the dimension of the cascade to convert the globally dependent user representation matrix Mapping to the frequency domain:

[0056]

[0057] Low-frequency signals usually represent long-term trends and stable patterns in information propagation. For example, a core user group that pays attention to a topic for a long time or information slowly spreads over a long period of time, forming a stable propagation path. In this case, the propagation signal is manifested as a low-frequency component in the frequency domain. We use the low-frequency signal (common signal) extracted by LPC[·] to represent users with a high degree of cascade matching. High-frequency signals represent short-term fluctuations and rapid changes in information propagation. The high-frequency signal (local fluctuation signal) extracted by HPC[·], which represents users with a low degree of cascade matching, does not change much and is more stable than the high-frequency signal. However, if only the low-frequency signal is used, it will lead to the problem of over-flattening, so we selectively use high-frequency signals and learn how to best fuse them.

[0058]

[0059] in, Contains the first c users participating in each cascade, Contains the remaining |c m |-c users. In order to better integrate the low-frequency and high-frequency components, a trainable parameter β is set to adjust the frequency ratio:

[0060]

[0061] 1.3 Decoupling Information Fusion Module

[0062] 1.3.1 Interactive Learning within Cascade

[0063] Given a target cascade c m , is the corresponding user representation found from the social homogeneity feature, arranged in the order of participation in the cascade. We expect to aggregate the information of the first k users to predict the next affected user u (k+1)However, although graph-based representation learning can capture the collaborative relationship between users, it is insufficient in analyzing the contextual interactions within the cascade. Given the outstanding performance of Transformer in sequential tasks such as sequence recommendation, we use a multi-head self-attention module to explore the local diffusion dependencies within the cascade:

[0064]

[0065] h1 is the number of heads of multi-head attention, Mask is a mask matrix introduced to avoid information leakage, which is 0 when i≤j, otherwise -∞. is a learnable parameter matrix. Finally, two layers of fully connected layers are used in the feedforward neural network to capture the characteristics of the interactive information in the aggregation cascade:

[0066]

[0067] W3 and W4 are learnable parameter matrices, b1 and b2 are bias parameters, is also obtained in a similar manner.

[0068] 1.3.2 Attention Fusion Module

[0069] Inspired by multimodal information fusion and other works, in order to obtain a more expressive embedding representation, we consider fusing features at two levels. Due to the information redundancy and information loss caused by gated fusion, this paper introduces an attention fusion layer. First, we use and Generate keys and queries respectively, thus obtaining the attention matrix and

[0070]

[0071]

[0072] is a learnable parameter matrix. Then the two are combined to obtain the associated attention scores:

[0073]

[0074] Finally, the outputs of all attention heads are connected and sent to the feedforward layer, where the MLP and ReLU activation functions are combined to further capture nonlinear features:

[0075]

[0076] W5 and W6 are learnable parameter matrices, b3 and b4 are bias parameters, and h2 is the number of attention heads. This strategy not only effectively solves the above problems, but also avoids unnecessary attention randomness caused by mixed associations of heterogeneous embeddings, as well as inflexible gradients and uncertain cross-relationships between different features, achieves reasonable and stable self-attention, and enables flexible gradients to adaptively learn various information in different scenarios to improve the expressiveness of the model.

[0077] 1.4 Cascade Propagation Prediction Module

[0078] The probability of the target user participating in the cascade is calculated by the following formula:

[0079]

[0080] W p Is X m Parameter matrix mapped to a specific space, Mask p Used to block users who are activated before prediction. The present invention uses the cross entropy loss function as a constraint:

[0081]

[0082] θ is the set of all trainable parameters, y ij Represents the true label, if u j is the i-th participant in cascade c m For users with ij =1, otherwise y ij =0.

[0083] 2. Experimental results and analysis

[0084] In order to further evaluate and analyze the effectiveness of the proposed model, the present invention conducts extensive experiments on FFIDP and its variants on four real datasets.

[0085] 2.1 Experimental Setup

[0086] 2.1.1 Experimental Dataset

[0087] We evaluated the performance of the model on four publicly available datasets. The statistics of the datasets are shown in Table 1:

[0088] 1) Douban: This dataset was collected by Zhong et al. from the Douban website, where users can follow other users and share, forward, and comment on their favorite books or movies. A book or a movie is considered an information item, and if a user marks it as read or watched, the user is considered to have participated in the dissemination of this information.

[0089] 2) Memetracker: collected news stories and blog posts from August to October 2008 on online websites and collected the most frequently used quotes and phrases as memes. Each meme is an information item and each URL of the website is a user, i.e. this dataset has no underlying social graph.

[0090] 3) Android: This dataset was collected by Sankar et al. from the community question-and-answer website StackExchanges. It contains users’ interactions on Android topics in different channels, and the questions and answers between users constitute user friendship relationships.

[0091] 4) Christianity: This dataset comes from the same source as Android and contains users’ interactions on the StackExchanges website on Christian topics.

[0092] Table 1 Dataset statistics

[0093]

[0094] 2.1.2 Evaluation Indicators According to the definition of information micro-diffusion prediction, this task can be regarded as a retrieval problem. Therefore, this paper uses the commonly used model ranking performance evaluation indicators Map@K and Hits@K in information retrieval tasks to evaluate the model. In this study, we choose K = [10, 50, 100] for evaluation. Higher Map@K and Hits@K scores indicate that the model has better prediction performance.

[0095] 1) Map@K: Calculates the average precision of the top K prediction results and takes into account the order of users in the ranked prediction results.

[0096] 2) Hits@K: Calculate the probability that the first K prediction results contain actual infected users, regardless of the ranking order.

[0097] 2.1.3 Parameter settings

[0098] All experiments of the present invention are conducted on a 12GB GeForce GTX 4070Ti (GPU). 80% of the cascades are randomly selected for training, 10% for validation, and the remaining 10% for testing for each dataset. The maximum cascade length is 200. The baseline model retains the settings provided in the original paper. The present invention implements the FFIDP model in PyTorch and optimizes the parameters using the Adam optimizer with a learning rate of 0.001, a packet loss rate of 0.3, a batch size of 64, and an embedding dimension of 64. In addition, we use a two-layer graph convolutional network to learn user social homogeneity and a single-layer hypergraph attention convolutional network to learn user representations at the cascade level. The number of self-attention heads is set to 14.

[0099] 2.2 Comparative Experiment

[0100] 2.2.1 Comparison Model

[0101] To verify the effectiveness of FFIDP, we compare it with the following state-of-the-art models.

[0102] 1) Topo-LSTM: The sequential cascade is reconstructed into a directed acyclic graph and the standard LSTM model is extended to simulate the information diffusion process.

[0103] 2) SNIDSA: It learns cascaded sequential information through RNN and jointly learns heterogeneous information representation by computing the structural attention of diffusion paths, thereby exploring diffusion paths and social networks.

[0104] 3) DeepDiffuse: Utilizes RNN and attention mechanism to sequentially model the cascade of previously affected users.

[0105] 4) FOREST: A multi-scale diffusion prediction model based on reinforcement learning is proposed. The model uses GRU to learn sequence features and GCN to extract network structure information, thereby achieving effective prediction of the information diffusion process.

[0106] 5) NDM: Cascade modeling uses deep learning techniques such as attention mechanisms and convolutional networks to make relaxed independence assumptions to alleviate long-term dependencies.

[0107] 6) Inf-VAE: Social homogeneity is embedded through GNN, and a co-attention fusion network is designed based on variational autoencoder to integrate social and temporal variables.

[0108] 7) DyHGCN: Encode temporal information into social graphs and dynamic diffusion graphs to construct heterogeneous graphs and dynamically learn user preferences.

[0109] 8) MS-HGAT: A series of hypergraphs are constructed to dynamically describe the interactions between users and cascades, and memory and gating mechanisms are used to effectively store and fuse features.

[0110] 9) MINDS: Use macro prediction results to guide micro predictions, enhance the intrinsic connection between the two tasks, and introduce adversarial training and orthogonality constraints to reduce feature redundancy and ensure the integrity of shared features.

[0111] 2.2.2 Performance Comparison

[0112] The present invention compares FFIDP with the baseline model on four public datasets. The results are shown in Tables 2 and 3. The best scores are indicated in bold and the second best scores are indicated in underline. The following conclusions can be drawn from the analysis of the experimental results:

[0113] The proposed model significantly outperforms all state-of-the-art baselines in terms of Hits@K and Map@K metrics. Compared with the second-best models MSHGAT and MINDS, FFIDP does not use a series of sub-graphs to obtain local and global information, but instead constructs a global diffusion graph to learn the deep relationship between cascades and users from a frequency domain perspective, and uses the attention mechanism to effectively fuse features.

[0114] FFIDP can effectively mine social homogeneity and social influence factors from cascade information, thereby more accurately predicting future target users. In general, FOREST and Inf-VAE graph-based methods are better than NDM based on CNN and attention mechanisms. Learning user representations from diffusion graphs can more effectively improve prediction accuracy. Models that only focus on static user relationships or focus on modeling propagation sequence information do not perform as well as models such as DyHGCN and MINDS that simultaneously model social relationships and information cascades.

[0115] The experimental results of FFIDP verify the effectiveness of the proposed representation model in the diffusion prediction task. Compared with the method based on time domain modeling, the method of fusing cascaded frequency domain features shows higher accuracy. In Douban, Android and Memetracker datasets, our model achieves an average performance improvement of more than 12% and 5% on the Map@K and Hits@K indicators, respectively. In particular, in the Douban dataset, FFIDP improves the Map@10 score by 23%. For the Memetracker dataset, despite the lack of social graphs, FFIDP still outperforms previous models by modeling the diffusion graph, demonstrating its flexibility. In the Christianity dataset, although FFIDP is 2% lower than MINDS in the Map@100 indicator, it performs more competitively in other indicators. Overall, FFIDP consistently outperforms all baseline models, showing its significant advantage in information diffusion prediction.

[0116] Table 2 Prediction scores of Map@K on four real datasets (%)

[0117]

[0118] Table 3 Prediction scores of Hits@K on four real datasets (%)

[0119]

[0120] 2.3 Ablation Experiment

[0121] In order to verify the effectiveness of each module, ablation experiments are conducted on four datasets. Several variants of the model are as follows:

[0122] 1) w / o GF removes the social graph and learns user representation only on the diffusion hypergraph, no longer considering the impact of social homogeneity on users.

[0123] 2) w / o GD removes the diffusion hypergraph and learns user representation only on the friendship graph, no longer considering the impact of social influence on users.

[0124] 3) w / o FFT removes the Fourier transform. After the diffusion graph passes through the hypergraph convolution, it no longer undergoes Fourier transform and directly performs cascade internal context information learning.

[0125] 4) w / oAF removes the attention fusion mechanism and uses the gated fusion mechanism to fuse the two types of user features.

[0126] The results of the ablation experiments are shown in Tables 4 and 5. It can be seen that after removing the social graph or diffusion hypergraph, the model performance is significantly reduced, proving the fundamental role of the two in modeling user dependency. For the Christianity and Android datasets, the score of modeling only the diffusion hypergraph is higher than that of modeling only the social connection, which shows that diffusion dependency has a more significant effect than social connection. The Douban dataset is the opposite, because in the Douban dataset, two people are considered friends only when they interact more than 20 times, so the social connection of this dataset contains more information. Since the Memetracker dataset does not have a social graph, the user representation is obtained by random initialization after removing the diffusion hypergraph. When the social graph and diffusion hypergraph are used at the same time, but the Fourier transform is no longer used, the performance of the model is not as good as the original model, because the matching degree between the user and the cascade will affect the path of information propagation. Not using the Fourier transform will limit the model's ability to separate short-term fluctuations and long-term trends, filter noise, and extract features, resulting in reduced model performance. Compared with the gated fusion mechanism, the attention fusion strategy captures complex dependencies through dynamic weight allocation to achieve refined information fusion and other advantages, which verifies the idea proposed in the present invention: the latter will lead to information redundancy and loss of important features, while the former effectively avoids this problem.

[0127] Table 4 Ablation experiments on four datasets (Map@K)

[0128]

[0129] Table 5 Ablation experiments on four datasets (Hits@K)

[0130]

[0131] In order to further verify the effectiveness of Fourier transform, the present invention also conducts experiments on high-frequency and low-frequency information respectively. The results are as follows Figure 2 As shown. w / o FFT means not using Fourier transform, w / o LP means not using low-frequency information, w / o HP means not using high-frequency information, and wFFT means using Fourier transform. It is not difficult to see that the lack of any information in Fourier transform will lead to a decline in model performance. Low-frequency information and high-frequency information complement each other and jointly affect a cascade propagation path. When low-frequency information is discarded on the Memetracker dataset, the model performance drops sharply. This may be because the high-frequency information contains a lot of noise, which affects the prediction accuracy.

[0132] Overall, the results of the ablation study demonstrate the importance of each component. By integrating these components, FFIDP can achieve the best experimental results.

[0133] 2.4 Hyperparameter Sensitivity Experiment

[0134] In this section, we further conduct sensitivity analysis experiments on key parameters on four datasets to determine their impact on the prediction performance.

[0135] 2.4.1. Impact of Embedding Dimension

[0136] This paper studies the impact of embedding dimension on model performance in the dimensions of {8, 16, 32, 64, 128, 256}. The results are as follows: Figure 3 As shown. The results show that compared with the other two baseline models in terms of Hits@K scores, the characterization method proposed in the present invention has better characterization capabilities in different dimensions. And the score of FFIDP on the Douban dataset continues to increase with the increase of dimension, because larger dimensions may more accurately represent complex diffusion relationships. However, when the dimension exceeds 128, the performance of MSHGAT and MINDS reaches saturation or even overfits, resulting in a decrease in model performance, which shows that the model proposed in the present invention has stronger learning ability. However, for the sake of fairness, all model experiments are compared using a unified dimension (64 dimensions).

[0137] 2.4.2 Impact of the Maximum Cascade Length

[0138] In order to explore the impact of cascade length on performance, we conducted experiments under the maximum cascade lengths of {50, 100, 200, 300}. The results are as follows Figure 4 As shown in the figure, regardless of the cascade length, FFIDP always outperforms other models. Since the increase of cascade length will increase the difficulty of prediction, the performance of the model usually decreases with the increase of cascade length. However, FFIDP is stable after the performance drops to a certain level, indicating that the model has strong robustness.

[0139] 2.4.3 Impact of training set ratio

[0140] The number of training sets may affect the performance of the model. Therefore, this section further explores how they affect the prediction performance when the training set ratio varies in the range of {0.2, 0.4, 0.5, 0.6, 0.8}, where 0.2 means using 20% ​​of the dataset for training and keeping the validation set and test set unchanged. The experimental results on the four datasets are shown in Figure 5 As shown in the figure, Hits@K is selected as the evaluation indicator. With the increase of training data, the performance of FFIDP on the Android, Memetracker and Christian datasets gradually improves until it stabilizes, and only a small amount of training data is needed to achieve good performance; for the Douban dataset, 80% of the data is needed for training before the performance is significantly improved.

[0141] Compared with related technologies, the information microscopic diffusion prediction model integrating cascaded frequency domain features provided by the present invention has the following beneficial effects:

[0142] This paper proposes a method to model information propagation in the frequency domain, which has a unique spectrum representation and fusion method, which can effectively capture the frequency and frequency domain intensity of feature changes, which is crucial for FFIDP to correctly learn cascade features. Extensive experiments show that the proposed method achieves advanced performance. Our future research includes exploring frequency-based insights and mechanisms, and further studying the effectiveness and interpretability of spectrum in information cascade prediction.

[0143] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An information microscopic diffusion prediction model integrating cascaded frequency domain features, characterized in that: include: User social homogeneity learning module, used to learn users’ social relationships through graph convolutional networks; A frequency domain learning module is used to obtain interaction-based user embeddings via diffusion hypergraphs and then learn concatenated frequency features in the Fourier domain; The decoupled information fusion module is used to learn the interactive features within the cascade, fuse the user embeddings at two levels, and retain the effective information to the maximum extent; Cascade propagation prediction module is used to calculate the probability of candidate nodes being infected.

2. The information microscopic diffusion prediction model integrating cascaded frequency domain features according to claim 1 is characterized in that: The learning of the social homogeneity of users comprises the following steps: Build a social relationship graph based on the user's social relationships; A two-layer graph convolutional network is used to learn user social homogeneity on the social relationship graph and obtain the social homogeneity representation of all users.

3. The information microscopic diffusion prediction model integrating cascaded frequency domain features according to claim 1 is characterized in that: The frequency domain learning module includes: Diffused hypergraph convolution, obtaining global information through the hypergraph attention network; User and cascade matching learning, the user representation matrix is ​​mapped to the frequency domain through Fourier transform.

4. The information microscopic diffusion prediction model integrating cascaded frequency domain features according to claim 1 is characterized in that: The decoupling information fusion module includes: Interactive learning within the cascade, using a multi-head self-attention module to explore local diffusion dependencies within the cascade; The attention fusion module retains effective information by calculating the correlation score of two vectors.

5. The information microscopic diffusion prediction model integrating cascaded frequency domain features according to claim 1 is characterized in that: The cascade propagation prediction module comprises: Calculate the probability of the target user participating in the cascade; The cross entropy loss function is used as a constraint.

Images