A talent scientific and technological innovation portrait construction method based on multi-source heterogeneous data fusion

Abstract

The present application relates to the technical field of data retrieval and digital twinning, and discloses a talent scientific and technological innovation portrait construction method based on multi-source heterogeneous data fusion. In the digital twinning space, virtual sensor agents are deployed for structured projects, unstructured texts and time series behavior logs. By using a contrast learning framework and an orthogonal projection matrix, heterogeneous data features are decoupled into modal invariant features and modal specific features constrained in independent subspaces. A global inverted index is constructed based on the modal invariant features, and a local graph index is constructed based on the modal specific features. When a retrieval request is received, a candidate talent node is located through the global inverted index, and a physical data source virtual sensor agent is traced along the independent subspaces through the local graph index, and fusion features are returned. The present application blocks the dominant bias of single modal features and maintains the independent topological structure of each modal feature.

Description

Technical Field

[0001] This invention relates to the field of data retrieval and database structure digital twin technology, and discloses a method for constructing talent innovation profiles based on the fusion of multi-source heterogeneous data. Background Technology

[0002] Current technologies for constructing and retrieving talent and innovation profiles based on multi-source heterogeneous data typically extract features from heterogeneous data such as structured project data, unstructured text data, and time-series behavioral logs from the physical world, and then directly map them into the same latent space to generate a single fusion vector. When retrieving talent and innovation profiles, the system directly calculates the distance between the feature vector of the query request and the fusion vectors of various talents in the database, selecting nodes whose distances meet preset conditions as the search results. This conventional approach eliminates metric differences between different modalities through a unified latent space mapping, thereby enabling cross-modal data similarity calculation.

[0003] Based on the aforementioned existing technologies, forcibly mapping data from different modalities to the same latent space results in a single modality, with its massive data volume, dominating the fused vector and thus overshadowing the feature information of other modalities. When performing profile retrieval in the digital twin space, this feature coupling phenomenon causes the retrieval results to deviate from the multi-dimensional comprehensive evaluation objective. Furthermore, because the fusion process disrupts the independent topological structure of each modality's features, the retrieved fused vector cannot be back-linked to the specific physical data source, leading to a disconnect between the retrieval results within the digital twin system and the original data in the physical world. Summary of the Invention

[0004] The purpose of this invention is to provide a method for constructing a talent innovation profile based on the fusion of multi-source heterogeneous data, which can effectively solve the problems in the background technology mentioned above.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for constructing a talent innovation profile based on the fusion of multi-source heterogeneous data, the method comprising: deploying virtual sensor agents in a digital twin space for structured project data, unstructured text data, and time-series behavioral logs from the physical world respectively; The virtual sensor agent extracts features from the input multi-source heterogeneous data. The extracted features are separated into modality-invariant features and modality-specific features using a contrastive learning framework, and the modality-specific features are constrained in an independent subspace. A global inverted index is constructed based on the modality-invariant features, and a local graph index is constructed based on the modality-specific features and the independent subspace affiliation. When the digital twin space receives a talent innovation profile retrieval request, it locates the candidate talent node through the global inverted index, then traces the physical data source through the virtual sensor proxy along the independent subspace where the modality-specific features are located through the local graph index, and returns the fused features.

[0006] Preferably, before performing feature extraction on the input multi-source heterogeneous data through the virtual sensor agent, the method further includes: for the structured project data, the virtual sensor agent maps the project attribute fields to nodes and initial edges of a preset attribute graph structure to generate a first structured tensor; For the unstructured text data, the virtual sensor agent extracts the text sequence through word segmentation and inputs the text sequence into a pre-trained language model to extract the first text tensor; For the time-series behavior log, the virtual sensor agent segments the time-series behavior log according to a preset time window, converts the segmented time segments into frequency domain vectors, and generates a first time-series tensor. The first structured tensor, the first text tensor, and the first temporal tensor are used as input multi-source heterogeneous data.

[0007] Preferably, the step of using a contrastive learning framework to separate the extracted features into modality-invariant features and modality-specific features includes: constructing a positive sample pair from the first structured tensor, the first text tensor, and the first temporal tensor corresponding to the same talent node, and constructing a negative sample pair from the same type of tensor corresponding to different talent nodes; The positive sample pairs and the negative sample pairs are input into a dual-branch encoder network to extract modality mixture features; The modal mixing features are input into an orthogonal projection matrix set. By calculating the weighted sum of contrast loss and subspace orthogonal constraint loss, the modal mixing features are projected onto a common feature space and multiple independent subspaces. Modality-invariant features are generated in the common feature space, and modality-specific features are generated in the independent subspaces.

[0008] Preferably, the step of constructing a global inverted index based on the modality-invariant features includes: performing dimensionality reduction processing on the modality-invariant features in the common feature space, inputting the dimensionality-reduced modality-invariant features into a quantization hash network, and generating binary hash codes; Count the number of identical hash sequences in the binary hash code, and store multiple talent node identifiers corresponding to the same hash sequence in the same hash bucket; Establish a first-level mapping table between the binary hash code and the hash bucket, and a second-level inverted linked list between the hash bucket and the talent node identifier; When a new modality-invariant feature is generated, the new binary hash code is calculated and inserted into the corresponding hash bucket, and the secondary inverted linked list is updated.

[0009] Preferably, the step of constructing a local graph index based on the modality-specific features and the independent subspace affiliation includes: using the modality-specific features within the independent subspace as graph nodes, and calculating the cosine similarity between different graph nodes; Establish an edge between two graph nodes whose cosine similarity is greater than a preset similarity threshold, and use the cosine similarity as the initial weight of the edge; The physical data source category corresponding to the graph node is used as the attribute label of the graph node to construct a heterogeneous graph structure containing node features, node attribute labels and edge weights; Establish a bidirectional pointer mapping between the graph nodes in the heterogeneous graph structure and the talent node identifiers in the global inverted index.

[0010] Preferably, the step of locating candidate talent nodes through the global inverted index and then tracing the physical data source by virtual sensor proxy along the independent subspace where the modality-specific features are located through the local graph index includes: parsing the talent innovation profile retrieval request and converting the retrieval request into query-invariant features and query-specific features; Calculate the Hamming distance between the query invariant feature and each of the binary hash codes in the first-level mapping table, and identify the talent node in the hash bucket with the smallest Hamming distance as the candidate talent node; Through the bidirectional pointer mapping, the graph node corresponding to the candidate talent node is obtained in the heterogeneous graph structure. Based on the query-specific features and the attribute tags of the graph node, the target physical data source is locked by traversing along the connecting edges.

[0011] Preferably, after traversing along the connecting edges and locking the target physical data source, the method further includes: the virtual sensor agent receiving a source tracing decoding instruction for the target physical data source; If the target physical data source is the unstructured text data, the virtual sensor agent extracts the attention weight matrix of the modality-specific features in the locked graph nodes, maps the index positions of the top N weight values ​​in the attention weight matrix back to the word segmentation positions of the text sequence, and extracts the original text paragraphs corresponding to the word segmentation positions as the source tracing results. If the target physical data source is the time-series behavior log, the virtual sensor agent will input the locked modality-specific features into the inverse Fourier transform network to restore them to the time-series coordinate points within the preset time window.

[0012] Preferably, the weighted sum of the contrast loss and the subspace orthogonality constraint loss includes: calculating the subspace orthogonality constraint loss between the common feature space and each of the independent subspaces during each forward propagation process, and calculating the contrast loss of the positive sample pair within the common feature space; During backpropagation, the gradient matrix of the fully connected layer in the dual-branch encoder network is obtained; The subspace orthogonal constraint loss is applied to the gradient matrix with a penalty term, and the superimposed gradient matrix is ​​forcibly projected onto a tangent plane orthogonal to the basis vectors of the independent subspace using a projection operator. The weight parameters of the dual-branch encoder network are updated using the projected gradient matrix to prevent feature vectors in the independent subspaces from leaking into the common feature space.

[0013] Preferably, after the new binary hash code is inserted into the corresponding hash bucket, the method further includes: monitoring the number of talent node identifiers stored in the hash bucket; When the number exceeds a preset capacity threshold, all modality-invariant features in the hash bucket are extracted to form a feature set. Calculate the covariance matrix of the feature set, perform eigenvalue decomposition on the covariance matrix, and obtain the first principal component direction vector corresponding to the largest eigenvalue; All the binary hash codes in the hash bucket are divided along the first principal component direction vector, splitting the hash bucket into a first sub-hash bucket and a second sub-hash bucket, and the pointing relationship between the first-level mapping table and the second-level inverted linked list is updated.

[0014] Preferably, after constructing the heterogeneous graph structure including node features, node attribute labels and edge weights, the method further includes: when the physical world generates new multi-source heterogeneous data, the virtual sensor agent extracts the new modality-specific features; In the heterogeneous graph structure, locate all neighboring graph nodes that belong to the same independent subspace as the newly added modality's unique features; The newly added modality-specific features and the neighbor graph nodes are input into a graph attention network to calculate the attention coefficient of the newly added modality-specific features for each neighbor graph node; The initial weights of the edges are replaced with the attention coefficients, and new edges are added between the graph nodes corresponding to the new modality-specific features and the neighboring graph nodes in the heterogeneous graph structure to complete the incremental update of the local graph index.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention deploys virtual sensor agents in a digital twin space and utilizes a contrastive learning framework and orthogonal projection matrices to decouple the features of multi-source heterogeneous data into modality-invariant features and modality-specific features constrained within independent subspaces. A global inverted index is constructed based on the modality-invariant features, and a local graph index is constructed based on the modality-specific features. During retrieval, candidate nodes are located using the global inverted index, and the physical data source is traced along the independent subspaces using the local graph index. This mechanism prevents the dominant bias of a single modality feature during the fusion process, maintaining the independent topological structure of each modality feature.

[0016] 2. In the feature decoupling stage, the orthogonal constraint loss of the subspace is calculated, and the gradient matrix is ​​projected onto the orthogonal tangent plane using the projection operator during backpropagation, thus preventing the leakage of feature vectors from independent subspaces to the common feature space. In the source decoding stage, for text data, the attention weight matrix is ​​mapped back to the word segmentation position to extract the original paragraph; for time-series data, the inverse Fourier transform is used to restore the time-series coordinates, realizing a concrete restoration of the physical data source. In the index maintenance stage, when the number of nodes in the hash bucket exceeds a preset threshold, the hash bucket is split and updated using the covariance matrix and the direction vector of the first principal component; when new data enters, the attention coefficients are calculated through a graph attention network to update the edge weights of the local graph index, ensuring the index availability and topological accuracy of the system during continuous data writing. Attached Figure Description

[0017] Figure 1 This is the overall flowchart of the present invention; Figure 2 This is a flowchart of the multi-source heterogeneous data preprocessing and tensor generation process of the present invention; Figure 3 This is a flowchart illustrating the feature decoupling and comparative learning separation process of this invention. Figure 4 This is a flowchart illustrating the global inverted index construction and update process of this invention. Figure 5 This is a flowchart illustrating the construction and bidirectional mapping of the local graph index in this invention. Figure 6 This is a flowchart illustrating the retrieval request execution and physical data source tracing process of this invention. Detailed Implementation

[0018] This embodiment provides a mirror topology that maps all elements of the talent and science innovation management system in the physical world to a digital twin space. For the three core data sources generated in the physical world—structured project data, unstructured text data, and time-series behavioral logs—corresponding virtual sensor agents are deployed. Specifically, the digital twin space provides a computable and traceable virtual mapping carrier for the entire talent and science innovation process in the physical world. The virtual sensor agents are deployed at the data source access layer of the digital twin space. Each type of virtual sensor agent establishes a bidirectional data link with the mirror node of the corresponding data source in the physical world. This allows for unidirectional synchronous acquisition and feature extraction of raw data from the physical world, and reverse data tracing and restoration of raw data during the retrieval process.

[0019] Virtual sensor agents deployed for structured project data are bound to mirror nodes of the talent project management database, scientific and technological innovation achievement registration system, and intellectual property management platform in the physical world to achieve real-time synchronization and standardized processing of structured project attribute data. Virtual sensor agents deployed for unstructured text data are bound to mirror nodes of the academic paper database, project research and development report library, technical document storage system, and scientific and technological innovation defense material archiving node in the physical world to achieve the collection and semantic parsing of unstructured text content. Virtual sensor agents deployed for time-series behavior logs are bound to mirror nodes of the scientific and technological innovation platform operation log system, collaborative research and development behavior recording node, and intellectual property application time-series recording system in the physical world to achieve continuous collection and time-dimensional standardized processing of time-series behavior data.

[0020] The virtual sensor agent extracts features from the input multi-source heterogeneous data. A contrastive learning framework separates the extracted features into modality-invariant features and modality-specific features, constraining the modality-specific features into independent subspaces. After standardizing the raw data, the virtual sensor agent extracts initial features from the input data for the corresponding modality, generating initial feature tensors for that modality. All initial feature tensors are input into the contrastive learning framework deployed in the feature decoupling layer of the digital twin space. The contrastive learning framework decouples the input multimodal initial features into two types of feature components through orthogonal projection and constraint mechanisms of the feature space: one type is modality-invariant features shared across modalities, carrying core information about talent and scientific innovation attributes commonly reflected in different modalities; the other type is modality-specific features independent of each modality, carrying unique scientific and technological innovation behaviors and attribute information for that modality. During feature decoupling, the contrastive learning framework constrains the modality-specific features corresponding to different modalities into mutually orthogonal independent subspaces, ensuring no overlap or leakage of feature components between different independent subspaces.

[0021] A global inverted index is constructed based on modality-invariant features, while a local graph index is constructed based on modality-specific features and independent subspace affiliation. After feature decoupling, the index construction layer of the digital twin space constructs a two-level index structure based on the two types of decoupled features. The global inverted index is constructed based on modality-invariant features within the common feature space, with talent nodes as the smallest index unit. It establishes an inverted mapping relationship between the hash codes of modality-invariant features and talent node identifiers, used for rapid coarse screening and location of candidate talent nodes during the retrieval process. The local graph index is constructed based on modality-specific features within each independent subspace. It uses the modality-specific features of each modality as graph nodes and the similarity between features as edge weights to construct a heterogeneous graph structure. At the same time, it establishes a bidirectional mapping relationship between graph nodes and talent node identifiers in the global inverted index, used for refined screening based on modality-specific features and physical data source tracing during the retrieval process.

[0022] Please refer to the attached document. Figure 5 When the digital twin space receives a talent and innovation profile retrieval request, it locates candidate talent nodes through a global inverted index, and then traces the virtual sensor proxy of the physical data source along the independent subspace where the modality-specific features are located through a local graph index, and returns the fused features. After receiving the talent and innovation profile retrieval request from an external system, the retrieval service layer of the digital twin space first performs feature parsing on the retrieval request, generating query-invariant features and query-specific features consistent with the output format of the feature decoupling layer. The retrieval service layer first inputs the query-invariant features into the global inverted index, and quickly locates the set of candidate talent nodes that meet the core retrieval conditions through hash coding matching and distance calculation. After obtaining the set of candidate talent nodes, the retrieval service layer locates the corresponding modality graph nodes in the local graph index through a bidirectional mapping relationship. Based on the subspace affiliation of the query-specific features, it traverses along the graph structure edges in the corresponding independent subspace to complete the virtual sensor proxy tracing of the corresponding physical data source. Finally, it concatenates the modality-invariant features and modality-specific features locked by tracing to generate the fused features of the corresponding talent node and returns them.

[0023] Table 1 Virtual Sensor Agent Deployment and Function Configuration Table Structured project data virtual sensor agent Talent project management database, scientific and technological innovation achievement registration system, and mirror nodes of the intellectual property management platform Structured field cleaning, attribute graph structure mapping, and initial structured feature extraction. Incremental synchronous triggering access, synchronous updates when project data status changes. A two-dimensional structured tensor containing an attribute graph adjacency matrix and a node feature matrix. Unstructured text data virtual sensor agent Academic paper database, project research and development report database, technical document storage system, and mirrored nodes for archiving science and technology innovation defense materials. Text denoising, word segmentation, and semantic feature extraction Full synchronization plus incremental supplementation integration, feature updates are completed when new text archives are added. A one-dimensional text feature tensor with the same dimension as the output dimension of the pre-trained language model. Time-series behavior log virtual sensor agent The mirrored systems for the science and technology innovation platform operation log system, collaborative R&D behavior recording nodes, and intellectual property application timeline recording system. Timestamp standardization, time window segmentation, frequency domain feature transformation Streaming continuous access, batch feature extraction is completed according to a preset time period. A one-dimensional time-series frequency domain tensor, with dimensions matching the number of sampling points within a preset time window. This table clarifies the deployment configuration and core functional boundaries of the three types of virtual sensor agents in this embodiment. Each type of virtual sensor agent is only responsible for the access, preprocessing and initial feature extraction of the corresponding modality data source, ensuring that the access process of multi-source heterogeneous data is independent of each other, avoiding cross-interference of raw data of different modalities during the access stage, and providing independent data link support for the subsequent feature decoupling and tracing process.

[0024] This embodiment achieves independent access and feature extraction of multi-source heterogeneous data by deploying independent virtual sensor agents for three core data sources in the digital twin space. It decouples and separates multimodal features through a contrastive learning framework, constraining modality-specific features into mutually independent subspaces. Based on the two types of features after decoupling, a two-level index structure of global inverted index and local graph index is constructed respectively. During the retrieval process, the global inverted index is used to quickly locate candidate talent nodes, and then the local graph index is used to trace the source of physical data sources and return fused features. This blocks the cross-interference and dominant bias of different modal features during the fusion process, maintains the independent topological structure of each modality feature, and realizes the bidirectional association between fused features in the digital twin space and the original data sources in the physical world.

[0025] Please refer to the attached document. Figure 1 In a preferred embodiment, before extracting features from the input multi-source heterogeneous data, the virtual sensor agent first performs standardized preprocessing on the original data of the corresponding modality to generate an input tensor in a unified format. For structured project data, the virtual sensor agent first performs field cleaning and normalization on the original structured data, extracts project attribute fields, including but not limited to project leader identification, project level, project initiation time, project completion time, project outcome type, outcome transformation status, number of intellectual property rights, and number of academic papers published. The extracted project attribute fields are mapped to nodes and initial edges of a preset attribute graph structure to generate a first structured tensor.

[0026] In the preset attribute graph structure, talent nodes are used as core nodes, and project attribute fields are used as attribute nodes. Initial edges are established between the core node and each attribute node, and the weights of the initial edges are assigned based on the preset weight coefficients corresponding to the attribute fields. The first structured tensor is a two-dimensional tensor generated by concatenating the adjacency matrix of the preset attribute graph structure and the node feature matrix. The rows of the tensor correspond to the number of nodes, and the columns correspond to the node feature dimensions. For unstructured text data, the virtual sensor agent first performs noise reduction processing on the original text data, removing format symbols, invalid spaces, and non-semantic characters. Then, it extracts the text sequence through word segmentation. The word segmentation is performed using a word segmenter matched with the pre-trained language model. The length of the segmented text sequence is uniformly padded to the preset input length of the pre-trained language model. The padded text sequence is then input into the pre-trained language model to extract the sentence vector features corresponding to the text sequence, generating the first text tensor. The first text tensor is a one-dimensional feature tensor with the same dimension as the output dimension of the pre-trained language model.

[0027] For time-series behavior logs, the virtual sensor agent first performs timestamp standardization on the original logs, converting all timestamps of time-series behaviors into continuous time coordinates under a preset time base. The standardized logs are then segmented according to preset time windows, with each time window corresponding to a behavior sequence segment. These segmented time segments are then transformed into frequency domain vectors using Discrete Fourier Transform, generating a first time-series tensor. This first time-series tensor is a one-dimensional feature tensor with the same dimension as the frequency domain vector corresponding to the preset time window. The virtual sensor agent uses the generated first structured tensor, first text tensor, and first time-series tensor as input to multi-source heterogeneous data, feeding them into a contrastive learning framework for feature decoupling processing.

[0028] Please refer to the attached document. Figure 2 When using a contrastive learning framework to separate extracted features into modality-invariant features and modality-specific features, a sample pair set is first constructed based on the input multi-source tensors. The first structured tensor, first text tensor, and first temporal tensor corresponding to the same talent node are paired to construct positive sample pairs. The two tensors in a positive sample pair correspond to different modal features of the same talent node, sharing the same core talent innovation attributes. Similarly, the tensors of the same type corresponding to different talent nodes are paired to construct negative sample pairs. The two tensors in a negative sample pair correspond to the same modal feature of different talent nodes, possessing the same modal structure but corresponding to different core talent innovation attributes. The sample pair set contains all positive and negative sample pairs. The sample pair set is divided into multiple training batches according to a preset batch size and input into a dual-branch encoder network for feature extraction.

[0029] The dual-branch encoder network comprises two structurally identical encoder branches. Each encoder branch consists of three fully connected layers and a layer normalization module. The input dimension of the first fully connected layer matches the dimension of the input tensor, and the output dimension is a preset intermediate dimension. The input and output dimensions of the second fully connected layer are both preset intermediate dimensions. The input dimension of the third fully connected layer is a preset intermediate dimension, and the output dimension is a preset feature dimension. The layer normalization module is set at the output of each fully connected layer to normalize the output features, eliminating the interference of feature amplitude differences on the subsequent projection process. The two tensors in a positive sample pair are input into the two branches of the dual-branch encoder network, respectively, and the output is a modal mixture feature with two consistent dimensions. Similarly, the two tensors in a negative sample pair are input into the two branches of the dual-branch encoder network, respectively, and the output is a modal mixture feature with two consistent dimensions. All output modal mixture features are input into an orthogonal projection matrix group to complete the projection and decoupling of the feature space.

[0030] The orthogonal projection matrix set comprises a common projection matrix and multiple independent subspace projection matrices. The common projection matrix corresponds to a common feature space, and each independent subspace projection matrix corresponds to an independent subspace of a mode. The number of independent subspace projection matrices is consistent with the number of modes. Both the common projection matrix and each independent subspace projection matrix are orthogonal matrices, satisfying the constraint that the product of the matrix's transpose and itself is an identity matrix. Modal mixture features are input into the common projection matrix and the corresponding independent subspace projection matrices. By calculating the weighted sum of the contrastive loss and the subspace orthogonality constraint loss, the modal mixture features are projected onto the common feature space and the multiple independent subspaces. Mode-invariant features are generated in the common feature space, and mode-specific features corresponding to the respective modes are generated in the independent subspaces.

[0031] The contrast loss is calculated using the following expression: ; in, To compare the losses, The number of sample pairs within the training batch. This is the cosine similarity calculation function. For the first The encoder output feature corresponds to the first tensor in each positive sample pair. For the first The encoder output feature corresponds to the second tensor in each positive sample pair. For the first The encoder output features corresponding to the tensors in each negative sample pair This is a preset temperature coefficient used to adjust the sharpness of the similarity distribution.

[0032] The subspace orthogonal constraint loss is calculated using the following expression: ; in, Subspace orthogonal constraint loss, For the number of modes, For the common projection matrix, For the first The independent subspace projection matrix corresponding to each mode. For the Frobenius norm, Used to calculate the common feature space and the first The orthogonality deviation between independent subspaces is such that when two spaces are completely orthogonal, all elements of the matrix are 0, and the corresponding orthogonality constraint loss is 0.

[0033] The total loss function is a weighted sum of the contrast loss and the subspace orthogonality constraint loss, and is calculated using the following expression: ; in, For the total loss function, The preset orthogonal constraint loss weight coefficient is used to balance the proportion of contrast loss and subspace orthogonal constraint loss in the total loss.

[0034] During each forward propagation, the subspace orthogonality constraint loss between the common feature space and each independent subspace is calculated, as is the contrastive loss of positive sample pairs within the common feature space. The total loss function is obtained by weighted summing the contrastive loss and the subspace orthogonality constraint loss. During backpropagation, the gradient matrix of the fully connected layers in the bi-branch encoder network is calculated based on the total loss function. The gradient matrix contains the partial derivatives of the total loss function with respect to the weight parameters of the fully connected layers.

[0035] The subspace orthogonality constraint loss is superimposed with a penalty term on the gradient matrix. A projection operator then forces the superimposed gradient matrix onto a tangent plane orthogonal to the basis vectors of the independent subspaces. The projected gradient matrix contains only components aligned with the basis vectors of the common feature space, eliminating the influence of independent subspace feature components on the update of the common feature space parameters. The projected gradient matrix is ​​then used to update the weight parameters of the bi-branch encoder network, preventing feature vectors in the independent subspaces from leaking into the common feature space. This ensures that mode-specific features are completely constrained within their respective independent subspaces, achieving complete decoupling from mode-invariant features.

[0036] The gradient projection is calculated using the following expression: ; in, The gradient matrix after projection. This is the original gradient matrix after adding the penalty term. For the first The independent subspace projection matrix corresponding to each mode. For the first The projection operator for each independent subspace completely removes the components in the original gradient matrix that are parallel to the basis vectors of the independent subspaces, and retains only the components that are orthogonal to the basis vectors of the independent subspaces, ensuring that the parameter update process does not cause the features in the independent subspaces to leak into the common feature space.

[0037] Table 2. Parameter Configuration Table for Multi-Source Heterogeneous Data Preprocessing and Feature Extraction Structured project data Field cleaning, attribute graph mapping, concatenation of adjacency matrix and feature matrix Original structured project field set, unique identifier for talent nodes (Total number of nodes × Feature dimension) Two-dimensional tensor Attribute graph preset node type, initial edge weight coefficient, node feature dimension Unstructured text data Text denoising, word segmentation, sequence imputation, and pre-trained language model encoding Original unstructured text content, unique text identifier (Pre-trained model output dimension) One-dimensional tensor Tokenizer type, maximum sequence length, and pre-trained language model weights Time-series behavior log Timestamp standardization, time window segmentation, discrete Fourier transform Original time-series behavior logs, timestamp baseline values (Frequency domain vector dimension) One-dimensional tensor Time window length, number of sampling points, number of Fourier transform points This table clarifies the preprocessing flow, input and output parameters, and fixed configurations for the three types of modal data in this embodiment. It ensures that the original data of different modalities are preprocessed to generate input tensors with uniform format and matching dimensions, providing standardized input for feature extraction and subsequent feature decoupling processes of the dual-branch encoder network. At the same time, the setting of fixed parameters ensures the consistency and reproducibility of the preprocessing process.

[0038] This embodiment designs corresponding preprocessing procedures for three different types of raw data, generates standardized input tensors, provides training samples for the contrastive learning framework by constructing positive and negative sample pairs, extracts modality mixture features based on a dual-branch encoder network, and decouples the modality mixture features into modality-invariant features in a common feature space and modality-specific features in independent subspaces by constraining the orthogonal projection matrix group and the total loss function. During backpropagation, the gradient matrix is ​​orthogonally constrained by the projection operator, which blocks the leakage of features in independent subspaces to the common feature space, achieves complete decoupling of different modality features, and avoids cross-interference and dominant bias between different modality features.

[0039] Please refer to the attached document. Figure 3 In a preferred embodiment, after decoupling features to obtain modality-invariant features and modality-specific features, a global inverted index and a local graph index are constructed respectively. When constructing the global inverted index based on modality-invariant features, the modality-invariant features in the common feature space are first subjected to dimensionality reduction processing. The dimensionality reduction processing is completed using principal component analysis, extracting principal component components whose cumulative contribution rate reaches a preset threshold from the modality-invariant features to generate dimensionality-reduced modality-invariant features. The dimension of the dimensionality-reduced modality-invariant features is consistent with the preset hash code length. The dimensionality-reduced modality-invariant features are input into a quantization hash network to generate binary hash codes. The quantization hash network consists of two fully connected layers and a symbolic activation function. The input dimension of the first fully connected layer is consistent with the dimension of the dimensionality-reduced modality-invariant features, and the output dimension is consistent with the preset hash code length. The input and output dimensions of the second fully connected layer are both the preset hash code length. The symbolic activation function is set at the output of the second fully connected layer, converting the continuous value features output by the fully connected layer into binary sequences with values ​​of +1 or -1, and then mapping the binary sequences into binary hash codes in 0-1 format.

[0040] The number of identical hash sequences in binary hash codes is counted. Multiple talent node identifiers corresponding to the same hash sequence are stored in the same hash bucket. Each hash bucket corresponds to a unique binary hash sequence, and the talent node identifiers stored in the hash bucket all have the same binary hash code, corresponding to similar modality-invariant features. A first-level mapping table between binary hash codes and hash buckets is established, where the key of the first-level mapping table is the binary hash sequence, and the value is a pointer to the storage address of the corresponding hash bucket. A second-level inverted linked list between hash buckets and talent node identifiers is established, where the head node of the second-level inverted linked list is the storage address of the hash bucket, and each node in the linked list corresponds to a talent node identifier. The linked list nodes are arranged in order of the talent node identifiers' entry time. When a new modality-invariant feature is generated, the same dimensionality reduction and quantized hash encoding process is followed to calculate the new binary hash code. The corresponding hash bucket is located through the first-level mapping table, and the new talent node identifier is inserted into the tail of the corresponding hash bucket's second-level inverted linked list, completing the update of the second-level inverted linked list.

[0041] The Hamming distance between two binary hash codes is calculated using the following expression: ; in, The Hamming distance between two binary hash codes. The length of the binary hash code, To query the hash code of the first Bit binary value, For the hash code to be matched The binary value, Hamming distance, ranges from 0 to... The smaller the value, the higher the feature similarity between the two hash codes.

[0042] After inserting the new binary hash code into the corresponding hash bucket, the number of talent node identifiers stored in the hash bucket is monitored in real time. When the number of talent node identifiers stored in the hash bucket exceeds a preset capacity threshold, the modality-invariant features corresponding to all talent nodes in the hash bucket are extracted to form a feature set. The covariance matrix of the feature set is calculated, and eigenvalue decomposition is performed on the covariance matrix to obtain the first principal component direction vector corresponding to the largest eigenvalue. The first principal component direction vector is the direction of the feature component with the largest variance in the feature set, representing the direction of the largest difference among all modality-invariant features in the hash bucket.

[0043] All binary hash codes within the hash bucket are divided along the first principal component direction vector. Talent node identifiers corresponding to binary hash codes whose feature projection values ​​are greater than the mean of the first principal component direction vector are assigned to the first sub-hash bucket, while those corresponding to binary hash codes whose feature projection values ​​are less than or equal to the mean of the first principal component direction vector are assigned to the second sub-hash bucket, thus completing the hash bucket split. After the hash bucket split, the storage address pointers corresponding to the original binary hash codes in the first-level mapping table are updated, the storage entries of the original hash buckets are deleted, and new binary hash code and storage address pointer entries corresponding to the first and second sub-hash buckets are added. Simultaneously, the pointing relationships of the second-level inverted linked lists are updated, establishing associations between the two sub-hash buckets and their corresponding talent node identifier linked lists.

[0044] The covariance matrix of the feature set is calculated using the following expression: ; in, Let be the covariance matrix of the feature set. This represents the number of talent nodes within the hash bucket. For the first The dimensionality-reduced mode-invariant feature vectors corresponding to each talent node. The mean vector of all eigenvectors in the feature set is given. The dimension of the covariance matrix is ​​the same as the dimension of the mode-invariant features after dimensionality reduction. The elements in the matrix correspond to the covariance between the two feature dimensions.

[0045] The eigenvalue decomposition of the covariance matrix is ​​expressed by the following expression: ; in, This is the eigenvector matrix of the covariance matrix, where each column of the matrix corresponds to an eigenvector. It is a diagonal matrix, and the elements on the diagonal are the eigenvalues ​​of the covariance matrix. The eigenvalues ​​are arranged in descending order, the first diagonal element is the largest eigenvalue, and the corresponding first column eigenvector is the first principal component direction vector.

[0046] Please refer to the attached document. Figure 4When constructing a local graph index based on modality-specific features and independent subspace affiliation, modality-specific features within independent subspaces are used as graph nodes. Each graph node corresponds to a unique talent node and a unique modality category, and the feature vector of the graph node is the modality-specific feature vector within the corresponding independent subspace. Cosine similarity is calculated between different graph nodes, measuring the degree of similarity between the modality-specific features corresponding to two graph nodes. Two graph nodes with a cosine similarity greater than a preset similarity threshold are connected by an undirected edge, using the calculated cosine similarity as the initial weight of the edge. The physical data source category corresponding to a graph node is used as the attribute label of the graph node, corresponding to the modality category, to identify the original physical data source type corresponding to the graph node.

[0047] Based on all graph nodes, node attribute labels, edges, and edge weights, a heterogeneous graph structure is constructed, incorporating node features, node attribute labels, and edge weights. Graph nodes with different attribute labels correspond to different independent subspaces within this heterogeneous structure, while graph nodes with the same attribute label reside in the same independent subspace. A bidirectional pointer mapping is established between graph nodes in the heterogeneous graph structure and talent node identifiers in the global inverted index. Each graph node stores the address pointer of its corresponding talent node identifier in the global inverted index, and each talent node identifier in the global inverted index stores the address pointers of its corresponding modal graph nodes within the heterogeneous graph structure, thus achieving a bidirectional association between the global inverted index and the local graph index.

[0048] The cosine similarity between features of two graph nodes is calculated using the following expression: ; in, Cosine similarity between modality-specific features corresponding to two graph nodes. For the first The feature vector of each graph node For the first The feature vector of each graph node The dot product of two eigenvectors. and These are the L2 norms of the two feature vectors, and the cosine similarity ranges from -1 to 1. The closer the value is to 1, the higher the similarity between the two feature vectors.

[0049] After constructing a heterogeneous graph structure that includes node features, node attribute labels, and edge weights, when new multi-source heterogeneous data is generated in the physical world, the corresponding virtual sensor agent completes the preprocessing, feature extraction, and decoupling of the new data, extracting new modality-specific features. Each new modality-specific feature corresponds to a unique talent node and a unique modality category, possessing a clear and independent subspace affiliation. Within the heterogeneous graph structure, all neighboring graph nodes belonging to the same independent subspace as the new modality-specific features are located. These neighboring graph nodes are all existing graph nodes that share the same attribute labels as the new modality-specific features.

[0050] The newly added modality-specific features and the feature vectors of all neighboring graph nodes are input into the graph attention network. The attention coefficient of the newly added modality-specific features for each neighboring graph node is calculated. This attention coefficient measures the degree of association between the newly added modality-specific features and their corresponding neighboring graph nodes. The calculated attention coefficients are used to replace the initial weights of the edges, and the graph nodes corresponding to the newly added modality-specific features are added to the heterogeneous graph structure. Simultaneously, new edges are established between the new graph nodes and neighboring graph nodes whose attention coefficients are greater than a preset threshold. The weights of these new edges are the corresponding calculated attention coefficients, completing the incremental update of the local graph index.

[0051] The graph attention coefficient is calculated using the following expression: ; in, To add graph nodes For neighbor graph nodes Attention coefficient This is a learnable linear transformation weight matrix used to perform a linear transformation on the input feature vector. This is a feature vector concatenation operation. This represents the weight vector of a single-layer feedforward neural network. It is a non-linear activation function. To add graph nodes The attention coefficients of all neighboring graph nodes are normalized using the softmax function, and the sum of the attention coefficients of all neighboring graph nodes is 1.

[0052] Table 3. Two-level index structure configuration and update rules Global Inverted Index First-level mapping table, hash bucket, second-level inverted linked list Hash code length, hash bucket capacity threshold, principal component cumulative contribution rate threshold Based on the construction of quantized hash codes with modality invariant features, talent nodes with the same hash code are stored in the same hash bucket, establishing a two-level mapping relationship. When a new feature is added, it is inserted into the second-level inverted linked list of the corresponding hash bucket; when the hash bucket is overcapacity, it is split based on principal component analysis, and the mapping table and linked list pointers are updated. Local Image Index Heterogeneous graph nodes, undirected edges, bidirectional pointer mapping Similarity threshold, attention coefficient threshold, number of independent subspaces Using modality-specific features as graph nodes and feature similarity as edge weights, a heterogeneous graph is constructed, establishing a bidirectional pointer mapping with the global inverted index. When adding new features, attention coefficients are calculated using a graph attention network, edge weights are updated, and new graph nodes are connected to edges that meet the thresholds to complete the incremental update. This table clarifies the structural configuration, construction rules, and update mechanism of the global inverted index and the local graph index in this embodiment. It clearly defines the functional boundaries and collaborative relationship between the two levels of indexes. The global inverted index is responsible for the rapid coarse screening of candidate talent nodes, while the local graph index is responsible for fine-grained screening and data source tracing. The two levels of indexes work together through bidirectional pointer mapping. At the same time, the clear update rules ensure the availability and accuracy of the indexes during continuous data writing.

[0053] This embodiment constructs a global inverted index with a two-level mapping structure based on decoupled modality-invariant features. It implements binary encoding of features through a quantized hash network and ensures the retrieval efficiency of the global inverted index in massive data scenarios through a hash bucket splitting mechanism. Based on modality-specific features and independent subspace affiliation, it constructs a local graph index with a heterogeneous graph structure. It implements incremental updates of the local graph index through a graph attention network and achieves bidirectional association between the global inverted index and the local graph index through bidirectional pointer mapping, providing index support for candidate node location and data source tracing during the retrieval process.

[0054] In a preferred embodiment, when the digital twin space receives a talent and innovation profile retrieval request, it completes the retrieval and tracing process based on the constructed global inverted index and local graph index. When locating candidate talent nodes using the global inverted index and then tracing the physical data source via virtual sensor proxies along the independent subspaces containing modality-specific features using the local graph index, the received talent and innovation profile retrieval request is first parsed. The retrieval conditions and constraint parameters in the retrieval request are extracted, and the retrieval conditions are transformed into query-invariant features and query-specific features consistent with the output format of the feature decoupling layer. Query-invariant features correspond to the core talent and innovation attribute conditions in the retrieval request, carrying the general retrieval requirements across modalities; query-specific features correspond to the modality-specific constraint conditions in the retrieval request, possessing a clear independent subspace affiliation and corresponding to a specific physical data source category.

[0055] Calculate the query binary hash code corresponding to the query invariant feature, and calculate the Hamming distance between the query binary hash code and each binary hash code in the first-level mapping table. Sort the hash buckets in ascending order of Hamming distance, and select all talent nodes in the hash buckets with the smallest predetermined number of Hamming distances as candidate talent nodes, generating a set of candidate talent nodes. All talent nodes in the candidate talent node set possess modality-invariant features that match the core conditions of the search request, thus satisfying the core requirements of the search.

[0056] By using bidirectional pointer mapping between the global inverted index and the local graph index, the modal graph nodes corresponding to each talent node in the candidate talent node set are obtained in the heterogeneous graph structure. Based on the attribute labels corresponding to the query-specific features and the independent subspace affiliation, the graph nodes in the corresponding independent subspace are locked. The graph structure edges in the independent subspace are traversed. During the traversal, the similarity between the query-specific features and the graph node features is used for filtering, and the graph node with the highest matching degree with the query-specific features is locked. Based on the binding relationship between the attribute labels of the graph nodes and the virtual sensor agent, the target physical data source is locked.

[0057] Please refer to the attached document. Figure 6After traversing along the edges and locking onto the target physical data source, the retrieval service layer of the digital twin space generates a source tracing decoding instruction for the target physical data source. This instruction is then sent to the corresponding bound virtual sensor agent. Upon receiving the instruction, the virtual sensor agent, based on the modality-specific features corresponding to the locked graph nodes, completes the restoration of the original physical data source. If the target physical data source is unstructured text data, the virtual sensor agent extracts the attention weight matrix output by the pre-trained language model corresponding to the modality-specific features in the locked graph nodes. Each element in the attention weight matrix corresponds to the weight value of each word segmentation position in the text sequence. A higher weight value indicates a higher contribution of the text content at that word segmentation position to the modality-specific features. The index positions of the top N weight values ​​in the attention weight matrix are mapped back to the word segmentation positions in the text sequence, and the original text segments corresponding to the word segmentation positions are extracted and returned as the source tracing result. N is the preset number of source text segments. If the target physical data source is a time-series behavior log, the virtual sensor agent will input the frequency domain vector corresponding to the locked modal-specific features into the inverse Fourier transform network to restore it into time-series coordinate points within a preset time window. The time-series coordinate points contain timestamps and corresponding behavior event identifiers. Based on the restored time-series coordinate points, a continuous time-series behavior sequence is generated and returned as the tracing result.

[0058] The inverse Fourier transform of a time-series frequency domain vector is calculated using the following expression: ; in, For the restored first The magnitude of each time-series coordinate point The number of sampling points within the preset time window, The first characteristic corresponding to the locked modality Each frequency domain vector component The inverse Fourier transform, using the imaginary unit, restores the frequency domain vector to time-series coordinates in the time domain, thus restoring the original data of the time-series behavior log.

[0059] The index mapping calculation for text attention weights uses the following expression: ; in, This is the set of the top N word segmentation index positions in the sorted sequence. This is the one-dimensional weight vector corresponding to the attention weight matrix. This is a descending sorting function that outputs the index positions corresponding to the weight values ​​from largest to smallest. To obtain the top N index positions of the sorted results, the corresponding original text segments are extracted by mapping this set of index positions back to the word segmentation positions of the original text sequence.

[0060] After tracing the target physical data source and restoring the original data, the modality-invariant features of the corresponding talent node and the locked modality-specific features are spliced ​​together to generate the fusion feature of the talent node. The fusion feature contains the core attribute information of the talent's scientific and technological innovation and the exclusive feature information of the corresponding modality. At the same time, the content of the original physical data source restored by tracing is associated and bound with the fusion feature and returned to the initiating end of the retrieval request.

[0061] Table 4. Mapping Table of Retrieval and Source Tracing Process Nodes Search Request Parsing Talent Innovation Profile Search Request none Extract search criteria and constraint parameters, and convert them into query-invariant features and query-specific features. Standardized query feature set Candidate node coarse screening Query invariant features Global Inverted Index Calculate the Hamming distance between the query hash code and the hash codes in the database, and extract the talent nodes in the hash bucket with the smallest Hamming distance. Candidate Talent Node Set Refined screening and traceability Query unique features and candidate talent node set Local Image Index By locating the corresponding graph node through bidirectional pointer mapping and traversing the edges along independent subspaces, the target physical data source is locked. Target graph nodes and corresponding physical data source identifiers Source tracing and decoding Target graph nodes, source tracing decoding instructions Virtual sensor agent Based on the target physical data source type, perform text paragraph extraction or time-series coordinate reconstruction respectively. Original data source tracing results Results returned Modality invariant features, locked mode-specific features, and tracing results none Feature concatenation is linked with the source tracing results to generate the final returned content. The results of integrating the characteristics and correlations of talent innovation profiles This table clarifies the mapping relationship of the entire process nodes of talent and science and technology innovation profile retrieval and tracing in this embodiment. It clearly defines the input and output, corresponding index structure and core processing logic of each process stage, ensuring that the retrieval process first completes efficient candidate node coarse screening through global inverted index, and then completes refined tracing processing through local graph index, thus achieving a balance between retrieval efficiency and tracing accuracy.

[0062] This embodiment addresses talent and innovation profile retrieval requests by parsing and transforming retrieval features. It achieves rapid coarse screening and positioning of candidate talent nodes through a global inverted index, locks the target physical data source through heterogeneous graph structure traversal of the local graph index, designs corresponding source tracing and decoding mechanisms for target physical data sources of different modalities, realizes the restoration of original data from unstructured text data and time-series behavior logs, and finally returns the fused features associated with the original data source. This realizes the reverse positioning and concrete restoration of fused features in the digital twin space to the original data source in the physical world.

Claims

1. A method for constructing a talent innovation profile based on multi-source heterogeneous data fusion, characterized in that, The method includes: deploying virtual sensor agents in a digital twin space for structured project data, unstructured text data, and time-series behavior logs from the physical world, respectively; The virtual sensor agent extracts features from the input multi-source heterogeneous data. The extracted features are separated into modality-invariant features and modality-specific features using a contrastive learning framework, and the modality-specific features are constrained in an independent subspace. A global inverted index is constructed based on the modality-invariant features, and a local graph index is constructed based on the modality-specific features and the independent subspace affiliation. When the digital twin space receives a talent innovation profile retrieval request, it locates the candidate talent node through the global inverted index, then traces the physical data source through the virtual sensor proxy along the independent subspace where the modality-specific features are located through the local graph index, and returns the fused features.

2. The method for constructing a talent innovation profile based on multi-source heterogeneous data fusion according to claim 1, characterized in that, Before performing feature extraction on the input multi-source heterogeneous data through the virtual sensor agent, the method further includes: for the structured project data, the virtual sensor agent maps the project attribute fields to nodes and initial edges of a preset attribute graph structure to generate a first structured tensor; For the unstructured text data, the virtual sensor agent extracts the text sequence through word segmentation and inputs the text sequence into a pre-trained language model to extract the first text tensor; For the time-series behavior log, the virtual sensor agent segments the time-series behavior log according to a preset time window, converts the segmented time segments into frequency domain vectors, and generates a first time-series tensor. The first structured tensor, the first text tensor, and the first temporal tensor are used as input multi-source heterogeneous data.

3. The method for constructing a talent innovation profile based on multi-source heterogeneous data fusion according to claim 2, characterized in that, The method of using a contrastive learning framework to separate the extracted features into modality-invariant features and modality-specific features includes: constructing a positive sample pair by constructing the first structured tensor, the first text tensor, and the first temporal tensor corresponding to the same talent node, and constructing a negative sample pair by constructing the same type of tensor corresponding to different talent nodes; The positive sample pairs and the negative sample pairs are input into a dual-branch encoder network to extract modality mixture features; The modal mixing features are input into an orthogonal projection matrix set. By calculating the weighted sum of contrast loss and subspace orthogonal constraint loss, the modal mixing features are projected onto a common feature space and multiple independent subspaces. Modality-invariant features are generated in the common feature space, and modality-specific features are generated in the independent subspaces.

4. The method for constructing a talent innovation profile based on multi-source heterogeneous data fusion according to claim 3, characterized in that, The construction of a global inverted index based on the modality-invariant features includes: performing dimensionality reduction on the modality-invariant features in the common feature space, inputting the dimensionality-reduced modality-invariant features into a quantization hash network, and generating binary hash codes; Count the number of identical hash sequences in the binary hash code, and store multiple talent node identifiers corresponding to the same hash sequence in the same hash bucket; Establish a first-level mapping table between the binary hash code and the hash bucket, and a second-level inverted linked list between the hash bucket and the talent node identifier; When a new modality-invariant feature is generated, the new binary hash code is calculated and inserted into the corresponding hash bucket, and the secondary inverted linked list is updated.

5. The method for constructing a talent innovation profile based on multi-source heterogeneous data fusion according to claim 4, characterized in that, The construction of a local graph index based on the modality-specific features and the independent subspace affiliation includes: using the modality-specific features within the independent subspace as graph nodes, and calculating the cosine similarity between different graph nodes; Establish an edge between two graph nodes whose cosine similarity is greater than a preset similarity threshold, and use the cosine similarity as the initial weight of the edge; The physical data source category corresponding to the graph node is used as the attribute label of the graph node to construct a heterogeneous graph structure containing node features, node attribute labels and edge weights; Establish a bidirectional pointer mapping between the graph nodes in the heterogeneous graph structure and the talent node identifiers in the global inverted index.

6. The method for constructing a talent innovation profile based on multi-source heterogeneous data fusion according to claim 5, characterized in that, The step of locating candidate talent nodes through the global inverted index and then tracing the physical data source by virtual sensor proxy along the independent subspace where the modality-specific features are located through the local graph index includes: parsing the talent innovation profile retrieval request and converting the retrieval request into query-invariant features and query-specific features; Calculate the Hamming distance between the query invariant feature and each of the binary hash codes in the first-level mapping table, and identify the talent node in the hash bucket with the smallest Hamming distance as the candidate talent node; Through the bidirectional pointer mapping, the graph node corresponding to the candidate talent node is obtained in the heterogeneous graph structure. Based on the query-specific features and the attribute tags of the graph node, the target physical data source is locked by traversing along the connecting edges.

7. The method for constructing a talent innovation profile based on multi-source heterogeneous data fusion according to claim 6, characterized in that, After traversing along the connecting edges and locking the target physical data source, the method further includes: the virtual sensor agent receiving a source tracing and decoding instruction for the target physical data source; If the target physical data source is the unstructured text data, the virtual sensor agent extracts the attention weight matrix of the modality-specific features in the locked graph nodes, maps the index positions of the top N weight values ​​in the attention weight matrix back to the word segmentation positions of the text sequence, and extracts the original text paragraphs corresponding to the word segmentation positions as the source tracing results. If the target physical data source is the time-series behavior log, the virtual sensor agent will input the locked modality-specific features into the inverse Fourier transform network to restore them to the time-series coordinate points within the preset time window.

8. The method for constructing a talent innovation profile based on multi-source heterogeneous data fusion according to claim 3, characterized in that, The calculation of the weighted sum of the contrast loss and the subspace orthogonality constraint loss includes: calculating the subspace orthogonality constraint loss between the common feature space and each of the independent subspaces in each forward propagation process, and calculating the contrast loss of the positive sample pair in the common feature space. During backpropagation, the gradient matrix of the fully connected layer in the dual-branch encoder network is obtained; The subspace orthogonal constraint loss is applied to the gradient matrix with a penalty term, and the superimposed gradient matrix is ​​forcibly projected onto a tangent plane orthogonal to the basis vectors of the independent subspace using a projection operator. The weight parameters of the dual-branch encoder network are updated using the projected gradient matrix to prevent feature vectors in the independent subspaces from leaking into the common feature space.

9. The method for constructing a talent innovation profile based on multi-source heterogeneous data fusion according to claim 4, characterized in that, After the new binary hash code is inserted into the corresponding hash bucket, the method further includes: monitoring the number of talent node identifiers stored in the hash bucket; When the number exceeds a preset capacity threshold, all modality-invariant features in the hash bucket are extracted to form a feature set. Calculate the covariance matrix of the feature set, perform eigenvalue decomposition on the covariance matrix, and obtain the first principal component direction vector corresponding to the largest eigenvalue; All the binary hash codes in the hash bucket are divided along the first principal component direction vector, splitting the hash bucket into a first sub-hash bucket and a second sub-hash bucket, and the pointing relationship between the first-level mapping table and the second-level inverted linked list is updated.

10. The method for constructing a talent innovation profile based on multi-source heterogeneous data fusion according to claim 5, characterized in that, After constructing the heterogeneous graph structure containing node features, node attribute labels, and edge weights, the method further includes: when the physical world generates new multi-source heterogeneous data, the virtual sensor agent extracts the new modality-specific features; In the heterogeneous graph structure, locate all neighboring graph nodes that belong to the same independent subspace as the newly added modality's unique features; The newly added modality-specific features and the neighbor graph nodes are input into a graph attention network to calculate the attention coefficient of the newly added modality-specific features for each neighbor graph node; The initial weights of the edges are replaced with the attention coefficients, and new edges are added between the graph nodes corresponding to the new modality-specific features and the neighboring graph nodes in the heterogeneous graph structure to complete the incremental update of the local graph index.

Images