A multi-source heterogeneous data fusion digital intelligence talent portrait analysis system

Abstract

The application provides a multi-source heterogeneous data fusion digital talent portrait analysis system, relates to the field of artificial intelligence and big data processing technology, models talent ability and post demand as probability quality distribution, constructs cross-dimension substitution cost compensation impedance matrix based on skill ontology graph, introduces optimal transmission theory to solve the minimum cost path of surplus ability to missing dimension, realizes the post adaptation evaluation under the flexible substitution effect, solves the rigid adaptation problem that the traditional vector similarity algorithm cannot compensate across dimensions due to the dimension orthogonal assumption. At the same time, through the neighborhood co-occurrence frequency statistics and normalized point mutual information purification synergy affinity, the topological distance is adaptively discounted by bottleneck constraint and Gaussian decay gate screening optimal bridge node pair, eliminating the systematic distance overestimation bias caused by the shortest path algorithm ignoring the third party bridge skill, so that the compound talent obtains the actual adaptation score.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and big data processing technology, and more specifically, to a digital talent profiling and analysis system that integrates multi-source heterogeneous data. Background Technology

[0002] With the deepening of the digital transformation of the power industry, enterprises have put forward higher requirements for the selection and allocation of compound talents in cutting-edge fields such as ultra-high voltage construction and smart grid dispatch. The traditional evaluation method that relies on human experience can no longer meet the needs of large-scale and multi-dimensional talent screening. There is an urgent need to build a digital talent profile analysis system that can integrate multi-source heterogeneous data such as quantitative performance data, project development logs, and job description texts, so as to achieve accurate characterization of candidates' comprehensive abilities and objective assessment of job suitability.

[0003] However, existing job matching technologies generally employ vector metrics such as cosine similarity or Euclidean distance. These underlying assumptions require strictly orthogonal and independent feature dimensions, treating skills as isolated variables. This fails to simulate the flexible substitution effect in real-world scenarios where a significant advantage of one skill can effectively compensate for a deficiency in another, leading to a systematic underestimation of multi-skilled talent in rigid selection. While topological distance based on skill ontology graphs can be introduced to model the substitution cost between skill dimensions to address this rigid matching bottleneck, traditional shortest path algorithms follow an independent pairwise metric, calculating node distance only along the single link with the minimum edge weight. This implicitly assumes that the substitution cost is entirely determined by the shortest path and is independent of other nodes in the graph. This premise ignores the widespread phenomenon of skill synergy and bridging in industry practice. That is, when there is a third-party bridging skill in the graph that can co-occur with the target skill pair in high-frequency business, the difficulty of effective skill transfer at both ends will be significantly reduced due to the existence of the bridging skill. However, the existing algorithm is completely unaware of this and always outputs an uncorrected distance value, thereby introducing a systematic overestimation bias at the compensation cost matrix level. Ultimately, this results in a lower flexible matching score for multi-skilled talents, which directly contradicts the strategic needs of the power industry to select cross-domain digital talents.

[0004] Therefore, an optimized digital talent profiling and analysis system is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a digital talent profiling and analysis system based on multi-source heterogeneous data fusion, comprising: The feature extraction module is used to perform skill entity recognition and feature intensity mapping on the quantitative performance data, project development log text and standard job description text contained in the collected multi-source heterogeneous data to obtain talent ability distribution vector and job demand distribution vector. The impedance calculation module is used to perform graph node alignment and weighted shortest path distance calculation on the talent ability distribution vector and job requirement distribution vector based on the preset skill ontology graph to obtain the compensation impedance matrix that represents the ability substitution cost between different skill dimensions. The transmission planning module is used to perform flexible adaptive path planning on the talent capability distribution vector and job demand distribution vector, using the compensation impedance matrix as the cross-dimensional transportation cost function, to obtain the transmission planning tensor. The matching evaluation module is used to evaluate the flexible matching score and compensation effectiveness of the transmission planning tensor to obtain compensation insight data and flexible matching score. The report generation module is used to perform semantic parsing and business language transformation on the compensation logic relationships in the compensation insight data, and to integrate and encapsulate the flexible matching score with the corresponding strength and weakness compensation path to obtain a talent profile analysis report.

[0006] Compared with existing technologies, this application provides a digital talent profiling analysis system based on multi-source heterogeneous data fusion. It models talent capabilities and job requirements as probability quality distributions, constructs a compensation impedance matrix representing cross-dimensional substitution costs based on a skill ontology graph, and introduces optimal transmission theory to solve the globally minimum cost path for transferring capability quality from surplus dimensions to missing dimensions. This achieves a flexible substitution effect in talent-job fit assessment, solving the rigid fit problem of traditional vector similarity algorithms, which cannot perform capability compensation across skill dimensions due to the orthogonality assumption. Furthermore, addressing the deficiency of the shortest path algorithm in topological distance metrics in perceiving skill synergy bridging relationships, it purifies synergy affinity through neighborhood co-occurrence frequency statistics and normalized point mutual information. It also adaptively discounts and corrects the original distance by jointly selecting optimal bridging nodes using bottleneck constraints and Gaussian decay gating, eliminating the systematic distance overestimation bias caused by ignoring third-party bridging skills. This allows multi-skilled talents to obtain more business-relevant fit scores. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1 This is a system block diagram of a digital talent profiling analysis system based on multi-source heterogeneous data fusion according to an embodiment of this application; Figure 2 This is a data flow diagram of a digital talent profiling and analysis system that integrates multi-source heterogeneous data according to an embodiment of this application; Figure 3 This is a block diagram of the impedance calculation module in a digital talent profiling analysis system for multi-source heterogeneous data fusion according to an embodiment of this application. Figure 4 This is a block diagram of a path estimation unit in a digital talent profiling analysis system for multi-source heterogeneous data fusion according to an embodiment of this application; Figure 5 This is a block diagram of the transmission planning module in a digital talent profiling analysis system for multi-source heterogeneous data fusion according to an embodiment of this application; Figure 6 This is a block diagram of the matching and evaluation module in a digital talent profiling analysis system that integrates multi-source heterogeneous data according to an embodiment of this application. Detailed Implementation

[0009] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0010] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0011] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0012] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0013] Existing job-person matching technologies generally employ vector metrics such as cosine similarity or Euclidean distance. Their underlying assumption of orthogonality treats skills as isolated independent variables, failing to achieve cross-skill dimension compensation. This leads to a systematic underestimation of multi-skilled talent in rigid screening. Furthermore, even when introducing topological distance based on skill graphs to model alternative costs, traditional shortest path algorithms only calculate node distances along a single link, failing to perceive the synergistic effects arising from the presence of third-party bridging skills within the graph, further exacerbating the evaluation bias of multi-skilled talent. Therefore, this application proposes a digital talent profiling analysis system based on multi-source heterogeneous data fusion. The system first performs skill entity recognition and feature strength mapping on quantitative performance data, project development log text, and standard job description text from multi-source heterogeneous data, modeling talent capabilities and job requirements as probability quality distribution vectors respectively. Then, based on a pre-defined skill ontology graph, it constructs a compensation impedance matrix through graph node alignment and weighted shortest path distance calculation to quantify the substitution costs between different skill dimensions. This matrix is ​​then used as a cross-dimensional transport cost function, employing an entropy-regularized optimal transport algorithm to solve for the globally minimum cost transport path for transferring surplus capabilities to missing dimensions. Finally, through cost aggregation and compensation path analysis of the transport planning tensor, it outputs flexible matching scores and compensation insight data, which are then semantically transformed to generate a talent profile analysis report. Building on this, the system further introduces a collaborative bridging perception mechanism. Through neighborhood co-occurrence frequency statistics and normalized point mutual information, it refines the collaborative affinity between skills. Bottleneck constraints and Gaussian decay gating are used to jointly screen optimal bridging nodes, adaptively discounting the original topology distance to eliminate systematic distance overestimation bias caused by ignoring third-party bridging skills, ensuring that multi-skilled talents receive appropriate assessment results that align with actual business needs.

[0014] Figure 1 This is a system block diagram of a digital talent profiling analysis system based on multi-source heterogeneous data fusion according to an embodiment of this application. Figure 2 This is a data flow diagram illustrating a digital talent profiling and analysis system based on multi-source heterogeneous data fusion according to an embodiment of this application. Figure 1 and Figure 2As shown, a digital talent profiling analysis system 100 based on multi-source heterogeneous data fusion according to an embodiment of this application includes: a feature extraction module 110, used to perform skill entity recognition and feature intensity mapping on the quantitative performance data, project development log text, and standard job description text contained in the collected multi-source heterogeneous data to obtain talent ability distribution vectors and job requirement distribution vectors; and an impedance calculation module 120, used to perform graph node alignment and weighted shortest path distance calculation on the talent ability distribution vectors and job requirement distribution vectors based on a preset skill ontology graph to obtain the cost of ability substitution between different skill dimensions. The system includes: a compensation impedance matrix; a transmission planning module 130, which uses the compensation impedance matrix as a cross-dimensional transport cost function to perform flexible adaptation path planning on the talent capability distribution vector and job requirement distribution vector to obtain a transmission planning tensor; a matching evaluation module 140, which evaluates the flexible matching score and compensation effectiveness of the transmission planning tensor to obtain compensation insight data and flexible matching score; and a report generation module 150, which performs semantic parsing and business language conversion on the compensation logic relationship in the compensation insight data, and integrates and encapsulates the flexible matching score with the corresponding strength and weakness compensation path to obtain a talent profile analysis report.

[0015] In the aforementioned digital talent profiling analysis system 100 based on multi-source heterogeneous data fusion, the feature extraction module 110 is used to perform skill entity recognition and feature intensity mapping on the quantitative performance data, project development log text, and standard job description text contained in the collected multi-source heterogeneous data to obtain talent ability distribution vectors and job demand distribution vectors. It should be noted that, since the quantitative performance data, project development log text, and standard job description text in the multi-source heterogeneous data belong to different data structures and semantic systems, they cannot be directly used for subsequent cross-dimensional ability compensation calculations. Based on this, the technical solution of this application first performs skill entity recognition and feature intensity mapping on the multi-source heterogeneous data, and transforms the results into standardized vectors that satisfy probability quality distribution constraints. Through the above processing, heterogeneous data can be uniformly projected onto the same mathematical space, providing standardized input for subsequent flexible adaptation based on optimal transmission theory.

[0016] More specifically, in a concrete example of this application, taking the candidate evaluation for a smart grid dispatching position in the power industry as an example, the project development log text accumulated by the candidate in the project management platform and the standard job description text corresponding to the target position are first input into a named entity recognition network based on a pre-trained language model using the BERT architecture. This network performs sentence-by-sentence scanning and sequence annotation on the log text, identifying skill entity terms such as power system steady-state analysis, Python programming, and deep learning algorithm design. Simultaneously, the same entity recognition operation is performed on the standard job description text to extract the skill terms required for the position. Subsequently, all extracted results are processed by synonym merging and cleaning, unifying semantically equivalent but differently expressed terms into standardized skill identifiers, generating a talent skill entity set and a job requirement entity set respectively.

[0017] Specifically, the training process of the aforementioned named entity recognition network is as follows: First, a large amount of text corpus is collected from historical project documents, job descriptions, and technical reports in the power industry. Domain experts then annotate the skill entities appearing in the corpus word by word according to the BIO annotation system, marking each lexical unit as the starting position, internal position, or non-entity position of the skill entity, thus constructing a skill entity annotation corpus. Subsequently, a pre-trained BERT language model is used as the encoder base, and a conditional random field sequence annotation layer is connected to its top layer. The overall network parameters are then subjected to supervised fine-tuning training using the aforementioned annotation corpus. The model weights are iteratively updated by minimizing the negative log-likelihood loss function of the sequence annotations until the entity recognition precision and recall on the validation set converge to a preset threshold, thus completing the training of the named entity recognition network for power industry skill entities.

[0018] After obtaining the two entity sets mentioned above, the demand intensity weights of each skill entity in the job requirement entity set are calculated. Based on the word frequency-inverse text frequency algorithm, the frequency of each skill entity in the job description text and its scarcity in the global job corpus are statistically analyzed. The calculated weight values ​​of each dimension are then arranged to form the original job requirement vector. Simultaneously, the talent skill entity set is logarithmically weighted and aggregated based on quantitative performance data. The candidate's basic performance score is extracted from the HR system as the evaluation base. The cumulative frequency of each skill entity in the project development log is calculated using a formula... Calculate the ability mapping value for each skill dimension, where Based on performance score, For the first The cumulative frequency of occurrence of each skill To adjust the empirical hyperparameter of frequency gain, all capability mapping values ​​are arranged by dimension to form the original vector of talent capability.

[0019] After obtaining the original vectors of talent capabilities and job requirements, a probability mass distribution transformation is performed on both. This involves calculating the sum of the absolute values ​​of all elements in the vectors, dividing the value of each element by the sum of these absolute values, so that all elements of the transformed vector are non-negative and their sum is strictly equal to 1, satisfying the mathematical constraints of the probability mass distribution. This yields the talent capability distribution vector and the job requirement distribution vector, which serve as standardized inputs for subsequent optimal transmission flexible adaptation calculations.

[0020] In the aforementioned multi-source heterogeneous data fusion-based intelligent talent profiling analysis system 100, the impedance calculation module 120 is used to perform graph node alignment and weighted shortest path distance calculation on the talent capability distribution vector and job requirement distribution vector based on a preset skill ontology graph to obtain a compensation impedance matrix representing the cost of capability substitution between different skill dimensions. It should be noted that although the talent capability distribution vector and job requirement distribution vector obtained in the previous step are in the same mathematical space, the subsequent optimal transfer algorithm needs a clear cost basis to measure the ease of substitution between different skills when performing cross-dimensional capability transfer. This cost cannot be obtained from the distribution vector itself and needs to be modeled using external domain knowledge. Based on this, the technical solution of this application further performs graph node alignment and weighted shortest path distance calculation on all skill feature dimensions involved in the talent capability distribution vector and job requirement distribution vector based on a preset skill ontology graph, and transforms the topological distance into a compensation impedance matrix through nonlinear transformation. Through the above processing, each pair of skill dimensions can be assigned a quantified value of substitution cost with domain semantics, so that the subsequent transfer planning has reasonable cost constraints when performing capability transfer.

[0021] Figure 3 This is a block diagram of the impedance calculation module in a digital talent profiling analysis system based on multi-source heterogeneous data fusion, according to an embodiment of this application. Figure 3 As shown, the impedance calculation module 120 includes: a dimension alignment unit 121, used to perform skill dimension alignment and ontology reference mapping on the talent ability distribution vector and the job requirement distribution vector to obtain a skill feature set; a path estimation unit 122, used to perform graph topology-based shortest path estimation on the skill feature set based on a preset skill ontology graph to obtain a skill topology distance matrix; and a cost transformation unit 123, used to perform nonlinear compensation cost transformation on each distance element in the skill topology distance matrix to obtain a compensation impedance matrix.

[0022] In the aforementioned digital talent profiling analysis system 100 based on multi-source heterogeneous data fusion, the dimension alignment unit 121 is used to perform skill dimension alignment and ontology reference mapping on the talent capability distribution vector and the job requirement distribution vector to obtain a skill feature set. It should be noted that since the talent capability distribution vector and the job requirement distribution vector may cover skill dimensions that do not completely overlap, and the dimension identifiers have not yet been associated with knowledge nodes in the skill ontology graph, path distance calculation cannot be directly performed on the graph. Based on this, the technical solution of this application further performs skill dimension alignment and ontology reference mapping on the two distribution vectors to obtain a skill feature set. Through the above processing, all skill dimensions scattered in the two vectors can be uniformly anchored to graph nodes, providing a complete and consistent node range for subsequent topological distance calculations.

[0023] More specifically, in a concrete example of this application, the entire dimensional index of the talent capability distribution vector and the job requirement distribution vector are traversed separately to extract the skill identifiers corresponding to each dimension. Examples include skill terms such as power system steady-state analysis, Python programming, deep learning algorithm design, and job-specific requirements like UHV equipment operation and maintenance, and distributed architecture design. The skill identifiers of both are then combined to eliminate duplicates. Subsequently, each skill identifier in the union is semantically matched with knowledge nodes in a pre-defined power industry skill ontology graph. Successfully matched nodes are marked as active nodes, and their unique ID and hierarchical affiliation information in the graph are recorded. For skill identifiers that do not match directly, semantic backtracking is performed along the hierarchical relationships in the graph until the nearest equivalent node is located. After anchoring nodes in all dimensions, all active nodes and their graph IDs are encapsulated into a skill feature set, which serves as the node input range for subsequent shortest path distance calculations on the graph.

[0024] In the aforementioned multi-source heterogeneous data fusion-based intelligent talent profiling analysis system 100, the path estimation unit 122 is used to perform graph topology-based shortest path estimation on the skill feature set based on a preset skill ontology graph to obtain a skill topology distance matrix. It should be noted that, given that the previous step has anchored all skill dimensions to corresponding nodes in the skill ontology graph, but the topological relationships between nodes have not yet been quantified into calculable numerical distances, and the subsequent compensation cost transformation requires specific distance values ​​between nodes as the input basis, the technical solution of this application further performs graph topology-based shortest path estimation on all active nodes in the skill feature set based on a preset skill ontology graph to obtain a skill topology distance matrix. Through the above processing, a quantified distance value reflecting the semantic span of any two skill nodes in the skill feature set can be generated, providing a complete distance basis for the downstream nonlinear compensation cost transformation.

[0025] Figure 4 This is a block diagram of a path estimation unit in a digital talent profiling analysis system based on multi-source heterogeneous data fusion, according to an embodiment of this application. Figure 4 As shown, the path estimation unit 122 includes: a distance statistics subunit 1221, used to perform original topological distance calculation and neighborhood co-occurrence frequency statistics on the skill feature set based on a preset skill ontology graph to obtain an original topological distance matrix and a neighborhood co-occurrence frequency matrix; an affinity quantification subunit 1222, used to perform cooperative affinity quantification on the neighborhood co-occurrence frequency matrix based on normalized point mutual information to obtain a cooperative affinity matrix; and a distance correction subunit 1223, used to perform cooperative bridging gating topological distance correction on the original topological distance matrix based on the cooperative affinity matrix to obtain a skill topological distance matrix.

[0026] In the aforementioned multi-source heterogeneous data fusion-based intelligent talent profiling analysis system 100, the distance statistics subunit 1221 is used to perform original topological distance calculation and neighborhood co-occurrence frequency statistics on the skill feature set based on a preset skill ontology graph to obtain an original topological distance matrix and a neighborhood co-occurrence frequency matrix. It should be noted that, given that the Floyd-Warshall algorithm follows an independent pairwise distance metric, it only calculates the node distance along the single link with the minimum sum of edge weights. It cannot perceive the synergistic effect caused by the existence of third-party bridging skills in the graph. Furthermore, the shortest path distance it calculates still objectively reflects the basic substitution difficulty between skills, constituting a benchmark reference for subsequent correction calculations. Therefore, it is necessary to extract neighborhood co-occurrence information while retaining the basic distance results. Based on this, the technical solution of this application further performs original topological distance calculation and neighborhood co-occurrence frequency statistics on the skill feature set based on a preset skill ontology graph to obtain an original topological distance matrix and a neighborhood co-occurrence frequency matrix. Through the above processing, we can obtain the basic substitution difficulty quantification value between skills, and provide original frequency evidence for subsequent judgment on whether there is a collaborative bridging relationship between skill pairs.

[0027] More specifically, in a concrete example of this application, a set of skill features is received, all skill node pairs within it are traversed, and the shortest path distance between each pair of nodes is calculated using the Floyd-Warshall algorithm on the power industry skill ontology graph. All calculation results are then filled into a two-dimensional square matrix to generate the original topological distance matrix, whose elements... The calculation method is as follows: in, Skill nodes in the original topological distance matrix With skill nodes The shortest topological distance between them Connecting nodes in the graph With nodes The set of all feasible paths, For a specific path in the set, For path One of the edges in, For the edge The pre-defined semantic weights represent different semantic spans between skills. While calculating the initial distance, neighborhood co-occurrence information is extracted as the frequency basis for subsequent affinity modeling. For each skill node in the skill feature set, all neighboring nodes within its K-hop reachable range are extracted from the industry skill ontology graph. The number of nodes contained in the intersection of the K-hop neighborhoods of any two skill node pairs is counted. All statistical results are filled into a square matrix to generate a neighborhood co-occurrence frequency matrix, whose elements... The calculation method is as follows: in, Skill nodes in the neighborhood co-occurrence frequency matrix With skill nodes Size of the intersection of K-jump neighborhoods In the graph, nodes The set of all neighboring nodes within a K-hop reachable radius of the center. This is the hop radius hyperparameter for neighborhood search. For example, in the power industry skill map, the skills of power system steady-state analysis and smart grid dispatch optimization share multiple neighborhood nodes such as power flow calculation, load forecasting modeling, and dispatch strategy optimization within a 2-hop range. The large number of their neighborhood intersections means that the knowledge structure of these two skills in industry job practice is frequently covered by the same set of professional directions, providing original frequency evidence for subsequent judgment on whether there is a real collaborative bridging relationship between the two.

[0028] In the aforementioned multi-source heterogeneous data fusion-based intelligent talent profiling analysis system 100, the affinity quantification subunit 1222 is used to perform cooperative affinity quantification on the neighborhood co-occurrence frequency matrix based on normalized point mutual information to obtain a cooperative affinity matrix. It should be noted that, given that the neighborhood co-occurrence frequency obtained in the previous step only reflects the absolute number of co-occurrences of skill pairs, it is not possible to directly determine whether co-occurrence originates from a genuine business collaboration relationship. General basic skills, such as routine office software operations, are widely distributed and have significant overlap with the neighborhoods of almost all skills, but this does not represent genuine collaborative enhancement. Based on this, the technical solution of this application further performs cooperative affinity quantification on the neighborhood co-occurrence frequency matrix based on normalized point mutual information to obtain a cooperative affinity matrix. Through the above processing, it is possible to distinguish between random co-occurrence caused by universality and directional co-occurrence caused by business collaboration, and to extract truly collaborative signals that transcend randomness from the frequency.

[0029] More specifically, in a particular example of this application, the neighborhood co-occurrence frequency matrix is ​​received, and the total number of skill nodes in the graph is given by... First, calculate each skill node. marginal probability and the joint probability of each pair of skill nodes Then, the normalized point mutual information value is calculated for each element position in the matrix, and all the calculation results are filled into a square matrix to generate a cooperative affinity matrix, whose elements... The calculation method is as follows: in, Skills in the Co-affinity Matrix With skills The cooperative affinity, with a value range of A value closer to 1 indicates that the co-occurrence of the two in the local structure of the map far exceeds random expectations and there is a strong cooperative relationship. A value closer to 0 indicates that the co-occurrence is on par with the random level and there is no significant cooperative relationship. A negative value indicates that the two have a mutual exclusion tendency. Skill Node The marginal probability, For skill node pairs The joint probability, denominator As a normalization factor, it eliminates the influence of different frequency magnitudes on the absolute value of mutual information, ensuring that affinity is comparable across different skill pairs across magnitudes. For example, in real-world scenarios in the power industry, the normalized point mutual information value between power system steady-state analysis and smart grid dispatch optimization will be higher than the random level because they are frequently covered by the same set of professional directions in grid dispatch operations, possessing a real business collaboration basis. While the original co-occurrence frequency between power system steady-state analysis and conventional office software operations may not be low, after filtering with normalized point mutual information, their affinity will drop back to a level close to zero, thereby achieving accurate purification of real collaborative signals.

[0030] In the aforementioned multi-source heterogeneous data fusion-based intelligent talent profiling analysis system 100, the distance correction subunit 1223 is used to perform topological distance correction with collaborative bridging gating on the original topological distance matrix based on the collaborative affinity matrix to obtain the skill topological distance matrix. It should be noted that directly adding affinity to distance can distort the topological relationship, as collaborative bridging is an indirect effect via a third-party skill. Therefore, it is necessary to find the optimal bridging node in the graph that can simultaneously form strong collaboration with both ends of the target skill pair and quantify the resulting distance reduction. Furthermore, simply relying on high affinity to determine bridging effectiveness is insufficient; the actual topological distance between the bridging skill and the target skills at both ends must also be considered. Bridging that is too far apart lacks feasibility in business practice and should be subject to distance attenuation constraints. Based on this, the technical solution of this application further performs topological distance correction with collaborative bridging gating on the original topological distance matrix based on the collaborative affinity matrix to obtain the skill topological distance matrix. Through the above processing, adaptive distance correction can be achieved, automatically discounting when there is collaborative bridging and maintaining the original value when there is no collaborative bridging.

[0031] More specifically, in a concrete example of this application, both the original topological distance matrix and the cooperative affinity matrix are received simultaneously. For each pair of skills in the original topological distance matrix... Traverse all candidate bridging skill nodes in the collaborative affinity matrix. ( and ), calculate each candidate bridge node The bridging gain, take respectively with and The smaller affinity between nodes serves as a bottleneck constraint. The strength of a bridging link is determined by its weakest end, multiplied by a Gaussian attenuation gating factor constructed based on the original topological distance. Finally, the maximum gain value is selected from all candidate bridging nodes. (Bridging gain factor) The calculation method is as follows: in, For skills Through optimal bridging skills The larger the obtained bridging gain factor, the stronger the synergistic bridging effect. and These are the bridging candidate nodes in the cooperative affinity matrix. With skills and skills The degree of synergistic affinity between them Operator implementation bottleneck constraints ensure bridging skills The skill pair must have substantial synergy with both ends of the target skill pair, rather than just affinity with one side. In the power industry scenario, a skill that is highly synergistic with power system steady-state analysis but has no connection with deep learning algorithm design should not be considered a valid bridge. and These are the bridging candidate nodes in the original topological distance matrix. The shortest distance between the two skills. For Gaussian decay gated function, when bridging skill When either end is too far away in the topological graph, even if the affinity value is high, the gating will compress its gain to near zero, thereby suppressing unreasonable long-distance false bridging in the topology. In actual deployment, edge nodes with a topological distance greater than a preset threshold should not substantially correct the distance of the core skill pair. The gating width hyperparameter controls the effective radiation radius of the bridging effect. The operator selects the node with the largest gain among all candidate bridging nodes as the skill pair. Anchor the optimal bridge.

[0032] After obtaining the bridging gain factor, it is used to discount and correct the corresponding elements in the original topology distance matrix, generating the final skill topology distance matrix, whose elements... The calculation method is as follows: in, Skill pairs in the skill topology distance matrix after cooperative bridging correction The final distance directly replaces the output of the same name in sub-step 2.2 of the first embodiment and seamlessly connects to the logarithmic cost transformation in the downstream sub-step 2.3. This is the original shortest path distance before correction. To define the bridging discount strength hyperparameter, which controls the maximum correction magnitude of the cooperative bridging effect on the original distance. A non-negative truncation is applied to the bridging gain factor. When the gain of all candidate bridging nodes is non-positive, meaning there are no effective cooperative bridging skills for the current skill pair in the graph, the correction factor is set to zero, the denominator degenerates to 1, and the output value is strictly equal to the original distance. This ensures that the improvement mechanism does not introduce any bias in scenarios without bridging. In the practical scenario of talent assessment in the power industry, power system steady-state analysis is used as a skill. Deep learning algorithm design as a skill For example, if smart grid dispatch optimization is considered a skill... If a node is identified by the algorithm as the optimal bridging node and its gain value remains positive after being filtered by both bottleneck constraints and Gaussian gating, the final topological distance between the two nodes will be reasonably reduced. This will allow the composite candidate who possesses all three skills to obtain a more business-appropriate matching score in the subsequent optimal transmission flexibility adaptation.

[0033] In the aforementioned multi-source heterogeneous data fusion-based intelligent talent profiling analysis system 100, the cost transformation unit 123 is used to perform nonlinear compensation cost transformation on each distance element in the skill topology distance matrix to obtain a compensation impedance matrix. It should be noted that since the original distance values ​​in the skill topology distance matrix reflect the path length on the graph, their numerical scale does not directly match the transport cost required by the subsequent optimal transport algorithm in terms of dimensions and distribution characteristics. Furthermore, excessively large original distance values ​​for long-distance skill pairs can easily lead to cost explosion. Based on this, the technical solution of this application further performs nonlinear compensation cost transformation on each distance element in the skill topology distance matrix to obtain a compensation impedance matrix. Through the above processing, the topology distance can be smoothly mapped to a transport cost with a reasonable value range distribution, providing a numerically stable cost function for subsequent optimal transport solutions.

[0034] More specifically, in a particular example of this application, the distance value of each element in the skill topology distance matrix is ​​read one by one. Applying a logarithmic nonlinear transformation to it, the calculation formula is as follows: ,in To compensate for the impedance matrix of the first Line number The element values ​​of the column represent the ability from the skill dimension. Compensation to skill dimension The unit transportation cost The skill topology distance after collaborative bridging correction. The proportional sensitivity coefficient hyperparameter is used to adjust the sensitivity of the compensation mechanism to the skill span. When two skills are at the same node, the distance is 0, and the impedance after transformation is also 0, indicating that no cross-dimensional transport is needed. As the distance increases, the logarithmic function gradually slows down the impedance growth rate, avoiding the overflow of cost values ​​caused by linear accumulation for cross-knowledge domain skill pairs in the graph. For example, although the steady-state analysis of power systems and the design of deep learning algorithms have a large span in the graph, after logarithmic compression, their compensation costs are constrained within a reasonable range, and will not completely block the flow of capabilities between the two in subsequent transmission solutions due to excessively large values. After traversing all element positions and completing the transformation, the results are filled into a square matrix to generate a compensation impedance matrix, which serves as the input of the cost function for subsequent optimal transmission flexible adaptation.

[0035] In the aforementioned multi-source heterogeneous data fusion-based intelligent talent profiling analysis system 100, the transmission planning module 130 is used to perform flexible adaptive path planning on the talent capability distribution vector and job demand distribution vector using a compensation impedance matrix as the cross-dimensional transportation cost function to obtain a transmission planning tensor. It should be noted that, given that the compensation impedance matrix has quantified the transportation cost between any two skill dimensions, and the talent capability distribution vector and job demand distribution vector respectively represent the candidate's capability supply side and the job's capability demand side, these three elements already possess all the necessary components for constructing the optimal transmission problem. Therefore, a cross-dimensional capability flow allocation scheme that minimizes the total transportation cost needs to be solved. Based on this, the technical solution of this application further uses the compensation impedance matrix as the cross-dimensional transportation cost function to perform flexible adaptive path planning on the talent capability distribution vector and job demand distribution vector to obtain a transmission planning tensor. Through the above processing, the capability quality of the candidate's surplus dimensions can be transferred to the missing dimensions with the minimum total cost, generating a transmission planning tensor that records the specific flow allocation between each dimension, providing a complete mathematical planning basis for subsequent matching scoring and compensatory path analysis.

[0036] Figure 5 This is a block diagram of the transmission planning module in a digital talent profiling analysis system based on multi-source heterogeneous data fusion, according to an embodiment of this application. Figure 5 As shown, the transmission planning module 130 includes: a target construction unit 131, used to construct a regularized objective function based on the compensation impedance matrix as the basic cost term and superimposed with an information entropy regularization term, and simultaneously set row and column sum constraints based on the talent capability distribution vector and the job demand distribution vector to obtain a supply and demand boundary constraint set; an iterative solution unit 132, used to construct a Gibbs kernel matrix based on the regularized objective function, and alternately update the left and right dual scaling vectors under the row and column sum constraints of the supply and demand boundary constraint set through an iterative scaling algorithm until convergence, to obtain the optimal joint distribution matrix; and a tensor encapsulation unit 133, used to perform zero-setting filtering on the near-zero floating-point drift terms generated by iterative truncation in the optimal joint distribution matrix, and perform high-dimensional tensor encapsulation along the batch and scene feature dimensions to obtain a transmission planning tensor.

[0037] In the aforementioned multi-source heterogeneous data fusion-based intelligent talent profiling analysis system 100, the target construction unit 131 is used to construct a regularized objective function based on a compensation impedance matrix as the basic cost term and superimposed with an information entropy regularization term. Simultaneously, based on the talent capability distribution vector and the job demand distribution vector, row and column constraints are set respectively to obtain a supply-demand boundary constraint set. It should be noted that solving the optimal transmission problem requires a clear mathematical optimization objective and corresponding boundary constraints. The original linear cost minimization problem has high computational complexity in high-dimensional scenarios and may have multiple equivalent optimal solutions. Regularization is needed to ensure the strict convexity of the objective function and the uniqueness of the solution. Based on this, the technical solution of this application further constructs a regularized objective function based on a compensation impedance matrix as the basic cost term and superimposed with an information entropy regularization term. Simultaneously, based on the talent capability distribution vector and the job demand distribution vector, row and column constraints are set respectively to obtain a supply-demand boundary constraint set. Through the above processing, the flexible adaptation problem can be transformed into a strictly convex optimization problem with a unique global optimal solution, laying a mathematical foundation for subsequent iterative solutions.

[0038] More specifically, in a particular example of this application, a joint distribution matrix to be solved is first established. Each element Representatives from the candidates' first The first skill dimension is directed towards the job. The capacity and quality allocation quota for transporting items according to demand dimensions. This is used to compensate for elements in the impedance matrix. As a basic cost term, the total transportation cost across all dimension pairs is calculated, and the information entropy of the joint distribution matrix itself is superimposed as a regularization term to construct the regularization objective function: in, The allocation amount to be solved in the joint distribution matrix. To compensate for the impedance matrix from the skill dimension Towards the skill dimension The unit cost of transportation The total number of skill feature dimensions. This is the entropy regularization coefficient. A larger coefficient results in a faster solution but a smoother result, while a smaller coefficient results in a more accurate result but lower numerical stability. The basic transportation total cost item, The information entropy regularization term ensures the strict convexity of the objective function, guaranteeing a unique global optimum for the optimization problem. Simultaneously, the talent capability distribution vector is set as a row sum constraint of the joint distribution matrix, meaning the sum of elements in each row of the matrix strictly equals the candidate's capability quality in the corresponding skill dimension. Similarly, the job demand distribution vector is set as a column sum constraint, meaning the sum of elements in each column of the matrix equals the job's required quality in the corresponding dimension. These row and column sum constraints are encapsulated into a supply-demand boundary constraint set, ensuring the conservation of total quality during capability transfer.

[0039] In the aforementioned multi-source heterogeneous data fusion-based intelligent talent profiling analysis system 100, the iterative solution unit 132 is used to construct a Gibbs kernel matrix based on a regularized objective function, and then alternately update the left and right dual scaling vectors under the row and column sum constraints of the supply and demand boundary constraint set using an iterative scaling algorithm until convergence, thereby obtaining the optimal joint distribution matrix. It should be noted that since the regularized objective function constructed in the previous step is a convex optimization problem with equality constraints, a closed-form solution cannot be directly obtained analytically; a numerical iterative algorithm is needed for efficient solution. Based on this, the technical solution of this application further constructs a Gibbs kernel matrix based on the regularized objective function, and then alternately updates the left and right dual scaling vectors under the row and column sum constraints of the supply and demand boundary constraint set using an iterative scaling algorithm until convergence, thereby obtaining the optimal joint distribution matrix. Through the above processing, the optimal transmission scheme can be quickly approximated with lower computational complexity, obtaining the optimal joint distribution matrix recording the capacity flow allocation among various skill dimensions.

[0040] More specifically, in a particular example of this application, the Gibbs kernel matrix is ​​first constructed based on the compensation impedance matrix and the entropy regularization coefficient, and its elements are calculated as follows: ,in The first in the Gibbs kernel matrix Line number Column elements, To compensate for the unit transportation cost at the corresponding position in the impedance matrix, The entropy regularization coefficient is then used. The left dual scaling vector is subsequently initialized. With right dual scaling vector When all elements are 1, the iteration scaling loop begins. In each iteration, the talent ability distribution vector is first used as the initial value. The left scaling vector is updated by dividing by the product of the Gibbs kernel matrix and the current right scaling vector, i.e. Then, using the job demand distribution vector The right scaling vector is updated by dividing by the product of the transpose of the Gibbs kernel matrix and the updated left scaling vector. ,in This represents the row sum and constraints of the talent capability distribution vector, i.e., the supply and demand boundary constraint set. The job requirement distribution vector is defined by columns and constraints. For the current iteration, the division operation is an element-wise division between vectors. The above alternating left and right update process continues until the change in the scaling vector between two adjacent iterations converges to a preset threshold, at which point the iteration stops. The optimal joint distribution matrix is ​​then calculated using the converged left and right scaling vectors and the Gibbs kernel matrix. ,in and This is the final left and right scaling vector after convergence. An operator for converting a vector into a diagonal matrix. This is the optimal joint distribution matrix that satisfies the supply and demand boundary constraints. Each element records the actual flow of ability quality that candidates transfer from a specific skill dimension to a specific job requirement dimension.

[0041] In the aforementioned multi-source heterogeneous data fusion-based intelligent talent profiling analysis system 100, the tensor encapsulation unit 133 is used to perform zero-setting filtering on near-zero floating-point drift terms generated by iterative truncation in the optimal joint distribution matrix, and to perform high-dimensional tensor encapsulation along the batch and scene feature dimensions to obtain the transmission planning tensor. It should be noted that after a finite number of rounds of truncation by the iterative scaling algorithm, a very small amount of floating-point drift noise will remain at the positions of some theoretically zero elements in the optimal joint distribution matrix. If this noise is not removed, it will be misjudged as valid flow during subsequent compensation path extraction. Furthermore, the two-dimensional matrix structure cannot meet the engineering requirements of batch evaluation of multiple candidates and parallel processing of multiple job scenarios. Based on this, the technical solution of this application further performs zero-setting filtering on near-zero floating-point drift terms generated by iterative truncation in the optimal joint distribution matrix, and performs high-dimensional tensor encapsulation along the batch and scene feature dimensions to obtain the transmission planning tensor. Through the above processing, the interference of numerical noise on downstream compensation path identification can be eliminated, and the results can be organized into a standard data structure that supports batch computation.

[0042] More specifically, in a concrete example of this application, all elements in the optimal joint distribution matrix are scanned one by one, and elements with absolute values ​​lower than a preset floating-point precision threshold are forcibly set to zero. This threshold is set to a very small positive number of the same order of magnitude as the convergence precision of the iterative scaling algorithm. After this filtering, only non-zero flow components with actual physical meaning are retained in the matrix, that is, the dimension pairs where capacity transfer actually occurred and their corresponding flow values. The matrix as a whole exhibits sparsity characteristics. Subsequently, based on the cleaned two-dimensional matrix, the batch dimension is expanded along the 0th axis to identify evaluation instances of different candidates, and the scene feature dimension is expanded along the 3rd axis to identify matching scenarios of different target positions. The expanded data is encapsulated into a 4th-order tensor structure to generate a transport planning tensor, which serves as the input for subsequent flexible matching score calculation and compensation effectiveness evaluation.

[0043] In the aforementioned multi-source heterogeneous data fusion-based intelligent talent profiling analysis system 100, the matching evaluation module 140 is used to perform flexible matching scoring and compensation effectiveness evaluation on the transmission planning tensor to obtain compensation insight data and flexible matching scores. It should be noted that, given that the transmission planning tensor records the raw flow distribution data between various skill dimensions, it has not yet been transformed into a quantitative score and interpretable compensation logic that can be directly used for talent selection decisions, thus failing to meet the dual requirements of the business side for matching conclusions and compensation paths. Based on this, the technical solution of this application further performs flexible matching scoring and compensation effectiveness evaluation on the transmission planning tensor to obtain compensation insight data and flexible matching scores. Through the above processing, the underlying mathematical programming results can be condensed into scalar scores reflecting the overall suitability and structured insight data revealing specific strengths and weaknesses in compensation relationships, providing scoring basis and interpretive material for the generation of the final talent profiling analysis report.

[0044] Figure 6 This is a block diagram of the matching and evaluation module in a digital talent profiling analysis system based on multi-source heterogeneous data fusion, according to an embodiment of this application. Figure 6 As shown, the matching evaluation module 140 includes: a cost aggregation unit 141, used to perform total transmission cost aggregation calculation on the transmission planning tensor to obtain the global transportation cost and flow component matrix; a score mapping unit 142, used to perform nonlinear mapping on the global transportation cost to obtain the flexible matching score; and an insight extraction unit 143, used to extract all non-zero cross-dimensional compensation paths and their traffic proportions after shielding the main diagonal of the flow component matrix to obtain compensation insight data.

[0045] In the aforementioned multi-source heterogeneous data fusion-based intelligent talent profiling analysis system 100, the cost aggregation unit 141 is used to perform total transmission cost aggregation calculations on the transmission planning tensor to obtain the global transportation cost and flow component matrix. It should be noted that since the transmission planning tensor stores the flow allocation data between all dimensions in a high-dimensional encapsulation form, it needs to be unpacked and multiplied and accumulated with the corresponding unit cost item by item to obtain a scalar indicator measuring the overall transportation difficulty. Simultaneously, the matrix form of the flow components needs to be preserved for subsequent compensation path extraction. Based on this, the technical solution of this application further performs total transmission cost aggregation calculations on the transmission planning tensor to obtain the global transportation cost and flow component matrix. Through the above processing, the high-dimensional tensor can be reduced to a scalar value reflecting the overall adaptation cost and a two-dimensional matrix retaining the flow details between each dimension, respectively serving subsequent score mapping and insight extraction.

[0046] More specifically, in a concrete example of this application, the transport planning tensor is sliced ​​along the batch and scenario dimensions to extract the two-dimensional optimal joint distribution matrix corresponding to the current candidate and the target position as the flow component matrix. Each non-zero element in this matrix represents the actual capacity quality flow transported from a certain skill dimension to a certain position requirement dimension. Subsequently, the flow component matrix and the compensation impedance matrix are subjected to element-wise Hadamard product to calculate the local compensation cost generated on each transport path. Then, all local costs are globally summed to obtain the global transport cost, calculated as follows: in The cost of global transport is the Wasserstein distance. The total number of skill feature dimensions. To compensate for the impedance matrix from the skill dimension Towards Dimension The unit cost of transportation This assigns the optimal flow rate quota to the corresponding position in the flow component matrix. A smaller global transport cost indicates a smaller overall deviation between the candidate's ability distribution and the job requirement distribution, and a higher degree of flexible adaptation.

[0047] In the aforementioned multi-source heterogeneous data fusion-based intelligent talent profiling analysis system 100, the scoring mapping unit 142 is used to perform a nonlinear mapping on the global transfer cost value to obtain a flexible matching score. It should be noted that since the global transfer cost value ranges from zero to positive infinity, it cannot be directly used as a job-person matching percentage indicator that is understandable to the business side. It needs to be mapped to a bounded interval to facilitate horizontal comparison and ranking between different candidates. Based on this, the technical solution of this application further performs a nonlinear mapping on the global transfer cost value to obtain a flexible matching score. Through the above processing, the continuous and unbounded cost value can be transformed into a standardized score between 0 and 1, intuitively reflecting the overall suitability between the candidate and the target position after considering the ability compensation effect.

[0048] More specifically, in a particular example of this application, the global transport cost is received and nonlinearly mapped using an exponential decay function, calculated as follows: ,in The score is for flexible matching, and its value range is [value range missing]. , Value for global data transfer For attenuation control parameters, For the natural constant An exponential function with base 0. When the global transfer cost is 0, meaning the candidate's ability distribution perfectly matches the job requirement distribution without any cross-dimensional transfer, the flexible matching score is 1; as the cost increases, the score decreases exponentially. Attenuation control parameter. Different settings are made according to different job sequences. For example, for R&D positions related to smart grid dispatching, the tolerance for compensatory skills across disciplines is relatively high. A smaller value is used to relax the rate of score decay, while a larger value is used for power safety and compliance positions to increase the sensitivity of the score to relocation costs, ensuring the business applicability of the flexible matching score.

[0049] In the aforementioned digital talent profiling analysis system 100 based on multi-source heterogeneous data fusion, the insight extraction unit 143 is used to extract all non-zero cross-dimensional compensation paths and their flow ratios after masking the main diagonal of the flow component matrix to obtain compensation insight data. It should be noted that since the elements on the main diagonal of the flow component matrix represent the direct digestion of capabilities within the same dimension rather than cross-dimensional compensation, only the non-zero elements off-diagonal reflect the compensation paths where the phenomenon of compensating for weaknesses with strengths truly occurs. Therefore, it is necessary to distinguish between the two and extract compensation-related feature information. Based on this, the technical solution of this application further extracts all non-zero cross-dimensional compensation paths and their flow ratios after masking the main diagonal of the flow component matrix to obtain compensation insight data. Through the above processing, it is possible to identify which specific strengths a candidate has compensated for which weaknesses, and the intensity of that compensation, providing structured material for the interpretability of subsequent reports.

[0050] More specifically, in a concrete example of this application, the receiving flow component matrix is ​​configured by setting all elements on the main diagonal to zero to mask the internal digestion flow within the same dimension, retaining all non-zero off-diagonal elements in the matrix. Each non-zero element corresponds to a cross-dimensional compensation path, with its row index identifying the skill dimension on the capacity supply side, its column index identifying the job dimension on the demand receiving side, and its element value identifying the actual capacity quality flow transported along that path. For example, the non-zero flow from the power system steady-state analysis dimension to the UHV equipment operation and maintenance dimension in the flow component matrix indicates that the candidate has compensated for the job shortage in the latter with the capacity surplus of the former. Subsequently, the proportion of the sum of all cross-dimensional compensation flows to the total flow is calculated as the compensation contribution, and the calculation method is as follows: in For the compensation contribution, the numerator is the sum of the flows of all off-diagonal elements, i.e., the total mass flow of actual cross-dimensional compensation, and the denominator is the global total flow. The extracted coordinates of all compensation path nodes, the flow values ​​of each path, and the compensation contribution are structured and packaged to generate compensation insight data.

[0051] In the aforementioned multi-source heterogeneous data fusion-based intelligent talent profiling analysis system 100, the report generation module 150 is used to perform semantic analysis and business language transformation on the compensation logic relationships in the compensation insight data, and integrate and encapsulate the flexible matching score and the corresponding strength and weakness compensation path to obtain a talent profiling analysis report. It should be noted that since both the compensation insight data and the flexible matching score are stored in numerical matrix coordinates and scalar form, business decision-makers cannot directly understand the candidate's ability compensation logic and job suitability conclusions from them. It is necessary to transform the mathematical results into human-readable business language and integrate them into a standardized decision document. Based on this, the technical solution of this application further performs semantic analysis and business language transformation on the compensation logic relationships in the compensation insight data, and integrates and encapsulates the flexible matching score and the corresponding strength and weakness compensation path to obtain a talent profiling analysis report. Through the above processing, the output of the underlying algorithm can be transformed into a structured report containing scoring conclusions and explanations of compensation paths, which can be directly used for talent selection decisions.

[0052] More specifically, in a concrete example of this application, the system first receives compensation insight data, reads the coordinates of all cross-dimensional compensation path nodes and their corresponding flow values, and inputs the source dimension skill identifier, target dimension skill identifier, and flow intensity of each compensation path into a natural language generation engine. Based on a preset business logic semantic template, the engine transforms the mathematical path coordinates into descriptive business statements. For example, it transforms the flow records from the power system steady-state analysis dimension to the UHV equipment operation and maintenance dimension into the semantic expression that the candidate's accumulated capabilities in power system steady-state analysis can effectively compensate for their lack of experience in UHV equipment operation and maintenance. Simultaneously, it transforms the compensation contribution into a general conclusion that the candidate's comprehensive capabilities present cross-domain composite compensation characteristics. All generated semantic text is then summarized and packaged into a profile diagnostic corpus.

[0053] Subsequently, the system receives the profiling diagnostic corpus and flexible matching scores, and performs adaptive layout rendering and formatting based on a pre-set talent selection report template. The flexible matching scores are filled into the template's scoring summary area, compensatory explanatory semantic text is filled into the capability analysis area, and cross-dimensional compensation paths are rendered as flow annotations in the template's capability radar chart area. The thickness and color depth of the annotation lines are adaptively adjusted according to the flow intensity of each compensation path. After completing the content filling and layout adaptation for all areas, the report is formatted and packaged to generate the final talent profiling analysis report.

[0054] In summary, a digital talent profiling analysis system based on multi-source heterogeneous data fusion according to embodiments of this application is presented. It abstracts talent capabilities and job requirements into probability quality distributions and introduces optimal transmission theory to achieve flexible compensation matching across skill dimensions, solving the rigid adaptation problem of traditional vector similarity algorithms, which cannot model the substitution effect between skills due to the orthogonal assumption of dimensions. The system first performs skill entity recognition and intensity mapping on quantitative performance data, project logs, and job description text to generate distribution vectors of talent capabilities and job requirements. Then, it calculates the topological distance between skills based on the skill ontology graph and constructs a compensation impedance matrix as a transport cost function. The optimal transmission is solved through entropy regularization to obtain the transmission planning tensor. Finally, the tensor is matched with a score mapping and a compensatory path is extracted to generate a talent profiling analysis report containing flexible scoring and compensation explanations. Simultaneously, addressing the problem of distance overestimation in shortest path algorithms due to their inability to perceive skill collaboration bridging relationships, neighborhood co-occurrence statistics and normalized point mutual information quantification are introduced. Collaborative bridging gating is used to adaptively correct the topological distance, reasonably reducing the cost of cross-domain capability transfer for multi-skilled talents.

Claims

1. A digital talent profiling and analysis system that integrates multi-source heterogeneous data, characterized in that, include: The feature extraction module is used to perform skill entity recognition and feature intensity mapping on the quantitative performance data, project development log text and standard job description text contained in the collected multi-source heterogeneous data to obtain talent ability distribution vector and job demand distribution vector. The impedance calculation module is used to perform graph node alignment and weighted shortest path distance calculation on the talent ability distribution vector and job requirement distribution vector based on the preset skill ontology graph to obtain the compensation impedance matrix that represents the ability substitution cost between different skill dimensions. The transmission planning module is used to perform flexible adaptive path planning on the talent capability distribution vector and job demand distribution vector, using the compensation impedance matrix as the cross-dimensional transportation cost function, to obtain the transmission planning tensor. The matching evaluation module is used to evaluate the flexible matching score and compensation effectiveness of the transmission planning tensor to obtain compensation insight data and flexible matching score. The report generation module is used to perform semantic parsing and business language transformation on the compensation logic relationships in the compensation insight data, and to integrate and encapsulate the flexible matching score with the corresponding strength and weakness compensation path to obtain a talent profile analysis report.

2. The digital talent profiling and analysis system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The feature extraction module includes: The entity recognition unit is used to perform skill entity extraction and synonym merging and cleaning on project development log text and standard job description text respectively to obtain the talent skill entity set and the job requirement entity set. The intensity mapping unit is used to calculate the demand intensity weight of the set of job demand entities to obtain the original vector of job demand, and to perform logarithmic weighted aggregation on the set of talent skill entities based on quantitative performance data to obtain the original vector of talent ability. The distribution transformation unit is used to perform probability quality distribution transformation on the original vector of talent ability and the original vector of job requirements to obtain the distribution vector of talent ability and the distribution vector of job requirements.

3. The digital talent profiling and analysis system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The impedance calculation module includes: The dimension alignment unit is used to perform skill dimension alignment and ontology reference mapping on the talent ability distribution vector and the job requirement distribution vector to obtain a set of skill features. The path estimation unit is used to perform graph topology-based shortest path estimation on the skill feature set based on a preset skill ontology graph to obtain a skill topology distance matrix. The cost transformation unit is used to perform nonlinear compensation cost transformation on each distance element in the skill topology distance matrix to obtain the compensation impedance matrix.

4. The digital talent profiling and analysis system based on multi-source heterogeneous data fusion according to claim 3, characterized in that, The cost transformation unit includes: performing nonlinear compensation cost transformation on each distance element in the skill topology distance matrix using the following formula: in, This represents the shortest path distance between nodes in the skill topology distance matrix. This is the proportional sensitivity coefficient hyperparameter.

5. The digital talent profiling and analysis system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The transmission planning module includes: The target construction unit is used to construct a regularized objective function based on the compensation impedance matrix as the basic cost term and superimposed with the information entropy regularization term. At the same time, based on the talent ability distribution vector and the job demand distribution vector, row and column constraints are set respectively to obtain the supply and demand boundary constraint set. The iterative solution unit is used to construct the Gibbs kernel matrix based on the regularization objective function, and to alternately update the left and right dual scaling vectors under the row and column constraints of the supply and demand boundary constraint set through an iterative scaling algorithm until convergence, thereby obtaining the optimal joint distribution matrix. Tensor encapsulation unit is used to perform zero-setting filtering on near-zero floating-point drift terms in the optimal joint distribution matrix caused by iterative truncation, and to perform high-dimensional tensor encapsulation along the batch and scene feature dimensions to obtain the transport planning tensor.

6. The digital talent profiling and analysis system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The matching evaluation module includes: The cost aggregation unit is used to perform total transportation cost aggregation calculation on the transportation planning tensor to obtain the global transportation cost and flow component matrix; The scoring mapping unit is used to perform a non-linear mapping on the global transport cost to obtain the flexible matching score; The insight extraction unit is used to extract all non-zero cross-dimensional compensation paths and their traffic proportions after masking the main diagonal of the flow component matrix to obtain compensation insight data.

7. The digital talent profiling and analysis system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The report generation module includes: The semantic parsing unit is used to perform semantic parsing of the compensation logic on the compensation insight data to obtain the profile diagnosis corpus package; The report encapsulation unit is used to adaptively render and format the compensatory explanatory semantic text, quantitative flexible scoring, and cross-dimensional compensation path contained in the profile diagnosis corpus based on a preset talent selection report template to obtain a talent profile analysis report.

8. The digital talent profiling and analysis system based on multi-source heterogeneous data fusion according to claim 3, characterized in that, The path estimation unit includes: The distance statistics subunit is used to perform original topological distance calculation and neighborhood co-occurrence frequency statistics on the skill feature set based on the preset skill ontology graph to obtain the original topological distance matrix and the neighborhood co-occurrence frequency matrix. The affinity quantification subunit is used to perform cooperative affinity quantification on the neighborhood co-occurrence frequency matrix based on normalized point mutual information to obtain the cooperative affinity matrix. The distance correction subunit is used to perform cooperative bridging gating on the original topological distance matrix based on the cooperative affinity matrix to obtain the skill topological distance matrix.

Images