Information cascade prediction method and apparatus based on spatiotemporal characteristics and content preferences

Abstract

This invention provides a method and apparatus for information cascading prediction based on spatiotemporal features and content preferences. The method includes: acquiring a user's social relationship data and using a GCN model to generate social relationship embedding features; then using a GRU model to construct spatiotemporal features of the user, and performing information cascading feature construction on these features to obtain spatiotemporal cascading features; acquiring the information dissemination content that the user will participate in, constructing multimodal joint features based on the information dissemination content, and determining interest influence factors. Finally, determining the dot product similarity between the user's spatiotemporal cascading features and the interest influence factors, and predicting the user's participation probability for the information dissemination content based on the dot product similarity. This application solves the technical problem of poor prediction performance in information cascading prediction tasks in the prior art.

Description

Technical Field

[0001] This invention relates to the field of information dissemination technology, and in particular to an information cascading prediction method and apparatus based on spatiotemporal characteristics and content preferences. Background Technology

[0002] As social networks have become increasingly dominant on the internet, the way information is disseminated has changed accordingly. Understanding how information spreads on social network platforms can help with applications such as information recommendation and public opinion control. The process of information diffusion is also known as cascading. Cascading mainly records relevant information about when information is transmitted from the communicator to the receiver. Information cascading prediction tasks aim to predict key attributes and characteristics of the future by observing early cascading events, thereby determining the likelihood of participating in information dissemination.

[0003] For information cascading prediction tasks, recent research has proposed neural network-based prediction methods. These methods model the information cascading prediction task by considering the social relationships between users, and simultaneously learn the structural features of social relationship graphs and dynamic diffusion graphs to improve the accuracy of information cascading prediction. However, neural network-based prediction methods face the following challenges: First, there is the problem of modeling user relationships. Since information in social networks is generated and spread based on user interaction behaviors, users are the decisive factor in the information dissemination process. Users establish follow relationships, share content, and forward content, enabling information to flow through user interactions. Constructing complex relationships between users is quite difficult. Second, there is the problem of modeling the spatiotemporal features of the cascade. The cascade structure changes over time, and considering the dynamic features of the cascade during the modeling process is also a challenge. Third, there is the problem of modeling information content features. The dissemination process of information is also affected by the content of the information itself. It is necessary to combine information content in information cascading prediction tasks. However, the information dissemination content in current social networks includes multimodal data, including text and images. These three challenges affect the prediction accuracy of information cascading prediction tasks, resulting in poor prediction results. Summary of the Invention

[0004] This invention provides a method and apparatus for information cascading prediction based on spatiotemporal features and content preferences. On the one hand, it combines user social relationships to model and construct user spatiotemporal cascading features. On the other hand, it constructs user interest and preference features based on the information dissemination content. Combining these two features, it performs information cascading prediction, solving the technical problem of low prediction accuracy and poor prediction results in existing information cascading prediction tasks.

[0005] This invention provides an information cascading prediction method based on spatiotemporal features and content preferences, comprising the following specific steps:

[0006] The user's social relationship data is obtained, and the GCN model is called to generate the social relationship embedding features of the user.

[0007] Based on the social relationship embedding features, the GRU model is invoked to construct user spatiotemporal features, and the user spatiotemporal features are then used to construct features based on information cascading to obtain user spatiotemporal cascaded features.

[0008] The system obtains the information dissemination content that users will participate in, constructs multimodal joint features based on the information dissemination content, and determines interest influence factors based on the multimodal joint features, wherein the interest influence factors are used to characterize the degree of user content preference for the information dissemination content;

[0009] Determine the dot product similarity between the user's spatiotemporal cascaded features and the interest influencing factors, and predict the user's participation probability for the information dissemination content based on the dot product similarity.

[0010] In some embodiments, the step of constructing user spatiotemporal features based on the social relationship embedding features by calling the GRU model includes:

[0011] Obtain user historical cascaded data and divide the user historical cascaded data into multiple cascaded subgraphs according to timestamps;

[0012] The social relationship embedding features are aggregated with the cascaded subgraph to obtain the cascaded subgraph embedding features corresponding to each timestamp;

[0013] Determine the temporal embedding feature corresponding to each timestamp;

[0014] The cascaded subgraph embedding features and the temporal embedding features are input into the GCN model for spatial feature extraction to obtain a spatial feature sequence.

[0015] The GRU model is invoked to extract features from the spatial feature sequence to obtain the user's spatiotemporal features.

[0016] In some embodiments, the step of constructing user spatiotemporal concatenated features based on information concatenation of the user spatiotemporal features includes:

[0017] The social relationship embedding features are gated and fused with the user spatiotemporal features to obtain a fused spatiotemporal feature sequence.

[0018] Determine the position embedding code corresponding to each sequence element in the fused spatiotemporal feature sequence, and concatenate the position embedding code with the sequence element to obtain the spatiotemporal concatenated feature sequence;

[0019] The spatiotemporal cascaded feature sequence is input into the Transformer model for mask attention mechanism processing to obtain the user spatiotemporal cascaded features.

[0020] In some embodiments, the information dissemination content includes visual images and corresponding text fragments, and the construction of multimodal joint features based on the information dissemination content includes:

[0021] The CLIP model's image encoder is invoked to encode each visual image, thereby obtaining initial image features;

[0022] The initial image features are input into the Transformer model for embedding encoding to obtain the image encoding features of the visual image;

[0023] The CLIP model's text encoder is invoked to encode the text segment corresponding to each visual image, obtaining text encoder features, and the text encoding features are merged to obtain the overall text encoding features;

[0024] The image coding features and the overall text coding features are gated and fused to obtain multimodal joint features.

[0025] In some embodiments, the step of inputting the initial image features into a Transformer model for embedding encoding to obtain the embedding encoded features of the visual image includes:

[0026] Each of the initial image features is concatenated as a sequence element to obtain an embedding sequence, and the CLS tag is inserted as the first sequence element into the head of the embedding sequence to obtain an image embedding sequence;

[0027] Determine the position code of each sequence element in the image embedding sequence, and concatenate the position code with the sequence element to obtain the concatenated feature;

[0028] Each spliced ​​feature in the image embedding sequence is input into the Transformer model for encoding processing to obtain the embedding encoding features of the visual image.

[0029] In some embodiments, determining the interest influence factor based on the multimodal joint features includes:

[0030] Based on the user's historical interaction data, the user's participation information embedding features are extracted, and the Transformer model is called to encode the participation information embedding features to obtain the user interest embedding features.

[0031] The user interest embedding features are subjected to average pooling to obtain average pooled features;

[0032] The feature similarity between the average pooling feature and the multimodal joint feature is determined, and the feature similarity is used as an interest influence factor.

[0033] In some embodiments, the participation probability is obtained by calling an information cascade prediction model, and the training method of the information cascade prediction model includes:

[0034] The user's social relationship data samples, information dissemination content samples, user historical cascade data samples, and user historical interaction data samples are used as model training samples;

[0035] The training samples of the model are input into the information cascade prediction model for forward propagation to obtain the predicted participation probability of users for the information dissemination content samples;

[0036] Determine the actual participation probability of users in the sample of the information dissemination content;

[0037] Based on the predicted participation probability and the actual participation probability, a cross-entropy loss function is constructed, and the cross-entropy loss function is used to backpropagate in the information cascade prediction model to update the parameters of the information cascade prediction model.

[0038] The present invention also provides an information cascade prediction device based on spatiotemporal features and content preferences, the device specifically including the following modules:

[0039] The acquisition module is used to acquire the user's social relationship data and call the GCN model to generate the social relationship data to obtain the user's social relationship embedding features;

[0040] The construction module is used to construct user spatiotemporal features by calling the GRU model based on the social relationship embedding features, and to construct user spatiotemporal cascaded features based on information concatenation of the user spatiotemporal features.

[0041] The determination module is used to obtain the information dissemination content that the user will participate in, construct multimodal joint features based on the information dissemination content, and determine the interest influence factor according to the multimodal joint features, wherein the interest influence factor is used to characterize the user's content preference for the information dissemination content;

[0042] The prediction module is used to determine the dot product similarity between the user's spatiotemporal cascaded features and the interest influencing factors, and to predict the user's participation probability for the information dissemination content based on the dot product similarity.

[0043] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the information cascading prediction method based on spatiotemporal features and content preferences as described above.

[0044] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the information cascading prediction method based on spatiotemporal features and content preferences as described above.

[0045] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the information cascade prediction method based on spatiotemporal features and content preferences as described above.

[0046] The information cascading prediction method based on spatiotemporal features and content preferences provided by this invention, on the one hand, combines user social relationship data for modeling, and then combines user spatiotemporal features to construct user spatiotemporal cascading features. On the other hand, it constructs multimodal joint features based on information dissemination content, and then determines user interest and preference features. Finally, it combines spatiotemporal cascading features and interest and preference features to perform information cascading prediction, thus solving the technical problem of low prediction accuracy and poor prediction results in existing information cascading prediction tasks. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0048] Figure 1 This is a flowchart illustrating the information cascading prediction method based on spatiotemporal features and content preferences provided by the present invention.

[0049] Figure 2 This is a schematic diagram of the information cascade graph embedding layer based on spatiotemporal features provided by the present invention.

[0050] Figure 3 This is a schematic diagram illustrating the principle of the visual information processing module for extracting image coding features provided by the present invention.

[0051] Figure 4 This is a schematic diagram illustrating the training process of the information cascade prediction model provided by this invention.

[0052] Figure 5 This is a schematic diagram of the structure of the information cascade prediction device based on spatiotemporal features and content preferences provided by the present invention.

[0053] Figure 6 This is a schematic diagram of the physical structure of an electronic device provided by the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0055] The following is combined with Figure 1 This invention describes an information cascading prediction method based on spatiotemporal features and content preferences. Figure 1 This is a flowchart illustrating the information cascading prediction method based on spatiotemporal features and content preferences provided by the present invention, as shown below. Figure 1 As shown, the method includes the following steps.

[0056] In step 101, the user's social relationship data is obtained, and the GCN model is called to generate the social relationship data to obtain the user's social relationship embedding features.

[0057] First, user features need to be extracted, taking into account their social relationships. Each user has different social relationships, and each user has varying preferences for different information content. Therefore, this invention acquires user social relationship data, treats all users as nodes, and uses the social relationships between users as undirected edges based on the social relationship data. This constructs a user social network graph based on the nodes and undirected edges, allowing information content to propagate freely between different user nodes in the graph. Since graph convolutional neural networks (GCNs) are suitable for graph-structured data, this invention uses a GCN model to generate social relationship data, obtaining the user's social relationship embedding features. That is, the features of each node (user) in the user social network graph are extracted using the GCN model to obtain the user's social relationship embedding features, represented as... The extraction process is represented by the following formulas (1) and (2):

[0058] (1)

[0059] (2)

[0060] In the above formulas (1) and (2), This indicates that the node (user) is in the GCN model. Vector representation of the layer (i.e., social relationship embedding features). The input to the first layer of the GCN model is a randomly initializeable feature vector, and N represents the number of nodes (users), and d represents the dimension of the feature vector. The adjacency matrix of the GCN model is derived from the user's social network graph, and E represents the identity matrix. Representing the adjacency matrix The matrix obtained by combining the identity matrix E, express The degree matrix, Degree matrix The matrix obtained by combining the identity matrix E, This represents a trainable weight matrix. This represents the activation function of the network layer in the GCN model.

[0061] By using step 101, the user's social relationship data is used to construct the user's social relationship embedding features using the GRU model. This can accurately represent the user's features under complex social network relationships, solving the technical problem that it is difficult to construct complex relationships between users in existing neural network-based prediction methods.

[0062] In step 102, based on the social relationship embedding features, the GRU model is called to construct user spatiotemporal features, and the user spatiotemporal features are constructed based on information cascading to obtain user spatiotemporal cascaded features.

[0063] The social relationship embedding features of users are generated through the GCN model, thus determining the user's feature vector representation. Next, based on the social relationship embedding features, the GRU model is called to construct the user's spatiotemporal features. Since the user's spatiotemporal features need to be fused with user cascade information, the user cascade information needs to be extracted first.

[0064] In this embodiment of the invention, the user's historical cascaded data is first obtained, and then the user's historical cascaded data is divided into multiple cascaded subgraphs according to the timestamp. Since there is spatial and temporal correlation between each cascaded subgraph, the modeling process of the dynamically changing information cascaded graph is transformed into an aggregation problem of spatial and temporal features.

[0065] See Figure 2 , Figure 2 This is a schematic diagram of the information cascaded graph embedding layer based on spatiotemporal features provided by the present invention, such as... Figure 2 As shown, a multi-layer GCN network is used to aggregate user information and cascaded information. In the input layer, the social relationship embedding features are aggregated with the cascaded subgraphs to obtain the cascaded subgraph embedding features corresponding to each timestamp, denoted as... , where t represents the timestamp. Since each cascaded subgraph is segmented according to its timestamp, each timestamp also corresponds to a cascaded subgraph embedding feature. For example... Figure 2 In the diagram, the embedding features of n cascaded subgraphs are represented as " , … The corresponding n timestamps are represented as "t, t-1, ..., tn".

[0066] Next, we determine the time embedding feature corresponding to each timestamp, that is, we construct the corresponding time embedding feature based on the timestamp. For example... Figure 2 As shown, for each of the n timestamps, a corresponding temporal embedding feature is constructed. In the spatial feature extraction stage, the cascaded subgraph embedding feature and the temporal embedding feature are input into the GCN model for spatial feature extraction, resulting in a spatial feature sequence. Considering that the Gated Recurrent Unit (GRU) can handle long sequences well and has a simpler structure, the GRU model is invoked in the spatiotemporal feature extraction stage to extract features from the spatial feature sequence, obtaining the user's spatiotemporal features. This allows the GRU model to understand the dynamic changes in the data within the spatial feature sequence, enabling the user's spatiotemporal features to better simulate the user's feature representation at the corresponding timestamp.

[0067] like Figure 2 As shown, at the timestamp t-1 of the spatiotemporal feature extraction stage, the user's spatiotemporal features are obtained in the (t-1)th gated recurrent unit (GRU) of the GRU model. Then, it is input into the GCN model. The GCN model then incorporates the temporal embedding features of the next timestamp t. The spatial feature sequence at time stamp t is extracted, and then the spatial feature sequence at time stamp t is input into the t-th gated recurrent unit to obtain the user's spatiotemporal features at time stamp t. Therefore, at the subsequent timestamp t+1, the corresponding user spatiotemporal characteristics The construction process can be expressed as the following formulas (3) and (4):

[0068] (3)

[0069] (4)

[0070] In the above formula (3), t represents the timestamp. This represents the time embedding feature corresponding to the timestamp t. , The dimension representing the temporal embedding feature. Represents the spatiotemporal characteristics of a user at timestamp t. At the first timestamp (t=0), the user's spatiotemporal characteristics Using social relationship embedding features (L represents the number of layers in the GCN network) is initialized. This represents the adjacency matrix in the user's social network graph at timestamp t+1. N represents the number of nodes (users) in the user's social network relationship graph. express The processing functions of the network layers in the model. represents the activation function of the gated unit in the GRU model, and GRU represents the processing function of the gated unit in the GRU model.

[0071] In constructing user spatiotemporal features, this invention integrates corresponding time-embedded features based on timestamps, building upon user social relationship embedding features. This ensures that the user spatiotemporal features accurately represent the user's characteristics at each timestamp. Furthermore, it considers the potential dynamic changes in user spatiotemporal features between timestamps, utilizing the temporal characteristics of the GRU model to continuously output the user spatiotemporal features corresponding to each timestamp in chronological order, thereby further improving the accuracy of user spatiotemporal feature representation. This addresses the technical problem in existing neural network-based prediction methods where the cascaded structure changes over time, making it difficult to construct cascaded dynamic features during the modeling process.

[0072] After determining the user's spatiotemporal features, feature construction based on information cascading is performed on these features to obtain user spatiotemporal cascaded features. To obtain the user's content preference features at different timestamps, this embodiment of the invention utilizes a gating fusion mechanism to fuse user social information with user spatiotemporal features, then concatenates randomly initialized position embedding features, and finally uses a masked self-attention mechanism to learn sequence representations.

[0073] Specifically, for each timestamp corresponding to the user's time features, the social relationship embedding features and the user's spatiotemporal features are first gated and fused to obtain a fused spatiotemporal feature sequence.

[0074] The calculation process for gating fusion for each timestamp corresponding to the user's time characteristics is shown in the following formula (5):

[0075] (5)

[0076] In the above formula (5), This represents the processing function for gating fusion. This represents the embedded features of a user's social relationships. Represents the spatiotemporal characteristics of users. The weight parameter represents the weight of the social relationship embedding feature and the spatiotemporal feature. The calculation formula is as follows: Formula (6):

[0077] (6)

[0078] In the above formula (6), This represents the natural exponential function. Represents weight parameters The corresponding weight matrix. Indicates the embedded features of user social relationships The corresponding coding features are calculated using the following formula (7). Representing user spatiotemporal characteristics The corresponding coding features are calculated using the following formula (8):

[0079] (6)

[0080] (7)

[0081] In formulas (6) and (7) above, Indicates the embedded features of user social relationships The corresponding weight matrix, Representing user spatiotemporal characteristics The corresponding weight matrix, tanh represents the hyperbolic tangent function.

[0082] Using the above formulas (5) to (8), the social relationship embedding features and user spatiotemporal features under each timestamp can be gated and fused. Then, the results of the gated fusion under each timestamp are spliced ​​together to obtain the fused spatiotemporal feature sequence.

[0083] Next, determine the position embedding code corresponding to each sequence element in the fused spatiotemporal feature sequence, denoted as . Then, the position embedding code is concatenated with the sequence elements to obtain the concatenated feature sequence. Here, that is, the corresponding position feature is incorporated into each sequence element. Thus, each sequence element in the concatenated feature sequence is denoted as h, and is expressed as follows (9):

[0084]

[0085] In the above formula (9), This represents the embedded features of a user's social relationships. Represents the spatiotemporal characteristics of users. This indicates the result obtained by gating fusion under timestamps. This indicates the corresponding position embedding encoding.

[0086] Finally, the cascaded feature sequence is input into the Transformer model for mask attention processing to obtain the user's spatiotemporal cascaded features.

[0087] Considering that the Transformer model is an attention-based model, it can learn the sequence feature representation of spatiotemporal cascaded feature sequences through attention mechanisms. Therefore, the Transformer model is called here to extract features from the cascaded feature sequences to obtain the user's spatiotemporal cascaded features H. Furthermore, since some nodes (users) in the user's social network graph may already be active users—for example, a user may have already propagated (forwarded or shared) information to some other users before the current timestamp—the probability of the user continuing to propagate information to other users after the current timestamp is very small. Therefore, in the cascaded feature sequence, some sequence elements corresponding to certain timestamps do not represent the user's specific features and are irrelevant sequence elements. Therefore, this invention employs a masked self-attention mechanism, first masking the cascaded feature sequence, that is, masking some sequence elements in the cascaded feature sequence, and then inputting the masked cascaded feature sequence back into the Transformer model. The multi-head attention network in the Transformer model is used to process the cascaded feature sequence using a self-attention mechanism, obtaining the cascaded features processed by each head of the multi-head attention network. The processing procedure of the mask self-attention mechanism is expressed as the following formula (10):

[0088] (10)

[0089] In the above formula (10), X represents the sequence element in the concatenated feature sequence obtained through sequence masking. , , Let X represent the weight matrices corresponding to the query vector, key vector, and value vector of sequence element X in the self-attention mechanism, respectively. The processing function for the self-attention mechanism in the Transformer model is expressed as the following formula (11):

[0090] (11)

[0091] In the above formula (11), Q, K, and V represent the query vector, key vector, and value vector in the attention mechanism, respectively, and d represents the vector dimension. This represents the sequence element used for sequence masking in a cascaded feature sequence. The normalized exponential function is commonly used to represent the attention mechanism processing.

[0092] Using the above formulas (10) and (11), the cascaded features of each head in the multi-head attention network can be calculated. Next, the cascade sequence can be extracted from the user's historical cascade data. Then, the cascade feature corresponding to the user at the last timestamp can be found based on the cascade sequence, thereby determining the cascade feature at each timestamp. The cascaded features are concatenated in the order of concatenation and then spliced ​​together to obtain the overall cascaded feature combination. The overall cascaded feature combination is then processed by the Multilayer Perceptron (MLP) in the Transformer model to obtain the final user spatiotemporal cascaded feature H, which is expressed as the following formula (12):

[0093] (12)

[0094] In the above formula (12), MLP represents the processing function of the multilayer perceptron. This indicates the number of heads in the multi-head attention network of the Transformer model. This represents the overall cascaded feature combination. The final user spatiotemporal cascaded feature H can accurately characterize the dynamic cascaded features of users in complex social network relationships under spatiotemporal conditions.

[0095] Therefore, through steps 101 and 102 above, by using users' social relationship data and constructing user social relationship embedding features using the GRU model, user characteristics under complex social network relationships can be accurately represented. When constructing user spatiotemporal features, based on the user social relationship embedding features, corresponding time embedding features are incorporated according to timestamps. This allows the user spatiotemporal features to accurately express the user's characteristics at each timestamp, and also considers the potential dynamic changes in user spatiotemporal features between timestamps. Utilizing the temporal characteristics of the GRU model, the user spatiotemporal features corresponding to each timestamp are continuously output in time-stamp order, further improving the accuracy of user spatiotemporal feature representation. Finally, corresponding cascaded information is incorporated into the user spatiotemporal features, and the final user spatiotemporal cascaded features are extracted using the Transformer model, accurately representing the dynamic cascaded features of users under complex social network relationships under temporal and spatial conditions.

[0096] See also Figure 1 In step 103, the information dissemination content that the user will participate in is obtained, multimodal joint features are constructed based on the information dissemination content, and interest influence factors are determined based on the multimodal joint features.

[0097] After constructing the spatiotemporal cascaded features of users (i.e., the dynamic cascaded features of users in complex social network relationships) in step 102, the content preference features of users are then determined to characterize the content preferences of users for information dissemination content, thereby predicting the probability of users participating in information dissemination content.

[0098] First, it is necessary to acquire the information dissemination content. In this embodiment of the invention, the information dissemination content belongs to the collected information graphic data, specifically including visual images and corresponding text fragments. For example, the information dissemination content can be a post, an advertisement, or a news article, containing text descriptions and corresponding images. The text descriptions can be divided into multiple text fragments according to the images, with each text fragment describing one image. Here, the images are visual pictures, and the text fragments describing the images are the text fragments corresponding to the visual images. Since the information graphic data involves both text and image forms, the information dissemination content is also called multimodal joint information. Therefore, this invention constructs multimodal joint features based on the information dissemination content, which will be explained in detail below.

[0099] To address multimodal joint information, this invention employs a Contrastive Language-Image Pre-training (CLIP) model to extract multimodal joint features. For visual images, issues such as inconsistent sizes and color mismatches exist in information dissemination content. Therefore, visual images are transformed to adapt to the CLIP model. First, the shape of the visual image is changed to a uniform width and height of 224, using cubic spline interpolation for image scaling. Then, the visual image data is normalized, and the data range is transformed to a standard normal distribution. Finally, the processed visual image is input into the CLIP model's image encoder. Encoding through the image encoder yields the corresponding image encoded features. For information dissemination content (e.g., a post, denoted as...),... The j-th visual image corresponding to ) can be denoted as The initial image features can be obtained through the CLIP model image encoder, denoted as... If there is no corresponding visual image in the information content, all initial image features are initialized to a zero vector.

[0100] Considering that information dissemination on social networks can contain multiple visual images, each with varying importance, and that the position of these visual images influences the overall message, this invention constructs a visual information processing module based on the Transformer model to represent the differences in importance and position among visual images. This module further extracts image features from the initial image features obtained from the CLIP model, yielding the image encoding features of the visual images within the information dissemination content. Here, the initial image features are referred to as... The image is input into the Transformer model for embedding encoding to obtain the image encoding features of the visual image.

[0101] See details Figure 3 , Figure 3 This is a schematic diagram illustrating the principle of the visual information processing module for extracting image coding features provided by the present invention. The following is a detailed explanation. Figure 3 Let's explain in detail the feature extraction process of the visual information processing module. For example... Figure 3 As shown, firstly, based on the distribution of visual images in the information dissemination content, the initial image features of each visual image are determined. As sequence elements, the sequence elements are then concatenated to obtain the embedded sequence. For example, based on the distribution of images in a post, the initial image features of each image are used as sequence elements and concatenated to obtain the embedded sequence of the images.

[0102] In order to aggregate the information features of multiple visual images in the information dissemination content, that is...

[0103] This invention aggregates information from multiple images in a post to obtain global features.

[0104] Inserting the CLS tag as the first sequence element into the beginning of the embedding sequence yields the image embedding sequence; that is, adding the "[CLS]" tag to the beginning of the image embedding sequence. For example... Figure 3 As shown, "*" represents the CLS tag inserted at position 0 of the sequence. In addition, to avoid bias towards a particular visual image in the embedding sequence, the embedding of the "[CLS]" tag is set to be randomly initialized and continuously updated during training, which can encode the statistical characteristics of the entire embedding sequence.

[0105] like Figure 3 In the process of inserting CLS tags to obtain the image embedding sequence, the position encoding of each sequence element in the image embedding sequence is then determined. The positional encoding is then concatenated with the sequence elements to obtain the concatenated features. , which is represented by the following formula (13):

[0106] (13)

[0107] Finally, each concatenated feature in the image embedding sequence is input into the Transformer model for encoding to obtain the embedded encoding features of the visual image. Similar to the cascaded feature sequence processing described above, some visual images in the information dissemination content may have already been forwarded or shared by users and do not represent the specific features of the information dissemination content. These visual images belong to irrelevant sequence elements, so a masking mechanism is still used here. First, the image embedding sequence is sequence-masked, that is, some sequence elements in the image embedding sequence are randomly masked (excluding the inserted CLS tags), and then the masked image embedding sequence is input into the Transformer model. For example... Figure 3 As shown, the image embedding sequence is processed using a self-attention mechanism through a multi-head attention network in the Transformer Encoder. The result of the self-attention mechanism is then input into the final MLP layer, and the final output is the embedding encoding features of the visual image, denoted as... It is expressed as the following formula (14):

[0108] (14)

[0109] In the above formula (14), This represents the processing function for encoding the Transformer model. Represents the sequence element processed by the mask.

[0110] For text fragments corresponding to visual images, since each text fragment corresponds to one visual image, a sliding window approach can be used to extract the text fragment from each visual image. Then, the CLIP model's text encoder is called to encode the text fragment corresponding to each visual image, yielding the text encoder features, denoted as . Then, the text encoding features of each text segment are merged to obtain the overall text encoding features, denoted as... The calculation process is expressed as the following formula (15):

[0111] (15)

[0112] In the above formula (15), n represents the total number of text fragments obtained through the sliding window, i represents the i-th information content, and j represents the j-th text fragment in the information content.

[0113] Finally, the image encoding features are... Overall text encoding features Gated fusion is performed to obtain multimodal joint features. This invention uses a gated fusion mechanism to learn the importance of different modal information and the interaction between models, fusing image features and text features to obtain multimodal joint features, denoted as... Multimodal joint features can accurately express the information characteristics of information dissemination content, facilitating subsequent matching with user preferences and completing information dissemination. Gated fusion can be implemented using a neural network model or a feature fusion algorithm, and the calculation process is expressed as the following formula (16):

[0114] (16)

[0115] In the above formula (16), This represents the processing function for gating fusion.

[0116] Therefore, this embodiment of the invention extracts image features and text features of the information dissemination content separately, and takes into account the position of each visual image and the corresponding text fragment in the information dissemination content. It then fuses these image features and text features to obtain multimodal joint features, thereby accurately representing the information features of the information dissemination content. This solves the problem in existing neural network-based prediction methods where information dissemination content in social networks contains multimodal data including text and images, and where the information dissemination process is affected by the information's own content, making it difficult to model information content features.

[0117] After extracting the multimodal joint features of the information dissemination content, the interest influence factor is determined based on the multimodal joint features to express the user's interest characteristics.

[0118] Considering that user interests change over time, and that previously interacted information content influences later interests, the information content users participated in is arranged chronologically. To determine user interest characteristics, we first obtain historical user interaction data—records of interactions with previously participated information content. Then, we use an embedding feature extraction algorithm to extract interaction features from this historical data, which serve as the embedded features of the user's participation information, denoted as W. According to the above formula (16), the multimodal joint feature of each information dissemination content is denoted as... Therefore, the multimodal joint features of the information dissemination content to be predicted here can be denoted as w, and This means that predictions must be made for every piece of information dissemination.

[0119] Next, the Transformer model is called to encode the participating information embedding feature W to obtain the user interest embedding feature. Because user interests change over time, and the information disseminated in the early stages of interaction will affect the user's later interests, the purpose of calling the Transformer encoding layer again here is to learn the temporal characteristics of user interests and extract the user interest embedding features containing temporal characteristics. The encoding process is represented by the following formula (17):

[0120]

[0121] In the above formula (17), This represents the processing function of the Transformer encoding layer.

[0122] Given the large amount of historical user interaction data, user interest embedding features are crucial. The feature dimension is very large, therefore, before calculating user interest features, the user interest embedding features are... Average pooling is performed to obtain average pooling features. Then, the feature similarity I between the average pooling features and the multimodal joint features w is determined, and the feature similarity I is used as the interest influence factor. Here, the feature similarity can be calculated by Euclidean distance or cosine similarity. The higher the feature similarity, the more suitable the corresponding information dissemination content is with the user's interests, and the more likely the user is to participate in the information dissemination content and share or forward it. The lower the feature similarity, the less suitable the corresponding information dissemination content is with the user's interests, and the lower the probability that the user is likely to participate in the information dissemination content and share or forward it. The formula for calculating the interest influence factor is expressed as follows: Formula (18):

[0123]

[0124] In the above formula (18), This represents the processing function for the average pooling operation. This represents the embedded features representing user interests. This represents the multimodal joint features that need to be predicted. The function for calculating feature similarity is expressed as follows (19):

[0125] (19)

[0126] In the above formula (19), and Let d represent two feature vectors whose feature similarity needs to be calculated. , The feature dimension, where i represents the index of the feature dimension.

[0127] Through the above step 103, the present invention extracts multimodal joint features from information dissemination content, and then calculates user interest features based on user historical interaction data. In this way, the user's interest influence factor is calculated based on the multimodal joint features and user interest features, thereby integrating information content and user preferences and making a preliminary estimate of the likelihood of users participating in information dissemination content, which can further improve the accuracy of cascade prediction.

[0128] See also Figure 1 In step 104, the dot product similarity between the user's spatiotemporal cascaded features and the interest influencing factors is determined, and the probability of the user's participation in the information dissemination content is predicted based on the dot product similarity.

[0129] After determining the factors influencing user interests, the next step is to combine the dynamic cascading features of users in complex social network relationships (i.e., the final spatiotemporal cascading features of users obtained in step 102) to make a final prediction on whether users are likely to participate in information dissemination content.

[0130] In this embodiment of the invention, the dot product similarity between the user's spatiotemporal cascaded feature H and the interest influence factor I is directly calculated. Then, based on the dot product similarity, the user's participation probability for the information dissemination content is predicted, denoted as... The prediction process can be implemented using an MLP, as shown in the following formula (20):

[0131] (20)

[0132] In the above formula (20), The expression represents the dot product similarity between the user's spatiotemporal cascaded features H and the interest influence factor I, and MLP represents the processing function for predicting the participation probability of the multilayer perceptron.

[0133] Through the above step 104, the probability of user participation in information dissemination is finally predicted. Based on this participation probability, targeted information dissemination can be carried out to users to achieve accurate dissemination of information content.

[0134] In some embodiments, the probability of user participation in information dissemination content predicted in step 104 can also be obtained by calling an information cascade prediction model. The information cascade prediction model specifically includes modules such as the GCN model, GRU model, Transformer model, MLP, and CLIP model. The prediction process of the information cascade prediction model will be described below in conjunction with section 4. Figure 4 This is a schematic diagram illustrating the training process of the information cascade prediction model provided by this invention.

[0135] First, it is necessary to obtain training samples for the information cascade prediction model. This can be done by collecting user social relationship data samples, information dissemination content samples that users will participate in, user historical cascade data samples, and user historical interaction data samples, and using these samples as model training samples.

[0136] Then, the training samples from these models are input into the information cascade prediction model for forward propagation to obtain the predicted participation probability of users for the information dissemination content samples. The forward propagation process is described below.

[0137] Specifically, such as Figure 4 As shown, in the forward propagation process, initially, based on user social relationship data samples, the GCN model is called to generate user social relationship embedding features. Then, based on user historical cascaded data samples, cascaded subgraphs and cascaded sequences are generated respectively. By calling the GCN and GRU models, user temporal features are generated based on the cascaded subgraphs and social relationship embedding features. By calling the Transformer model, user spatiotemporal cascaded features are generated based on the cascaded sequences and user temporal features. Next, based on the information propagation content samples that users will participate in, the CLIP model is called to extract the multimodal joint features of the information propagation content samples. Then, based on user historical interaction data samples, the Transformer model is called to generate user interest embedding features. The next step is to calculate the similarity between user interest embedding features and multimodal joint features as an interest influence factor. Finally, the dot product similarity between the interest influence factor and user spatiotemporal cascaded features is calculated and input into the MLP to obtain the predicted participation probability, that is, the probability of user participation in the information propagation content samples. Thus, the forward propagation process of the information cascade prediction model ends.

[0138] Next, determine the actual participation probability of users for the information dissemination content sample, that is, determine the actual participation of users for the information dissemination content sample based on the actual information dissemination environment, denoted as y. Then, based on the predicted participation probability and the actual participation probability, construct the cross-entropy loss function Loss, the construction process of which is expressed as the following formula (21):

[0139] (twenty one)

[0140] In the above formula (21), n ​​represents the number of information dissemination content samples, and i represents the i-th information dissemination content sample. y represents the predicted participation probability obtained during the forward propagation of the information cascade prediction model, and y represents the actual participation probability of users for the information dissemination content sample.

[0141] After constructing the cross-entropy loss function, backpropagation is performed on the information cascade prediction model using the cross-entropy loss function. The backpropagation process is also the process of calculating the gradient. The gradient is optimized using the gradient descent algorithm to update the parameters of the information cascade prediction model, thereby causing the cross-entropy loss function to gradually converge.

[0142] The training process is also an iterative process. Each iteration involves forward and backward propagation. During forward propagation in each iteration, it's necessary to determine if the cross-entropy loss function (Loss) has significantly decreased. If it no longer decreases significantly, it indicates that the cross-entropy loss function has converged, and the predicted participation probability is output, ending the iteration and the training process. If it continues to decrease significantly, it means the cross-entropy loss function has not converged, and backward propagation continues to update the parameters of the information cascade prediction model.

[0143] Once trained, the information cascade prediction model can be used directly for prediction. Based on the input user information and information dissemination samples, it can directly predict the probability of user participation in the information dissemination content. Based on the predicted participation probability, it is possible to intuitively understand the dissemination of information on social network platforms, thereby enabling targeted information diffusion and completion of the cascade process, which can help with applications such as information recommendation and public opinion control.

[0144] Through the embodiments of this invention, on the one hand, user social relationship data is combined for modeling, and then user spatiotemporal features are combined to construct user spatiotemporal cascade features. On the other hand, multimodal joint features are constructed based on information dissemination content to determine user interest and preference features. Then, information cascade prediction is performed by combining spatiotemporal cascade features and interest and preference features. This overcomes the three difficulties of user relationship modeling, cascade spatiotemporal feature modeling, and information content feature modeling in existing neural network-based prediction methods, and solves the technical problem of low prediction accuracy and poor prediction effect in existing information cascade prediction tasks.

[0145] The following describes the information cascading prediction device based on spatiotemporal features and content preferences provided by the present invention. The information cascading prediction device based on spatiotemporal features and content preferences described below can be referred to in correspondence with the information cascading prediction method based on spatiotemporal features and content preferences described above.

[0146] Figure 5 This is a schematic diagram of the information cascade prediction device based on spatiotemporal features and content preferences provided by the present invention. Figure 5As shown, this invention provides an information cascade prediction device based on spatiotemporal features and content preferences. The device specifically includes the following modules: an acquisition module 501, a construction module 502, a determination module 503, and a prediction module 504. The acquisition module 501 acquires the user's social relationship data and uses a GCN model to generate social relationship embedding features. The construction module 502 uses the social relationship embedding features to construct spatiotemporal features based on a GRU model, and performs information cascade-based feature construction on these features to obtain spatiotemporal cascaded features. The determination module 503 acquires the information dissemination content the user will participate in, constructs multimodal joint features based on the information dissemination content, and determines an interest influence factor based on the multimodal joint features. The interest influence factor characterizes the user's content preference for the information dissemination content. The prediction module 504 determines the dot product similarity between the user's spatiotemporal cascaded features and the interest influence factor, and predicts the user's participation probability for the information dissemination content based on the dot product similarity.

[0147] Figure 6 This is a schematic diagram of the physical structure of an electronic device provided by the present invention, such as... Figure 6 As shown, the electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communications bus 640, wherein the processor 610, the communications interface 620, and the memory 630 communicate with each other through the communications bus 640. The processor 610 can call logic instructions in the memory 630 to execute an information cascade prediction method based on spatiotemporal features and content preferences. This method includes: acquiring a user's social relationship data and using a GCN model to generate embedded features of the social relationship data; constructing spatiotemporal features of the user based on the embedded features of the social relationship using a GRU model, and constructing spatiotemporal cascade features based on the spatiotemporal features of the user to obtain spatiotemporal cascade features; acquiring information dissemination content that the user will participate in, constructing multimodal joint features based on the information dissemination content, and determining an interest influence factor based on the multimodal joint features, wherein the interest influence factor is used to characterize the user's content preference for the information dissemination content; determining the dot product similarity between the user's spatiotemporal cascade features and the interest influence factor, and predicting the user's participation probability for the information dissemination content based on the dot product similarity.

[0148] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0149] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the information cascade prediction method based on spatiotemporal features and content preferences provided by the above methods. The method includes: acquiring a user's social relationship data and calling a GCN model to generate the social relationship data to obtain the user's social relationship embedding features; based on the social relationship embedding features, calling a GRU model to construct user spatiotemporal features, and performing information cascade-based feature construction on the user's spatiotemporal features to obtain user spatiotemporal cascade features; acquiring the information dissemination content that the user will participate in, constructing multimodal joint features based on the information dissemination content, and determining an interest influence factor according to the multimodal joint features, wherein the interest influence factor is used to characterize the user's content preference degree for the information dissemination content; determining the dot product similarity between the user's spatiotemporal cascade features and the interest influence factor, and predicting the user's participation probability for the information dissemination content based on the dot product similarity.

[0150] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the information cascade prediction method based on spatiotemporal features and content preferences provided by the above methods. The method includes: acquiring a user's social relationship data and calling a GCN model to generate the social relationship data to obtain the user's social relationship embedding features; based on the social relationship embedding features, calling a GRU model to construct user spatiotemporal features, and performing information cascade-based feature construction on the user's spatiotemporal features to obtain user spatiotemporal cascade features; acquiring information dissemination content that the user will participate in, constructing multimodal joint features based on the information dissemination content, and determining an interest influence factor according to the multimodal joint features, wherein the interest influence factor is used to characterize the user's content preference degree for the information dissemination content; determining the dot product similarity between the user's spatiotemporal cascade features and the interest influence factor, and predicting the user's participation probability for the information dissemination content based on the dot product similarity.

[0151] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0152] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An information cascade prediction method based on spatiotemporal features and content preferences, characterized in that, The method includes: The user's social relationship data is obtained, and the GCN model is called to generate the social relationship embedding features of the user. Based on the social relationship embedding features, the GRU model is invoked to construct user spatiotemporal features, and the user spatiotemporal features are then used to construct features based on information cascading to obtain user spatiotemporal cascaded features. The system obtains the information dissemination content that users will participate in, constructs multimodal joint features based on the information dissemination content, and determines interest influence factors based on the multimodal joint features, wherein the interest influence factors are used to characterize the degree of user content preference for the information dissemination content; Determine the dot product similarity between the user's spatiotemporal cascaded features and the interest influencing factors, and predict the user's participation probability for the information dissemination content based on the dot product similarity. The step of constructing user spatiotemporal features based on the social relationship embedding features and calling the GRU model includes: acquiring user historical cascaded data and segmenting the user historical cascaded data into multiple cascaded subgraphs according to timestamps; aggregating the social relationship embedding features and the cascaded subgraphs to obtain cascaded subgraph embedding features corresponding to each timestamp; determining the temporal embedding features corresponding to each timestamp; inputting the cascaded subgraph embedding features and the temporal embedding features into the GCN model for spatial feature extraction to obtain a spatial feature sequence; and calling the GRU model to extract features from the spatial feature sequence to obtain user spatiotemporal features. The step of constructing user spatiotemporal features based on information concatenation to obtain user spatiotemporal concatenated features includes: gating and fusing the social relationship embedding features with the user spatiotemporal features to obtain a fused spatiotemporal feature sequence; determining the position embedding code corresponding to each sequence element in the fused spatiotemporal feature sequence, and concatenating the position embedding code with the sequence element to obtain a spatiotemporal concatenated feature sequence; and inputting the spatiotemporal concatenated feature sequence into a Transformer model for mask attention mechanism processing to obtain user spatiotemporal concatenated features. The information dissemination content includes visual images and corresponding text fragments. The construction of multimodal joint features based on the information dissemination content includes: calling the CLIP model's image encoder to encode each visual image, obtaining initial image features; inputting the initial image features into a Transformer model for embedding encoding, obtaining image-encoded features of the visual images; calling the CLIP model's text encoder to encode the text fragments corresponding to each visual image, obtaining text encoder features, and merging the text encoding features to obtain overall text encoding features; and performing gated fusion of the image encoding features and the overall text encoding features to obtain multimodal joint features. The step of inputting the initial image features into a Transformer model for embedding encoding to obtain the embedding encoding features of the visual image includes: concatenating each initial image feature as a sequence element to obtain an embedding sequence, and inserting the CLS label as the first sequence element into the head of the embedding sequence to obtain an image embedding sequence; determining the positional encoding of each sequence element in the image embedding sequence, and concatenating the positional encoding with the sequence element to obtain a concatenated feature; and inputting each concatenated feature in the image embedding sequence into a Transformer model for encoding processing to obtain the embedding encoding features of the visual image. The step of determining the interest influence factor based on the multimodal joint features includes: extracting user participation information embedding features based on user historical interaction data, and calling the Transformer model to encode the participation information embedding features to obtain user interest embedding features; performing average pooling processing on the user interest embedding features to obtain average pooled features; determining the feature similarity between the average pooled features and the multimodal joint features, and using the feature similarity as the interest influence factor; The participation probability is obtained by calling an information cascade prediction model. The training method of the information cascade prediction model includes: using user social relationship data samples, information dissemination content samples, user historical cascade data samples, and user historical interaction data samples as model training samples; inputting the model training samples into the information cascade prediction model for forward propagation to obtain the predicted participation probability of the user for the information dissemination content sample; determining the user's actual participation probability for the information dissemination content sample; constructing a cross-entropy loss function based on the predicted participation probability and the actual participation probability, and performing backpropagation through the cross-entropy loss function in the information cascade prediction model to update the parameters of the information cascade prediction model.

2. An information cascade prediction device based on spatiotemporal characteristics and content preferences, characterized in that, The device includes: The acquisition module is used to acquire the user's social relationship data and call the GCN model to generate the social relationship data to obtain the user's social relationship embedding features; The construction module is used to construct user spatiotemporal features by calling the GRU model based on the social relationship embedding features, and to construct user spatiotemporal cascaded features based on information concatenation of the user spatiotemporal features. The determination module is used to obtain the information dissemination content that the user will participate in, construct multimodal joint features based on the information dissemination content, and determine the interest influence factor according to the multimodal joint features, wherein the interest influence factor is used to characterize the user's content preference for the information dissemination content; The prediction module is used to determine the dot product similarity between the user's spatiotemporal cascaded features and the interest influencing factors, and to predict the user's participation probability for the information dissemination content based on the dot product similarity. The step of constructing user spatiotemporal features based on the social relationship embedding features and calling the GRU model includes: acquiring user historical cascaded data and segmenting the user historical cascaded data into multiple cascaded subgraphs according to timestamps; aggregating the social relationship embedding features and the cascaded subgraphs to obtain cascaded subgraph embedding features corresponding to each timestamp; determining the temporal embedding features corresponding to each timestamp; inputting the cascaded subgraph embedding features and the temporal embedding features into the GCN model for spatial feature extraction to obtain a spatial feature sequence; and calling the GRU model to extract features from the spatial feature sequence to obtain user spatiotemporal features. The step of constructing user spatiotemporal features based on information concatenation to obtain user spatiotemporal concatenated features includes: gating and fusing the social relationship embedding features with the user spatiotemporal features to obtain a fused spatiotemporal feature sequence; determining the position embedding code corresponding to each sequence element in the fused spatiotemporal feature sequence, and concatenating the position embedding code with the sequence element to obtain a spatiotemporal concatenated feature sequence; and inputting the spatiotemporal concatenated feature sequence into a Transformer model for mask attention mechanism processing to obtain user spatiotemporal concatenated features. The information dissemination content includes visual images and corresponding text fragments. The construction of multimodal joint features based on the information dissemination content includes: calling the CLIP model's image encoder to encode each visual image, obtaining initial image features; inputting the initial image features into a Transformer model for embedding encoding, obtaining image-encoded features of the visual images; calling the CLIP model's text encoder to encode the text fragments corresponding to each visual image, obtaining text encoder features, and merging the text encoding features to obtain overall text encoding features; and performing gated fusion of the image encoding features and the overall text encoding features to obtain multimodal joint features. The step of inputting the initial image features into a Transformer model for embedding encoding to obtain the embedding encoding features of the visual image includes: concatenating each initial image feature as a sequence element to obtain an embedding sequence, and inserting the CLS label as the first sequence element into the head of the embedding sequence to obtain an image embedding sequence; determining the positional encoding of each sequence element in the image embedding sequence, and concatenating the positional encoding with the sequence element to obtain a concatenated feature; and inputting each concatenated feature in the image embedding sequence into a Transformer model for encoding processing to obtain the embedding encoding features of the visual image. The step of determining the interest influence factor based on the multimodal joint features includes: extracting user participation information embedding features based on user historical interaction data, and calling the Transformer model to encode the participation information embedding features to obtain user interest embedding features; performing average pooling processing on the user interest embedding features to obtain average pooled features; determining the feature similarity between the average pooled features and the multimodal joint features, and using the feature similarity as the interest influence factor; The participation probability is obtained by calling an information cascade prediction model. The training method of the information cascade prediction model includes: using user social relationship data samples, information dissemination content samples, user historical cascade data samples, and user historical interaction data samples as model training samples; inputting the model training samples into the information cascade prediction model for forward propagation to obtain the predicted participation probability of the user for the information dissemination content sample; determining the user's actual participation probability for the information dissemination content sample; constructing a cross-entropy loss function based on the predicted participation probability and the actual participation probability, and performing backpropagation through the cross-entropy loss function in the information cascade prediction model to update the parameters of the information cascade prediction model.

3. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the information cascade prediction method based on spatiotemporal features and content preferences as described in claim 1.

4. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the information cascade prediction method based on spatiotemporal features and content preferences as described in claim 1.

Images