Information diffusion prediction method based on cascading retrieval context learning
Patent Information
- Application Number
- CN202410585053.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-13
- Publication Date
- 2026-10-20
- Estimated Expiration
- 2044-05-13
AI Technical Summary
[0006]然而，目前的工作只能捕获有限的上下文信息，例如单个级联内的结构和时间模式，或通过各种嵌入技术(例如VAE、Hypergraph)改进用户表示
[0046]1、本发明探索了基于动态级联提示的级联信息扩散预测方法，并通过检索增强语境学习来获取级联间复杂的相互依赖关系，提升信息扩散预测效果；
Smart Images
Figure CN118485106B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of information dissemination technology, and relates to information diffusion and social network analysis based on deep learning, particularly to an information diffusion prediction technology based on retrieval and in-context learning. Background Technology
[0002] The rise of online social media, such as Twitter and Weibo, has enriched personal real-time social experiences and enabled users to access information that interests them. This information (such as microblogs and news) then spreads further through social networks, stimulating user engagement (such as comments, likes, and reposts) and creating information cascades. The goal of information diffusion prediction is to anticipate future users who may be interested in specific information by analyzing the cascading order of currently active users and other relevant information (such as social networks). This has significant economic value for understanding the information dissemination process and has attracted considerable interest from academia and industry. Information diffusion prediction plays a crucial role in its downstream applications, such as social recommendation, rumor detection, and popularity prediction. Due to the complexity of social relationships and the influence of time, the information dissemination process exhibits dynamic temporal changes, making accurate prediction of this process extremely challenging. Information diffusion prediction methods can be mainly divided into three categories: probability diffusion-based methods, feature engineering-based methods, and deep learning methods.
[0003] Information diffusion prediction methods based on probability diffusion assume that each user has an independent probability of information diffusion. However, this approach oversimplifies the complexity of the information dissemination process and ignores the diversity and interests among users, thus its performance on real-world datasets is not ideal.
[0004] Information diffusion prediction methods based on feature engineering predict information by studying and designing features such as cascading structures, time, user attributes, and information content. However, this method relies heavily on manually constructed features, and high-quality features require specialized domain knowledge and incur additional computational costs. Therefore, this method is not easily promoted and applied in new fields.
[0005] Deep learning-based information diffusion prediction is a comprehensive framework that learns user dynamics and related knowledge by modeling cascading sequences of active users. This method uses sequence models such as recurrent neural networks (RNNs), long short-term memory neural networks (LSTMs), and gated recurrent units (GRUs) to map the diffusion process onto social relationships and employs specific mechanisms such as gating and attention to model the cascading sequences. Furthermore, this method utilizes graph-based models to construct various local and global graphs to learn the diffusion structure.
[0006] However, current work can only capture limited contextual information, such as structure and temporal patterns within a single cascade, or improve user representations through various embedding techniques (e.g., VAEs, Hypergraphs). While these methods have greatly improved prediction accuracy, several key challenges remain unresolved: (1) Current methods often ignore the complex interdependencies between cascades. (2) They do not account for performance differences between different platforms; that is, models optimized for a specific social media platform often show significant performance degradation on datasets from other platforms, even if the diffusion process is essentially similar. Summary of the Invention
[0007] The purpose of this invention is to address the aforementioned technical problems in the existing technology by proposing an information diffusion prediction method based on cascade retrieval context learning. This method obtains the complex interdependencies between cascades through retrieval-enhanced context learning and embeds additional cue enhancement strategies and social relationship enhancements to improve prediction performance.
[0008] The invention's approach is as follows: First, a cascaded cue pool is constructed from the cascaded information sequences. Several similar cascades are retrieved as in-context learning (ICL) cues for each cascade. Then, a graph attention network is used to obtain the embedding representation of the information cascaded social network graph, thereby obtaining the dynamic embedding representation of the user. Combined with context learning, the cues are input into a pre-trained transformer to obtain the implicit representation of the cascades. A multilayer perceptron is then used to predict the next infected user. Finally, the backpropagation algorithm is used to optimize the model's parameters, ultimately achieving effective model prediction.
[0009] Based on the above-mentioned inventive concept, the information diffusion prediction method based on cascaded retrieval context learning provided by this invention includes the following steps:
[0010] S1 constructs a prompt pool based on the cascaded sequence of historical information, enhances the prompt pool, and obtains the user embedding of the prompt pool.
[0011] S2 retrieves several hints from the hint pool that match the query concatenation;
[0012] S3 obtains the query concatenation and several suggestion user embeddings based on the user embeddings in the suggestion pool; then generates the concatenation representation through a pre-trained transformer network; and finally fuses the concatenation representation with the user embeddings of the query concatenation to obtain the final representation.
[0013] S4 generates user prediction results based on the final representation and the user embedding of the cue pool, thus completing the information diffusion prediction.
[0014] The purpose of step S1 above is to construct a cue pool to capture cross-cascade dependencies, facilitating the retrieval of several similar cascades as prompts for each cascade to be retrieved using the retrieval algorithm. This allows the model to effectively utilize relevant past experience. This step also utilizes two cue enhancement strategies and a social relationship enhancement embedding method to enrich the context, thereby constraining the model to enhance information samples. Two enhancement strategies are used here to introduce noise into the prompts, aiming for effective model training. Specifically, these strategies are applied after constructing the cue pool and are divided into user masking and user reordering.
[0015] Step S1 includes the following sub-steps:
[0016] S11 divides each cascade in the historical information cascade into several sub-cascades;
[0017] S12 performs random masking and / or random reordering on each sub-cascade to generate user enhancement prompts; the user enhancement prompts and the original sub-cascades constitute the final prompt pool;
[0018] S13 is a multi-layer graph attention network based on a multi-head attention mechanism. It encodes the social relationship network formed by users in the cue pool to obtain the user embedding of the cue pool.
[0019] In step S11 above, the specific implementation method is as follows: When the total length is... Historical cascade In the middle, use a sliding window to display each cascading line. Divide into N segments of length W i Each child cascade. Then, all child cascades are placed into a cue pool. In the middle, the total size of the prompt pool is Each of them
[0020] In step S12 above, the enhanced user prompts include enhanced prompts for user blocking and enhanced prompts for user reordering.
[0021] For each sub-cascade, the random masking is implemented as follows: For each Random blocking For each user, γ represents the masking rate; each masked element is replaced with a special marker "[mask]", and the generated user masking enhancement prompt is represented as follows:
[0022] For each sub-cascade, the reordering is implemented as follows: to reduce the model's dependence on the user's order and enhance its robustness to new interactions, the position of p is changed randomly. j The order of consecutive subsequences, the length of which is Where β is the reordering rate; the generated user reordering enhancement prompt can be represented as:
[0023] The final cue pool is a collection of the original sub-cascades, user-masked enhanced cues, and user-reordered enhanced cues, i.e.
[0024] In step S13 above, user embeddings are generated using user dynamic embeddings with enhanced social relationships, assuming there is a trainable embedding matrix. Where N represents the total number of users in the prompt pool dataset, d E It is an adjustable dimension. Then, a multi-layer graph attention network (GAT) with a multi-head attention mechanism is used to encode the social relationship network graph. The user's social relationship embedding (i.e., user embedding) is represented as... Here, W k Denotes a set of independent trainable weight matrices; K represents the number of attention heads; A k Let represent the attention matrix calculated through the self-attention mechanism, and σ(·) represent the activation function; the initial input of GAT is... Repeat the above steps until user representations for all users in the detection pool are generated. Then, E and Connect them to form the final user embedding matrix Right now
[0025] The purpose of step S2 above is to retrieve several suggestion concatenations that match the query concatenations from the suggestion pool. This specifically includes the following sub-steps:
[0026] S21 divides the query cascade into several query sub-cascades, and uses these query sub-cascades to construct a query list;
[0027] S22 retrieves candidate suggestions from the suggestion pool that are identical to at least one user in each query sub-cascade, based on the query list;
[0028] S23 calculates the correlation score between each candidate suggestion and the corresponding query sub-cascade;
[0029] S24 selects several candidate suggestions as suggestions for matching the query sub-cascades, sorted by relevant scores in descending order.
[0030] In step S21 above, similar to historical cascading, the query cascading D is divided into several query sub-cascadings, and a query list is constructed using these sub-cascadings, denoted as... Let q represent the query sub-cascade.
[0031] In step S22 above, from the prompt pool Search and query sub-cascade The most relevant tips. First, use a Boolean query to filter out candidate tips that contain at least one user from the query. Where N can express The number of prompts in the text.
[0032] In step S23 above, the BM25 ranking function is used to calculate each candidate hint. The relevance score R relative to the query sub-cascade q is:
[0033]
[0034] Where k1 is a free parameter. This indicates a query for users in the child cascade q. exist Word frequency in.
[0035] IDF can be viewed as a penalty factor that reduces the importance of common users relative to rare users, meaning that matching with rare users will produce a stronger similarity signal than matching with common users. The IDF term is defined as:
[0036]
[0037] in, Represent a function used to count users. The number of times it appears in the candidate suggestions.
[0038] In step S24 above, after ranking the candidate prompts using BM25, the top K prompts are selected. As a hint for matching query sub-cascades, final context learning is performed. This is after retrieving the query list. After considering all query sub-cascades, a suggestion set was obtained. It includes A tip.
[0039] The purpose of step S3 above is to obtain the user embeddings of the query concatenation and several prompts, and to generate implicit representations through a pre-trained transformer network for prediction, thereby improving robustness across different datasets. Step S3 includes the following sub-steps:
[0040] S31 is based on user index u i =LookUp(u i The query list is found by searching the user embedding matrix U. The query user embedded for each user and hint set Each user's corresponding prompt is embedded in the user profile.
[0041] S32 will embed the query user and prompt users to embed Connect them to get the total user embedding
[0042] S33 embeds the total user data. Input a pre-trained transformer network (i.e., freeze the network parameters in the transformer network), and the output is the cascaded representation.
[0043] S34 will cascade representation Compressed into shape V cp Then embed it with the query user Connecting these layers and passing them through a linear layer with a softmax function, we obtain the final representation.
[0044] In step S4 above, the probability of all users being in q is determined by... Calculate, where m mask Used to block users who have already been activated.
[0045] Compared with existing technologies, the present invention has the following advantages:
[0046] 1. This invention explores a cascading information diffusion prediction method based on dynamic cascading prompts, and improves the information diffusion prediction effect by retrieving and enhancing contextual learning to obtain complex interdependencies between cascades.
[0047] 2. This invention introduces additional cue enhancement strategies and social relationship enhancement embedding information diffusion prediction methods into the model training by adding noise and enriching user features to obtain fine-grained aligned user embedding representations. Attached Figure Description
[0048] Figure 1 This is a schematic diagram of the information diffusion prediction process based on cascading retrieval context learning. Detailed Implementation
[0049] The present invention will be further described with reference to the accompanying drawings.
[0050] Example
[0051] This invention aims to address the problems existing in the prior art and proposes a framework for an information diffusion prediction method based on cascaded retrieval context learning. First, we construct a cascaded cue pool from the information cascade sequence to retrieve several similar cascades as in-context learning (ICL) cues for each cascade. Next, we utilize a graph attention network to obtain the embedded representation of the information cascade social network graph, thereby obtaining the dynamic embedded representation of the user. Combining context learning, the cues are input into a pre-trained transformer to obtain the implicit representation of the cascade, and then a multilayer perceptron is used to predict the next infected user. Finally, the backpropagation algorithm is used to optimize the model parameters to achieve effective model prediction.
[0052] Based on the above analysis, such as Figure 1 As shown, the information diffusion prediction method based on cascading retrieval context learning provided in this embodiment includes the following steps:
[0053] S1 constructs a prompt pool based on the cascaded sequence of historical information, enhances the prompt pool, and obtains the user embedding of the prompt pool.
[0054] The purpose of this step is to construct a cue pool to capture cross-cascade dependencies, enabling the retrieval algorithm to retrieve several similar cascades as prompts for each cascade to be retrieved, allowing the model to effectively utilize relevant past experience. This step also utilizes two cue enhancement strategies and a social relation enhancement embedding method to enrich the context, thereby constraining the model to enhance information samples.
[0055] This process includes the following sub-steps:
[0056] S11 divides each cascade in the historical information cascade into several sub-cascades.
[0057] The specific implementation method is as follows: In a total length of... Historical cascade In the middle, use a sliding window to display each cascading line. Divide into N segments of length W i Each child cascade. Then, all child cascades are placed into a cue pool. In the middle, the total size of the prompt pool is Each of them in Indicates sub-cascade p j The i-th user in the list.
[0058] S12 performs random masking and / or random reordering on each sub-cascade to generate user enhancement prompts; the user enhancement prompts and the original sub-cascades constitute the final prompt pool.
[0059] Two enhancement strategies (user masking and user reordering) are used here to introduce noise into the cues for efficient model training. Specifically, these strategies are applied after the cue pool is built to introduce noise and enhance the model's robustness.
[0060] For each sub-cascade, the random masking is implemented as follows: For each Random blocking For each user, γ represents the masking rate. Each masked element is replaced with a special marker "[mask]", and the generated user masking enhancement prompt is represented as...
[0061] For each sub-cascade, the reordering is implemented as follows: to reduce the model's dependence on the user's order and enhance its robustness to new interactions, the position of p is changed randomly. j The prompt is reordered based on the order of consecutive subsequences. A length of [length missing] is selected. The subsequences are randomly reordered, where β is the reordering rate. The generated user reordering enhancement hint can be represented as:
[0062] Therefore, user enhancement cues include user-masking enhancement cues and user-reordering enhancement cues. The final cue pool is a set of the original sub-cascades, user-masking enhancement cues, and user-reordering enhancement cues, i.e.
[0063] S13 is a multi-layer graph attention network based on a multi-head attention mechanism. It encodes the social relationship network formed by users in the cue pool to obtain the user embedding of the cue pool.
[0064] User embeddings are generated using user dynamic embeddings with enhanced social relationships. First, it is assumed that there is a trainable embedding matrix. Where N represents the total number of users in the prompt pool dataset, d E It is an adjustable dimension. Then, a multi-layer graph attention network (GAT) with a multi-head attention mechanism is used to encode the social relationship network graph. The user's social relationship embedding (i.e., user embedding) is represented as:
[0065]
[0066] Here, W k Denotes a set of independent trainable weight matrices; K represents the number of attention heads; A k Let represent the attention matrix calculated using the self-attention mechanism, and σ(·) represent the activation function. The initial input to GAT is... Repeat the above steps until user representations for all users in the detection pool are generated. Then, E and Connect them to form the final user embedding matrix Right now
[0067] S2 retrieves several hints from the hint pool that match the query cascade.
[0068] The purpose of this step is to retrieve several suggestion concatenations that match the query concatenations from the suggestion pool. This includes the following sub-steps:
[0069] S21 divides the query cascade into several query sub-cascades, and uses these query sub-cascades to construct a query list.
[0070] Similar to historical cascading, the query cascading D is divided into several query sub-cascadings, and a query list is constructed using these sub-cascadings, denoted as . Let q represent the query sub-cascade.
[0071] S22 retrieves candidate suggestions from the suggestion pool that are identical to at least one user in each query sub-cascade, based on the query list.
[0072] From the cue pool Search and query sub-cascade The most relevant tips. First, use a Boolean query to filter out candidate tips that contain at least one user from the query. Where N can express The number of prompts in the text.
[0073] S23 calculates the correlation score between each candidate suggestion and the corresponding query sub-cascade.
[0074] Each candidate suggestion is calculated using the BM25 ranking function. The relevance score R relative to the query sub-cascade q is:
[0075]
[0076] Where k1 is a free parameter. This indicates a query for users in the child cascade q. exist Word frequency in.
[0077] IDF can be viewed as a penalty factor that reduces the importance of common users relative to rare users, meaning that matching with rare users will produce a stronger similarity signal than matching with common users. The IDF term is defined as:
[0078]
[0079] in, Represent a function used to count users. The number of times it appears in the candidate suggestions.
[0080] S24 selects several candidate suggestions as suggestions for matching the query sub-cascades, sorted by relevant scores in descending order.
[0081] After ranking the candidate suggestions using BM25, select the top K suggestions. As a final contextual learning prompt. After searching the query list. After considering all query sub-cascades, a suggestion set was obtained. It includes A tip.
[0082] S3 obtains the user embeddings of the query concatenation and several prompts based on the user embeddings of the prompt pool; then generates the concatenation representation through a pre-trained transformer network; and finally fuses the concatenation representation with the user embeddings of the query concatenation to obtain the final representation.
[0083] The purpose of this step is to obtain user embeddings of query concatenation and several hints, and to generate implicit representations through a pre-trained transformer network for prediction, thereby improving robustness across different datasets.
[0084] This process includes the following sub-steps:
[0085] S31 is based on user index u i =LookUp(u i The query list is found by searching the user embedding matrix U. The query user embedded for each user and hint set Each user's corresponding prompt is embedded in the user profile.
[0086] S32 will embed the query user and prompt users to embed Connect them to get the total user embedding
[0087] S33 embeds the total user data. Input a pre-trained transformer network (i.e., freeze the network parameters in the transformer network), and the output is the cascaded representation.
[0088] S34 will cascade representation Compressed into shape V cp Then embed it with the query user Connecting these layers and passing them through a linear layer with a softmax function (in this embodiment, a multilayer perceptron with a softmax function is used) yields the final representation. W and b represent the parameter matrix and bias, respectively.
[0089] S4 generates user prediction results based on the final representation and the user embedding of the cue pool, thus completing the information diffusion prediction.
[0090] In this step, the probability of all users being in q is determined by... Calculate, where M mask Used to block users who have already been activated. The user with the highest probability is used as the next user to be predicted.
[0091] Application examples
[0092] The information diffusion prediction method based on cascaded retrieval context learning provided in the examples was applied to three different real datasets (Twitter, Douban, and Android). The first dataset is from reference [Hodas NO, Lerman K. The simple rules of social contagion[J]. Scientific reports, 2014, 4(1): 4343.], the second dataset is from reference [Zhong E, Fan W, Wang J, et al. Comsoc: adaptive transfer of user behaviors over composite social network[C] / / Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and datamining. 2012: 696-704.], and the third dataset is from reference [Sankar A, Zhang X, Krishnan A, et al. Inf-VAE: A variational autoencoder framework to integrate homophily and influence in diffusion prediction[C] / / Proceedings of the 13th international conference on web search and data]. The prediction performance on mining.2020:510-518. In the three datasets, the ratio of sample data in the training set, validation set, and test set is 8:1:1.
[0093] The information diffusion prediction method based on cascaded retrieval context learning provided in the above embodiments uses models including a Graph Attention Network (GAT), a Transformer network, and a multilayer perceptron. The GAT is trained using a social relationship network formed by users in the cue pool. The Transformer network is pre-trained, and its parameters are frozen. Based on these two trained networks, the model is trained and validated using the training and validation sets according to steps S1-S4 given in the embodiments to obtain predicted values. The parameters of the multilayer perceptron in the model are optimized using the backpropagation algorithm to minimize the true value y. q and predicted value The model utilizes cross-entropy loss to improve its predictive performance and robustness. The model is then tested on a test set.
[0094] Meanwhile, the information diffusion prediction method based on cascaded retrieval context learning provided by this invention is compared with seven different baseline models (DeepDiffuse, Topo-LSTM, NDM, RotDiff, DisenIDP, Inf-VAE, MS-HGAT). In this application example, Hits@100 and MAP@100 are used as evaluation indicators (the larger the value, the better the prediction effect). The prediction results are shown in Table 1.
[0095] Table 1: Performance of popularity prediction on the application example dataset
[0096]
[0097] The remaining methods in the table are described below:
[0098] Topo-LSTM: This method extends the standard LSTM model to dynamic DAG-LSTM to model the local propagation structure of dynamic directed acyclic graphs (DAGs) and perform information diffusion prediction. The implementation process is referenced in [Wang J, Zheng VW, Liu Z, et al. Topological recurrent neural network for diffusion prediction[C] / / 2017IEEE international conference on data mining(ICDM).IEEE,2017:475-484].
[0099] DeepDiffuse: This method combines time-point processes and LSTM models to model temporal information and user sequences for diffusion prediction. It also employs a Transformer module, which enhances the features of activated users learned by the LSTM model. The implementation process is referenced in [Islam MR, Muthiah S, Adhikari B, et al. Deepdiffuse: Predicting the 'who' and 'when' in cascades[C] / / 2018IEEE international conference on data mining(ICDM).IEEE,2018:1055-1060].
[0100] NDM: This method models the cascade diffusion process through self-attention mechanism and convolutional neural network, and reduces long-term dependence through relaxed independence assumption. Its implementation process is referenced in [Yang C, Sun M, Liu H, et al. Neural diffusion model for microscopic cascade study[J].IEEE Transactions on Knowledge and Data Engineering,2019,33(3):1128-1139].
[0101] Inf-VAE: This method designs a variational framework that combines a variational graph autoencoder and a collaborative attention network to model social relationships and sequence information. It considers both social homogeneity and temporal influence to predict the set of all active users. The implementation process is referenced in [Sankar A, Zhang X, Krishnan A, et al. Inf-VAE: A variational autoencoder framework to integrate homophily and influence indiffusion prediction[C] / / Proceedings of the 13th international conference on web search and data mining.2020:510-518].
[0102] MS-HGAT: This method designs a hypergraph attention network to obtain cascaded structure information from sequential hypergraphs, and then uses a gating fusion strategy to fuse hypergraph structure information and static social structure information to improve prediction performance. The implementation process is referenced in [Sun L, Rao Y, Zhang X, et al. MS-HGAT: memory-enhanced sequential hypergraph attention network for information diffusion prediction[C] / / Proceedings of the AAAI Conference on Artificial Intelligence.2022,36(4):4156-4164].
[0103] RotDiff: This method proposes a hyperbolic rotation representation model that models cascading diffusion patterns by performing rotational transformations in hyperbolic space. It can better describe the strength of user influence and the asymmetric characteristics in the diffusion process. The implementation process is referenced in [Qiao H, Feng S, Li X, et al. RotDiff: A Hyperbolic Rotation Representation Model for Information Diffusion Prediction[C] / / Proceedings of the 32nd ACM International Conference on Information and Knowledge Management.2023:2065-2074].
[0104] DisenIDP: This method proposes a self-supervised decoupling framework (DisenIDP), which improves the problems of user intent mixing and dynamic drift of cascade time effects by constructing an intent-aware hypergraph and extracting long-term and short-term cascade effects. The implementation process is referenced in [Cheng Z, Ye W, Liu L, et al. Enhancing Information Diffusion Prediction with Self-Supervised Disentangled User and Cascade Representations[C] / / Proceedings of the 32nd ACM International Conference on Information and Knowledge Management.2023:3808-3812].
[0105] As can be seen from the experimental results in Table 1, the information diffusion prediction method based on cascaded retrieval context learning provided by this invention can significantly improve the accuracy of predicting the next active user compared to other baseline models.
[0106] Therefore, the information diffusion prediction method based on cascading retrieval context learning proposed in this invention not only fully utilizes enhanced cascading retrieval prompts and pre-trained transformers, but also achieves accurate predictions through a trainable social relationship embedding module and additional linear layers. Furthermore, introducing noise further enhances the model's robustness. Extensive experiments were conducted on three real-world datasets, demonstrating the effectiveness and robustness of the proposed method. The performance improvements of the proposed scheme indicate that combining the advantages of context learning with the information diffusion prediction task is an effective way to simulate the cascading diffusion process of information and improve the prediction accuracy for the next activated user.
[0107] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.
Claims
1. An information diffusion prediction method based on cascading retrieval context learning, characterized in that, Includes the following steps: S1 constructs a cue pool based on the cascaded sequence of historical information, enhances the cue pool, and obtains the user embedding of the cue pool; this step includes the following sub-steps: S11 divides each concatenation in the historical information concatenation into several sub-concatenations; specifically, it is implemented by: […]. Historical cascade In the middle, use a sliding window to display each cascading line. Cut into lengths of of Individual cascades; Then put all the child cascades into a cue pool. In the middle, the total size of the prompt pool is Each of them ; S12 performs random masking and / or random reordering on each sub-cascade to generate user enhancement prompts; the user enhancement prompts and the original sub-cascades constitute the final prompt pool; S13 uses a multi-head attention mechanism-based multi-layer graph attention network to encode the social relationship network formed by users in the cue pool, thus obtaining the user embedding of the cue pool. S2 retrieves several hints from the hint pool that match the query concatenation; S3 obtains the query concatenation and several suggestion user embeddings based on the user embeddings in the suggestion pool; then generates the concatenation representation through a pre-trained transformer network; and finally fuses the concatenation representation with the user embeddings of the query concatenation to obtain the final representation. S4 generates user prediction results based on the final representation and the user embedding of the cue pool, thus completing the information diffusion prediction.
2. The information diffusion prediction method based on cascading retrieval context learning according to claim 1, characterized in that, In step S12, the enhanced user prompts include enhanced prompts for user blocking and enhanced prompts for user reordering; For each sub-cascade, the random masking is implemented as follows: For each Random blocking Users, among whom The masking rate; each masked element is replaced with a marker "[mask]", and the generated user masking enhancement prompt is represented as follows. ; For each sub-cascade, the reordering is implemented by randomly changing the positions. The order of consecutive subsequences, the length of which is ,in The reordering rate; the generated user reordering enhancement prompt can be represented as: ; The final cue pool is a collection of the original sub-cascades, user-masked enhanced cues, and user-reordered enhanced cues, i.e. .
3. The information diffusion prediction method based on cascading retrieval context learning according to claim 2, characterized in that, In step S13, user embeddings are generated using user dynamic embeddings with enhanced social relationships, assuming there is a trainable embedding matrix. ,in This indicates the total number of users in the suggestion pool dataset. It is an adjustable dimension; then a multi-layer graph attention network with a multi-head attention mechanism is used to encode the social relationship network graph; the user's social relationship embedding is represented as... ;here, This represents a set of independent trainable weight matrices; Indicates the number of attention heads; This represents the attention matrix calculated using the self-attention mechanism. This represents the activation function; the initial input to GAT is... Repeat the above steps until user representations for all users in the detection pool are generated. ; then, and Connect them to form the final user embedding matrix ,Right now = .
4. The information diffusion prediction method based on cascading retrieval context learning according to any one of claims 1 to 3, characterized in that, Step S2 specifically includes the following sub-steps: S21 splits the query concatenation into several query sub-concatenations, and uses these query sub-concatenations to construct a query list; S22 Based on the query list, retrieve candidate suggestions from the suggestion pool that are identical to each query sub-cascade containing at least one user; S23 Calculate the correlation score between each candidate suggestion and the corresponding query sub-cascade; S24 Select several candidate suggestions as suggestions for matching the query sub-cascades, sorted in descending order of relevant scores.
5. The information diffusion prediction method based on cascading retrieval context learning according to claim 4, characterized in that, In step S21, the query cascades. Divide the query into several sub-cascades, and construct a query list using these sub-cascades, denoted as . ;by Indicates query for sub-cascades. .
6. The information diffusion prediction method based on cascading retrieval context learning according to claim 5, characterized in that, In step S23, the BM25 ranking function is used to calculate each candidate hint. Relative to query sub-cascade Relevant scores ,Right now: ； in, These are free parameters. Indicates query sub-cascade users in exist Word frequency in; The IDF term is defined as: ； in, Represent a function used to count users. The number of times it appears in the candidate suggestions; Indicates candidate suggestions The number of prompts in the text; In step S24, after ranking the candidate suggestions using BM25, the top K suggestions are selected. As a hint for matching query sub-cascades; after retrieving the query list. After considering all query sub-cascades, a suggestion set was obtained. .
7. The information diffusion prediction method based on cascading retrieval context learning according to claim 4, characterized in that, Step S3 includes the following sub-steps: S31 Based on User Index By finding the user embedding matrix Find the query list The query user embedded for each user and hint set Each user's corresponding prompt is embedded in the user profile. ; S32 embeds the query user and prompt users to embed Connect them to get the total user embedding ; S33 embeds the total user The input to a pre-trained transformer network is the output of the cascaded representation. ; S34 represents the cascaded representation. Compressed into shape of Then embed it with the query user Connecting these layers and passing them through a linear layer with a softmax function, we obtain the final representation. .
8. The information diffusion prediction method based on cascading retrieval context learning according to claim 7, characterized in that, In step S4, all users... The probability of being in is determined by Calculation, where Used to block users who have already been activated.