A method, device and equipment for constructing a multi-source data fusion prediction model based on graph propagation

Abstract

The application discloses a kind of multi-source data fusion prediction model construction method, device and equipment based on graph propagation, belong to data fusion technical field.Method includes: with the enhanced data node feature of each type of data source as graph node, the information interaction path between any two enhanced data nodes is edge to establish fully connected directed graph structure;Enhanced data node feature is mapped to multi-source data and is obtained by enhancement;Information propagation and difference update are carried out to each node pair of graph structure, and graph node feature vector is obtained;Feature fusion is carried out to graph node feature vector, and global fusion feature vector is obtained;Graph node feature vector and global fusion feature vector are input into prediction network, and total loss of training is determined according to prediction result;Whether prediction model meets convergence condition is determined according to total loss, if yes, then output final prediction model, otherwise, parameter iterative update is carried out to prediction model until convergence is reached.The application can improve the prediction precision and generalization ability of fusion prediction model.

Description

Technical Field

[0001] This invention relates to the field of data fusion technology, and in particular to a method, apparatus and equipment for constructing a multi-source data fusion prediction model based on graph propagation. Background Technology

[0002] In complex system modeling and intelligent prediction tasks, there are often multiple data sources with different qualities, resolutions, or precisions. While high-fidelity data offers high accuracy, it is costly to acquire and has a limited sample size; low-fidelity data, although easily obtained, is noisy and prone to errors. Therefore, effectively integrating data of different fidelities to improve high-fidelity prediction capabilities is a crucial issue in the field of multi-source data modeling.

[0003] In related technologies, the multi-source data fusion prediction process usually requires manual pre-setting of the fidelity level and hierarchical relationship of different data sources. At the same time, the fusion process only supports one-way information migration and lacks a unified information interaction mechanism, which makes it difficult for the prediction accuracy and generalization ability of the finally constructed fusion prediction model to meet actual needs.

[0004] Therefore, there is an urgent need for a method, apparatus, and equipment for constructing a multi-source data fusion prediction model based on graph propagation to solve the above-mentioned technical problems. Summary of the Invention

[0005] This invention provides a method, apparatus, and device for constructing a multi-source data fusion prediction model based on graph propagation, which can effectively improve the prediction accuracy and generalization ability of the fusion prediction model. The technical solution is as follows: On the one hand, a method for constructing a multi-source data fusion prediction model based on graph propagation is provided, the method comprising: A fully connected directed graph structure is established using the enhanced data node features of each type of data source as graph nodes and the information interaction path between any two enhanced data nodes as edges; wherein, the enhanced data node features are obtained by inputting multi-source data that has undergone mapping processing into a preset data enhancement network; Based on the optimized bidirectional information propagation network, bidirectional information propagation and difference learning updates are performed on each node pair of the fully connected directed graph structure to obtain graph node feature vectors containing information from all data sources and differences in the structure of different data sources. The graph node feature vectors are subjected to multi-scale hierarchical feature fusion processing according to a preset fusion network to obtain a global fusion feature vector. The graph node feature vectors and the global fusion feature vectors are input into a preset prediction network, and the total training loss is determined based on the prediction results. Based on the total loss, determine whether the prediction model, consisting of the data augmentation network, the two-way information propagation network, the fusion network, and the prediction network, meets the preset convergence condition. If it does, output the final prediction model; otherwise, iterate and update the parameters of each network in the prediction model based on the total loss until the model converges.

[0006] On the other hand, a device for constructing a multi-source data fusion prediction model based on graph propagation is provided, the device comprising: A construction module is used to establish a fully connected directed graph structure with the enhanced data node features of each type of data source as graph nodes and the information interaction path between any two enhanced data nodes as edges; wherein, the enhanced data node features are obtained by inputting multi-source data that has undergone mapping processing into a preset data enhancement network; The propagation module is used to perform bidirectional information propagation and difference learning updates on each node pair of the fully connected directed graph structure according to the optimized bidirectional information propagation network, so as to obtain graph node feature vectors containing information from the full data source and differences in the structure of different data sources. The fusion module is used to perform multi-scale hierarchical feature fusion processing on the graph node feature vectors according to a preset fusion network to obtain a global fusion feature vector. The prediction module is used to input the graph node feature vector and the global fusion feature vector into a preset prediction network, and determine the total training loss based on the prediction results. The update module is used to determine whether the prediction model consisting of the data augmentation network, the bidirectional information propagation network, the fusion network, and the prediction network meets the preset convergence conditions based on the total loss. If so, the final prediction model is output; otherwise, the parameters of each network in the prediction model are iteratively updated based on the total loss until the model converges.

[0007] On the other hand, a computer device is provided, the computer device including a memory and a processor, the memory for storing computer programs, and the processor for executing the computer programs stored in the memory to implement the steps of the above-described method for constructing a multi-source data fusion prediction model based on graph propagation.

[0008] The technical solution provided by this invention can bring at least the following beneficial effects: By constructing nodes with learnable embeddings, without the need for manual definition of fidelity levels, it can adaptively mine potential hierarchical relationships of multi-source data and adapt to scenarios with unknown fidelity; then, through bidirectional information flow propagation and difference learning branches of a fully connected directed graph, it simultaneously realizes low-fidelity information transfer and high-fidelity reverse constraints, explicitly models the differences in data source structures, and constructs a unified multi-source information interaction framework; subsequently, through a multi-level fusion mechanism, it completes the progressive aggregation from fine-grained interaction to global representation, taking into account both common features and differentiated information; finally, through multi-objective loss joint training, it strengthens feature alignment and difference learning constraints, ultimately effectively improving the prediction accuracy and model generalization ability of high-fidelity data. This method addresses the core defects of existing multi-fidelity data fusion methods, such as reliance on manual hierarchical setting, unidirectional information transfer, lack of difference modeling, and inconsistent interaction mechanisms, solving the pain points of scarcity of high-fidelity data and low utilization of low-fidelity data in complex system modeling. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of a method for constructing a multi-source data fusion prediction model based on graph propagation, provided in an embodiment of the present invention. Figure 2 This is a structural diagram of a multi-source data fusion prediction model construction device based on graph propagation provided in an embodiment of the present invention; Figure 3 This is a hardware architecture diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0012] As mentioned earlier, the multi-source data fusion prediction process requires manual intervention and involves one-way information transmission, which makes it difficult for the final prediction accuracy and model generalization ability to meet application needs.

[0013] Based on this, the concept of the present invention is to construct a directed graph structure between multi-fidelity data to realize information dissemination, differential learning and hierarchical fusion across data sources, thereby overcoming the shortcomings of the existing technology.

[0014] The following describes the specific implementation of the above concept.

[0015] Please refer to Figure 1 This invention provides a method for constructing a multi-source data fusion prediction model based on graph propagation, the method comprising: Step 100: Establish a fully connected directed graph structure with the enhanced data node features of each type of data source as graph nodes and the information interaction path between any two enhanced data nodes as edges; wherein, the enhanced data node features are obtained by inputting multi-source data that has undergone mapping processing into a preset data enhancement network; Step 102: Perform bidirectional information propagation and difference learning updates on each node pair of the fully connected directed graph structure according to the optimized bidirectional information propagation network to obtain graph node feature vectors containing information from all data sources and differences in the structure of different data sources. Step 104: Perform multi-scale hierarchical feature fusion processing on the graph node feature vectors according to the preset fusion network to obtain a global fusion feature vector; Step 106: Input the graph node feature vector and the global fusion feature vector into a preset prediction network, and determine the total training loss based on the prediction results; Step 108: Determine whether the prediction model consisting of the data augmentation network, bidirectional information propagation network, fusion network and prediction network meets the preset convergence condition based on the total loss. If yes, output the final prediction model; otherwise, iterate and update the parameters of each network of the prediction model according to the total loss until the model converges.

[0016] In this embodiment of the invention, by constructing nodes with learnable embeddings, the fidelity levels can be adaptively mined from multi-source data without manual definition, adapting to scenarios with unknown fidelity. Then, through bidirectional information flow propagation and difference learning branches in a fully connected directed graph, low-fidelity information transfer and high-fidelity reverse constraints are simultaneously achieved, explicitly modeling differences in data source structures and constructing a unified multi-source information interaction framework. Subsequently, a multi-layered fusion mechanism completes the progressive aggregation from fine-grained interaction to global representation, taking into account both common features and differentiated information. Finally, through joint training with multi-objective loss, feature alignment and difference learning constraints are strengthened, ultimately effectively improving the prediction accuracy and model generalization ability of high-fidelity data. This method addresses the core shortcomings of existing multi-fidelity data fusion methods, such as reliance on manual hierarchical setting, unidirectional information transfer, lack of difference modeling, and inconsistent interaction mechanisms, solving the pain points of scarcity of high-fidelity data and low utilization of low-fidelity data in complex system modeling.

[0017] The following description Figure 1 The execution method of each step is shown.

[0018] First, for step 100, a fully connected directed graph structure is established with the enhanced data node features of each type of data source as graph nodes and the information interaction path between any two enhanced data nodes as edges.

[0019] In this embodiment of the invention, the enhanced data node features are obtained by inputting multi-source data that has undergone mapping processing into a preset data enhancement network.

[0020] Specifically, suppose there exists Different types of fidelity data ,in Indicates the first Data source class. Among them... Represents input features; For the corresponding tags; Indicates the sample size. High-fidelity data labels: As the primary forecast target; low-fidelity label As an auxiliary monitoring signal.

[0021] For the k-th type of data source, its input is first mapped to latent space features through an encoder. : in, For the k-th type of data source; For feature encoder; Since this embodiment is designed for scenarios without fidelity priors, a learnable embedding vector is introduced: The latent space feature data is summed with a pre-defined learnable vector to obtain enhanced data node features. : in, These are learnable embedding vectors.

[0022] After obtaining the enhanced data node features, the enhanced data node features of each data source type are used as graph nodes. The information exchange path between any two enhanced data nodes is an edge. Establish a fully connected directed graph structure : in, Each node corresponds to a data source.

[0023] The edge set is defined as: This indicates that there is an information exchange path between any two data sources.

[0024] Then, for step 102, bidirectional information propagation and difference learning updates are performed on each node pair of the fully connected directed graph structure according to the optimized bidirectional information propagation network to obtain graph node feature vectors containing information from all data sources and differences in the structure of different data sources.

[0025] In this embodiment of the invention, the graph node feature vector is obtained through the following steps: The node pairs consisting of target graph nodes and source graph nodes are concatenated along the feature dimension to obtain a concatenated vector that fuses the feature context of the two nodes.

[0026] Specifically, each node in the diagram corresponds to a type of data source. Since all nodes have been mapped to a unified latent feature space, they have the computational basis for cross-node interaction. Therefore, in order to achieve interaction between different data sources, any two nodes (such as source node i and target node j) are used to form a node pair. The source node representation and the target node representation are concatenated along the feature dimension to obtain a concatenated vector that fuses the dual-node feature context.

[0027] Furthermore, the splicing vector is input into a preset forward information flow attention network and a reverse information flow attention network, respectively, and the forward propagation information from the source graph node to the target graph node and the reverse calibration information from the target graph node to the source graph node are output sequentially.

[0028] Specifically, the concatenated vector is input into the forward information flow attention network, and the attention weights and forward transmission information from the source node to the target node are calculated through nonlinear transformation, realizing attention mapping between different data sources—the source node information with higher semantic matching degree has a greater contribution weight to the target node.

[0029] Positive dissemination of information It is calculated using the following formula: in, For feature splicing symbols; For source graph nodes; For the target graph nodes; It is a positive information flow attention network.

[0030] The process of positive information propagation can realize the one-way information injection from the auxiliary data source to the target data source; under the condition that no fidelity level is preset by humans, the model can automatically learn the information supply relationship between nodes and complete the semantic transfer across data sources.

[0031] The concatenated vector is input into the reverse information flow attention network, and the reverse attention weights are calculated through nonlinear transformation. Then, the feature information of the baseline node is weighted and compressed based on the weights to generate the constraint information for reverse propagation. The structural features of high-quality nodes are then injected into low-quality nodes.

[0032] Reverse calibration information It is calculated using the following formula: in, It is a reverse information flow attention network.

[0033] This backpropagation process utilizes the stable distribution of high-quality data to constrain and stabilize the representation learning process of low-quality nodes, reducing the negative impact of noise from low-fidelity data on the fusion results. Combined with the forward information flow, it forms a bidirectional interactive mechanism of "low-precision information supplementation and high-precision back-calibration," which is one of the innovative aspects of this embodiment.

[0034] Furthermore, based on the difference information of the node pairs, multi-layer nonlinear transformation processing is performed on the multi-dimensional difference features of the node pairs to obtain a difference vector representing the structural differences between the two types of data sources.

[0035] Specifically, in order to address the problem that traditional methods only force feature alignment and cannot characterize the inherent structural differences of data sources, this embodiment constructs three types of complementary difference features based on node pairs to comprehensively capture difference patterns.

[0036] First, the original baseline feature values ​​of the two nodes are used to provide the feature baseline context.

[0037] Next, the feature difference between the two nodes is calculated based on the baseline feature value to capture the linear numerical difference in the feature dimension; at the same time, the element-wise product of the two nodes is calculated to capture the non-linear interactive difference between the feature dimensions.

[0038] Subsequently, the baseline feature values, linear numerical differences, and nonlinear interaction differences are sequentially input into a preset neural difference network for multi-layer nonlinear transformation to obtain the difference vector. : In the formula, It is a neural differential network; This is an element-wise product.

[0039] Output This represents the difference vector between two data sources, explicitly encoding the structural differences in feature distribution, noise level, and accuracy level between different data sources. The output of this branch does not directly participate in the final prediction calculation; it is only used for self-supervised training through a dedicated difference learning loss. This helps the model distinguish between "common information shared by multiple sources" and "difference information unique to each data source," avoiding the loss of effective information caused by forced feature alignment.

[0040] Furthermore, based on the forward propagation information, reverse calibration information, and difference vectors of all node pairs related to the current propagation layer and the graph node to be updated, the feature vector of the graph node to be updated is subjected to multi-source information aggregation and residual update processing to obtain the updated graph node of the next propagation layer after fusion.

[0041] Specifically, after completing a single round of bidirectional propagation and difference calculation, the node representation of the next layer is generated through a node update mechanism, and finally, deep information fusion of the entire graph is achieved through multi-layer iteration.

[0042] For example, in the first l After the layer propagation is completed, the first node j... l The +1 layer means that the update will aggregate four types of information: the original node representation of the current layer, the forward propagation information from all other nodes, the backward constraint information from all other nodes, and the structural constraint information from the difference vector. In the formula, Represents a node j In the l Current feature representation in layer graph propagation, This represents the (l+1)th layer updated feature representation obtained by aggregating the information passed from other associated nodes by node j; It is a multilayer perceptron.

[0043] The update process preserves the core features of the node itself, while internalizing multi-source interaction information to achieve iterative upgrades of the representation.

[0044] The final step is to determine whether the number of iterations meets the preset threshold. If so, the iteration stops and the graph node feature vector containing information from all data sources and the structural differences between different data sources is output.

[0045] Specifically, when the number of iterations does not meet the preset requirement, the complete process of "forward propagation + backward propagation + difference calculation + node update" will be repeated until the preset number of propagation layers is reached. Shallow propagation only realizes direct interaction between pairs of nodes, while deep propagation can realize the full graph diffusion of information through intermediate nodes, allowing each node to eventually integrate effective information from all data sources.

[0046] The feature vectors of all nodes after the last layer of graph propagation are output as graph node feature vectors containing information from all data sources and differences in the structure of different data sources.

[0047] The above process unifies the multi-source interaction through a graph structure, and simultaneously realizes the triple capabilities of information migration, reverse calibration, and difference modeling, fundamentally solving the shortcomings of existing technologies such as unidirectional migration, non-differential modeling, and lack of a unified interaction framework.

[0048] For step 104, the graph node feature vectors are subjected to multi-scale hierarchical feature fusion processing according to the preset fusion network to obtain a global fusion feature vector.

[0049] In this embodiment of the invention, after obtaining the graph node feature vectors, a progressive aggregation structure from fine to coarse is adopted to perform hierarchical refinement and integration of multi-source node features, ultimately generating a global fusion feature vector that can be directly used for prediction. Specifically, this is achieved through the following steps: Based on the consistency requirements of graph node feature vectors and feature semantics, attention-weighted aggregation processing is performed on nodes with similar semantic patterns to obtain multiple local aggregated features corresponding to different semantic groups.

[0050] Specifically, firstly, local information groups are divided based on the semantic consistency of node features, and then the fusion weights for each node are adaptively learned through an attention mechanism: In the formula, Network weights; Node fusion weights.

[0051] By combining network weights to perform a weighted summation of node features within a local range, multiple single-node representations are aggregated into a local aggregated feature: In the formula, This is a local aggregation feature.

[0052] The weights will be automatically adjusted based on the quality of the information from each node, prioritizing the retention of valid information from high-quality data sources.

[0053] Furthermore, all local aggregated features are subjected to a dual-path concatenation process of global average pooling and global max pooling to obtain a fused global feature vector.

[0054] Specifically, the global fusion feature vector It is calculated in the following way: This embodiment employs a dual-pooling parallel strategy to perform global feature integration on all local aggregated representations for two purposes: first, to capture the overall statistical characteristics and common patterns shared by all data sources through global average pooling; and second, to extract the most discriminative salient features from each group through global max pooling. The results of both pooling methods together constitute the final global fused feature vector.

[0055] For step 106, the graph node feature vector and the global fusion feature vector are input into a preset prediction network, and the total training loss is determined based on the prediction results.

[0056] In this embodiment of the invention, the total training loss is determined through the following steps: The global fusion feature vector is input into a preset prediction head network, and a unique and effective high-fidelity prediction result is output. The high-fidelity task loss is calculated based on the output result.

[0057] Specifically, the prediction head network (a non-linear transformation network consisting of fully connected layers and activation functions) maps the globally fused features to an output space consistent with the dimension of the high-fidelity labels, generating the final high-fidelity prediction result. This result is the core objective output of the model and the only prediction branch that is effective in practical applications.

[0058] The high-fidelity task loss was calculated based on real high-fidelity data. : In the formula, This represents the total number of high-fidelity samples.

[0059] This loss is the core objective of model optimization, directly determining the accuracy of high-fidelity prediction, and has the highest weight in the total loss.

[0060] Furthermore, the feature vector of each graph node is input into a preset auxiliary prediction network, and the auxiliary prediction result of the corresponding data source is output. The auxiliary supervision loss is then calculated based on the output result.

[0061] Specifically, this embodiment configures an independent auxiliary prediction network for each type of data source, mapping the representation of the corresponding node to the output space of its respective label, and generating auxiliary prediction results for a single data source. .

[0062] By supervising predictions using a single data source, the representation learning process of each node is constrained, avoiding the drift of feature semantics from a single data source during multi-source information fusion and improving the overall training stability.

[0063] It is worth noting that this branch is only enabled during the training phase; it does not participate in computation once the model has been trained and put into practical application.

[0064] Auxiliary monitoring loss It is calculated using the following formula: This loss provides an independent supervision signal for the representation learning of each data source node, stabilizes the feature learning process of each node, and prevents feature degradation caused by multi-source fusion.

[0065] Furthermore, based on the node pairs and their corresponding difference vectors, the cross-fidelity consistency loss and difference learning loss are calculated.

[0066] Specifically, cross-fidelity consistency loss Calculations are performed using a contrastive learning approach: In the formula, This represents the similarity function.

[0067] The purpose of this loss function is to bring the representations (positive sample pairs) of the same object from different data sources as close as possible in the feature space, while keeping the representations (negative sample pairs) of different objects as far apart as possible, thereby achieving feature alignment across data sources.

[0068] Differential learning loss Self-supervised pseudo-labels constructed based on the difference vector and the difference between the real labels. To determine the pseudo-labels, the difference between the real labels from different data sources is used to supervise the learning process of the neural differential network, and the deviation between the difference vector and the pseudo-labels is calculated: This loss function is used to explicitly constrain the difference learning branch, allowing the model to accurately capture structural difference patterns between different data sources and improve the effectiveness of difference modeling.

[0069] The final step involves weighting and summing the high-fidelity task loss, auxiliary supervision loss, cross-fidelity consistency loss, and differential learning loss according to preset loss weights to calculate the total training loss.

[0070] Specifically, the four types of losses mentioned above are weighted and summed according to preset weights to obtain the overall optimization objective of the model, i.e., the total loss. : Among them, weight , , Adjustments can be made according to actual needs, which will not be elaborated here.

[0071] For step 108, determine whether the prediction model consisting of the data augmentation network, bidirectional information propagation network, fusion network and prediction network meets the preset convergence condition based on the total loss. If it does, output the final prediction model; otherwise, iterate and update the parameters of each network of the prediction model according to the total loss until the model converges.

[0072] In this embodiment of the invention, based on the total loss function, all learnable parameters of the entire model are iteratively optimized to achieve end-to-end joint training. The training includes the following process: Multi-source training data is input in batches, and forward computation is performed sequentially through node representation modeling, bidirectional information flow graph propagation, and multi-scale hierarchical fusion to obtain the main prediction result and each auxiliary prediction result. Based on the forward output, the four types of sub-losses are calculated sequentially and weighted to obtain the total loss; The gradient of the total loss is propagated back from the output layer layer by layer through the backpropagation algorithm, and the learnable parameters of all modules are updated synchronously, including the feature encoder, learnable embedding vector, bidirectional attention network, neural differential network, fusion layer weights, and parameters of each prediction head. Repeat the batch iteration process described above until the total loss converges and the high-fidelity prediction accuracy reaches the preset threshold, thus completing the model training.

[0073] During training, the learnable embedding vectors automatically learn the potential hierarchical relationships of different data sources as the gradient is updated, without the need to manually specify the fidelity level.

[0074] Please refer to Figure 2 This invention provides an apparatus for constructing a multi-source data fusion prediction model based on graph propagation. The apparatus includes: The construction module 200 is used to establish a fully connected directed graph structure with the enhanced data node features of each type of data source as graph nodes and the information interaction path between any two enhanced data nodes as edges; wherein, the enhanced data node features are obtained by inputting multi-source data that has undergone mapping processing into a preset data enhancement network; The propagation module 202 is used to perform bidirectional information propagation and difference learning updates on each node pair of the fully connected directed graph structure according to the optimized bidirectional information propagation network, so as to obtain graph node feature vectors containing information from the full data source and differences in the structure of different data sources. The fusion module 204 is used to perform multi-scale hierarchical feature fusion processing on the graph node feature vectors according to a preset fusion network to obtain a global fusion feature vector. The prediction module 206 is used to input the graph node feature vector and the global fusion feature vector into a preset prediction network, and determine the total training loss based on the prediction results. The update module 208 is used to determine whether the prediction model consisting of the data augmentation network, the bidirectional information propagation network, the fusion network and the prediction network meets the preset convergence conditions based on the total loss. If so, the final prediction model is output; otherwise, the parameters of each network in the prediction model are iteratively updated based on the total loss until the model converges.

[0075] In this embodiment of the invention, multi-source data that has undergone mapping processing is input into a preset data augmentation network to obtain augmented data node features, including: The multi-source data is input into a preset feature encoder and mapped to obtain latent space features. : in, For the k-th type of data source; For feature encoder; The latent space feature data is summed with a pre-defined learnable vector to obtain enhanced data node features. : in, These are learnable embedding vectors.

[0076] In this embodiment of the invention, the step of performing bidirectional information propagation and difference learning updates on each node pair of a fully connected directed graph structure according to an optimized bidirectional information propagation network to obtain graph node feature vectors containing information from all data sources and differences in the structures of different data sources includes: The node pairs consisting of target graph nodes and source graph nodes are concatenated along the feature dimension to obtain a concatenated vector that fuses the dual-node feature context. The spliced ​​vector is input into a preset forward information flow attention network and a reverse information flow attention network, respectively, and the forward propagation information from the source graph node to the target graph node and the reverse calibration information from the target graph node to the source graph node are output in sequence. Based on the difference information of the node pair, the multi-dimensional difference features of the node pair are subjected to multi-layer nonlinear transformation to obtain a difference vector representing the structural differences between the two types of data sources. Based on the forward propagation information, reverse calibration information, and difference vector of all node pairs related to the current propagation layer and the graph node to be updated, multi-source information aggregation and residual update processing are performed on the feature vector of the graph node to be updated to obtain the updated graph node of the next propagation layer after fusion. Determine whether the number of iterations meets the preset threshold. If so, stop the iteration and output the graph node feature vector containing information from all data sources and the structural differences between different data sources.

[0077] In this embodiment of the invention, the forward propagation information It is calculated using the following formula: in, For feature splicing symbols; For source graph nodes; For the target graph nodes; It is a positive information flow attention network; The reverse calibration information It is calculated using the following formula: in, It is a reverse information flow attention network.

[0078] In this embodiment of the invention, the step of performing multi-layer nonlinear transformation processing on the multi-dimensional difference features of the node pairs based on the difference information of the node pairs to obtain a difference vector characterizing the structural differences between the two types of data sources includes: Based on the baseline eigenvalues ​​of two graph nodes that are node pairs, the linear numerical difference between the two graph nodes in the feature dimension and the nonlinear interaction difference between the feature dimensions are calculated. The baseline feature values, linear numerical differences, and nonlinear interaction differences are sequentially input into a preset neural difference network for multi-layer nonlinear transformation to obtain the difference vector. : In the formula, It is a neural differential network; This is an element-wise product.

[0079] In this embodiment of the invention, the updated graph nodes of the next propagation layer are calculated using the following formula: In the formula, Represents a node j In the l Current feature representation in layer graph propagation, This represents the (l+1)th layer updated feature representation obtained by aggregating the information passed from other associated nodes by node j; It is a multilayer perceptron.

[0080] In this embodiment of the invention, the graph node feature vectors are subjected to multi-scale hierarchical feature fusion processing according to a preset fusion network to obtain a global fused feature vector, including: Based on the consistency requirements of graph node feature vectors and feature semantics, attention-weighted aggregation processing is performed on nodes with similar semantic patterns to obtain multiple local aggregated features corresponding to different semantic groups; All local aggregated features are concatenated using both global average pooling and global max pooling to obtain a fused global feature vector.

[0081] In this embodiment of the invention, the graph node feature vector and the global fusion feature vector are input into a preset prediction network, and the total training loss is determined based on the prediction results, including: The global fusion feature vector is input into a preset prediction head network, and a unique and effective high-fidelity prediction result is output. The high-fidelity task loss is calculated based on the output result. Each graph node feature vector is input into a pre-defined auxiliary prediction network, and the auxiliary prediction result corresponding to the data source is output. The auxiliary supervision loss is calculated based on the output result. Based on the node pairs and their corresponding difference vectors, the cross-fidelity consistency loss and difference learning loss are calculated. The total training loss is calculated by weighting and summing the high-fidelity task loss, auxiliary supervision loss, cross-fidelity consistency loss, and differential learning loss according to preset loss weights.

[0082] It should be noted that the graph propagation-based multi-source data fusion prediction model construction device provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the graph propagation-based multi-source data fusion prediction model construction device and the graph propagation-based multi-source data fusion prediction model construction method embodiment provided in the above embodiments belong to the same concept, and the specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0083] Embodiments of this application also provide a computer device, please refer to... Figure 3 The computer device includes a processor and a memory, the memory storing at least one instruction, at least one program, code set, or instruction set, wherein at least one instruction, at least one program, code set, or instruction set is loaded and executed by the processor to implement the graph propagation-based multi-source data fusion prediction model construction method provided in the above-described method embodiments.

[0084] For ease of description, the above systems or devices are described separately as various modules or units based on their functions. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware components.

[0085] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, 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 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 various embodiments or some parts of the embodiments of this application.

[0086] Finally, it should be noted that in this document, relational terms such as first, second, third, and fourth are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0087] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for constructing a multi-source data fusion prediction model based on graph propagation, characterized in that, The method includes: A fully connected directed graph structure is established using the enhanced data node features of each data source as graph nodes and the information interaction path between any two enhanced data nodes as edges. The enhanced data node features are obtained by inputting multi-source data that has undergone mapping processing into a preset data augmentation network, including: inputting the multi-source data into a preset feature encoder to map and obtain latent space features. : in, For the k-th type of data source; For feature encoder; The latent space feature data is summed with a pre-defined learnable vector to obtain enhanced data node features. : in, For learnable embedding vectors; Based on the optimized bidirectional information propagation network, bidirectional information propagation and difference learning updates are performed on each node pair of the fully connected directed graph structure to obtain graph node feature vectors containing information from all data sources and differences in the structures of different data sources, including: The node pairs consisting of target graph nodes and source graph nodes are concatenated along the feature dimension to obtain a concatenated vector that fuses the dual-node feature context. The spliced ​​vector is input into a preset forward information flow attention network and a reverse information flow attention network, respectively, and the forward propagation information from the source graph node to the target graph node and the reverse calibration information from the target graph node to the source graph node are output in sequence. Based on the difference information of the node pair, the multi-dimensional difference features of the node pair are subjected to multi-layer nonlinear transformation to obtain a difference vector representing the structural differences between the two types of data sources. Based on the forward propagation information, reverse calibration information, and difference vector of all node pairs related to the current propagation layer and the graph node to be updated, multi-source information aggregation and residual update processing are performed on the feature vector of the graph node to be updated to obtain the updated graph node of the next propagation layer after fusion. Determine whether the number of iterations meets the preset threshold. If so, stop the iteration and output the graph node feature vector containing information from all data sources and the structural differences between different data sources. The graph node feature vectors are subjected to multi-scale hierarchical feature fusion processing according to a preset fusion network to obtain a global fusion feature vector. The graph node feature vectors and the global fusion feature vectors are input into a preset prediction network, and the total training loss is determined based on the prediction results, including: The global fusion feature vector is input into a preset prediction head network, and a unique and effective high-fidelity prediction result is output. The high-fidelity task loss is calculated based on the output result. Each graph node feature vector is input into a pre-defined auxiliary prediction network, and the auxiliary prediction result corresponding to the data source is output. The auxiliary supervision loss is calculated based on the output result. Based on the node pairs and their corresponding difference vectors, the cross-fidelity consistency loss and difference learning loss are calculated. The total training loss is calculated by weighting and summing the high-fidelity task loss, auxiliary supervision loss, cross-fidelity consistency loss, and differential learning loss according to preset loss weights. Based on the total loss, determine whether the prediction model, consisting of the data augmentation network, the two-way information propagation network, the fusion network, and the prediction network, meets the preset convergence condition. If it does, output the final prediction model; otherwise, iterate and update the parameters of each network in the prediction model based on the total loss until the model converges.

2. The method as described in claim 1, characterized in that, The positive propagation information It is calculated using the following formula: in, For feature splicing symbols; For source graph nodes; For the target graph nodes; It is a positive information flow attention network; The reverse calibration information It is calculated using the following formula: in, It is a reverse information flow attention network.

3. The method as described in claim 2, characterized in that, The step involves performing multi-layer nonlinear transformation processing on the multi-dimensional difference features of the node pairs based on their difference information to obtain a difference vector characterizing the structural differences between the two types of data sources, including: Based on the baseline eigenvalues ​​of two graph nodes that are node pairs, the linear numerical difference between the two graph nodes in the feature dimension and the nonlinear interaction difference between the feature dimensions are calculated. The baseline feature values, linear numerical differences, and nonlinear interaction differences are sequentially input into a preset neural difference network for multi-layer nonlinear transformation to obtain the difference vector. : In the formula, It is a neural differential network; This is an element-wise product.

4. The method as described in claim 3, characterized in that, The updated graph nodes of the next propagation layer are calculated using the following formula: In the formula, Represents a node j In the l Current feature representation in layer graph propagation, This represents the (l+1)th layer updated feature representation obtained by aggregating the information passed from other associated nodes by node j; It is a multilayer perceptron.

5. The method as described in claim 1, characterized in that, The graph node feature vectors are subjected to multi-scale hierarchical feature fusion processing based on a preset fusion network to obtain a global fused feature vector, including: Based on the consistency requirements of graph node feature vectors and feature semantics, attention-weighted aggregation processing is performed on nodes with similar semantic patterns to obtain multiple local aggregated features corresponding to different semantic groups; All local aggregated features are concatenated using both global average pooling and global max pooling to obtain a fused global feature vector.

6. A device for constructing a multi-source data fusion prediction model based on graph propagation, characterized in that, The apparatus, used in the method of any one of claims 1-5, comprises: A construction module is used to establish a fully connected directed graph structure with the enhanced data node features of each type of data source as graph nodes and the information interaction path between any two enhanced data nodes as edges; wherein, the enhanced data node features are obtained by inputting multi-source data that has undergone mapping processing into a preset data enhancement network; The propagation module is used to perform bidirectional information propagation and difference learning updates on each node pair of the fully connected directed graph structure according to the optimized bidirectional information propagation network, so as to obtain graph node feature vectors containing information from the full data source and differences in the structure of different data sources. The fusion module is used to perform multi-scale hierarchical feature fusion processing on the graph node feature vectors according to a preset fusion network to obtain a global fusion feature vector. The prediction module is used to input the graph node feature vector and the global fusion feature vector into a preset prediction network, and determine the total training loss based on the prediction results. The update module is used to determine whether the prediction model consisting of the data augmentation network, the bidirectional information propagation network, the fusion network, and the prediction network meets the preset convergence conditions based on the total loss. If so, the final prediction model is output; otherwise, the parameters of each network in the prediction model are iteratively updated based on the total loss until the model converges.

7. A computer device, characterized in that, The computer device includes a memory and a processor. The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory to implement the steps of the method according to any one of claims 1-5.

Images