Knowledge Graph-Based Talent Skill Gap Diagnosis and Recommendation System

By constructing a two-layer talent knowledge graph and performing temporal decay and meta-capability decoupling, the problems of misdiagnosis of capability gaps and rating bias in cross-domain talent management are solved, and the accurate quantification of cross-domain capability gaps and the optimized recommendation of learning paths are realized.

CN123070901APending Publication Date: 2026-11-10STATE GRID HENAN INFORMATION & TELECOMM CO +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611251516.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-18
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing knowledge graph-based talent management systems suffer from misdiagnosis of competency gaps and rating biases when transferring talent across domains. They cannot effectively identify cross-domain transferable meta-competencies, and static graphs fail to characterize skill decay, leading to inaccurate learning resource recommendations.

Method used

A two-layer talent knowledge graph is constructed. The node state and edge weight are updated through the time-series decay module. The cross-layer information bottleneck is decoupled by the meta-capability extraction module to generate cross-domain meta-capability vectors. The optimized learning path is output through the path generation module.

Benefits of technology

It achieves accurate quantification of cross-domain capability gaps and efficient recommendation of learning paths, reduces misdiagnosis and rating bias in cross-domain mobilization, and improves the accuracy and efficiency of resource recommendation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure FT_1
    Figure FT_1
  • Figure FT_2
    Figure FT_2
  • Figure FT_3
    Figure FT_3
Patent Text Reader

Abstract

This application provides a knowledge graph-based talent capability gap diagnosis and recommendation system. First, it unifies multi-source data into a two-layer talent knowledge graph, simultaneously constructing a job capability subgraph and a learning resource subgraph, vertically linking surface-level business skills with underlying meta-capabilities within the same graph. Then, it updates node states and edge weights using a decay function determined by interaction timestamps and capability half-life, automatically degrading skills that have been inactive for a long time. Subsequently, it decouples the two-layer features through a comparative information network containing cross-layer information bottlenecks, filtering out domain-specific noise and extracting reusable meta-capability vectors across domains. These vectors are then transformed into a job capability gap matrix through nonlinear mapping and domain-adaptive distance measurement, avoiding gap exaggeration caused by subgraph isolation. Finally, using the gap matrix as the state and the learning resource subgraph as the restricted action space, it performs constrained path search and branch pruning, outputting a cross-domain learning path that satisfies prior relationships and has the optimal global benefit.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and graph data processing technology, and more specifically, to a knowledge graph-based talent gap diagnosis and recommendation system. Background Technology

[0002] As enterprises advance digital transformation and business lines iterate at an accelerated pace, cross-domain talent mobility has become a frequent occurrence within organizations. Traditional talent management methods, relying on supervisor experience and manual data entry, are no longer sufficient to support the needs of large-scale competency assessments. Enterprises accumulate a large amount of heterogeneous data from multiple sources, including job descriptions, employee performance, training records, and code submissions. If this data can be organized into a structured network, it is possible to quantitatively assess the gap between an employee's current capabilities and the requirements of their target position, and then match appropriate courses and practical projects accordingly.

[0003] Existing systems of this type typically organize graphs using triplets of employees, skills, and positions. They calculate capability gaps through path similarity or node vector distances in graph neural networks, and then retrieve learning resources based on similarity. While this approach is usable within the same domain, it leads to significant misdiagnosis of gaps when transferring between domains. This is because their graphs are often single-layered and flat, with skill nodes from different business domains at the same level and lacking topological pathways. Shortest paths or random walks cannot establish connections between them. At the representation level, the target position's skills and the employee's existing skills belong to non-overlapping local subgraphs, and the vectors obtained from neighbor aggregation fall into separate feature clusters with extremely low similarity, creating geometric isolation. More importantly, existing models lack mathematical abstraction of domain-invariant features independent of specific business operations, failing to reconstruct transferable meta-capabilities such as architectural design and distributed thinking from behavioral data. Ultimately, this results in senior engineers being judged as having almost no mismatch in their job transition needs. Furthermore, static graphs do not depict the decay of skills over time due to inactivity, further amplifying rating bias. In the path recommendation stage, even if causal inference is introduced to remove confounding variables, the effect estimation remains at the level of a single learning action. It assumes that resources are independent of each other, making it impossible to identify the synergistic amplification or cognitive overload inhibition relationship between adjacent courses, and it is also difficult to use a fixed threshold to adapt to the gap span of different employees. This results in low-scoring and efficient paths being mistakenly cut off, while high-scoring and inefficient paths are retained.

[0004] Therefore, an optimized knowledge graph-based talent gap diagnosis and recommendation system is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a knowledge graph-based talent skill gap diagnosis and recommendation system, which includes: The graph construction module is used to perform graph association modeling on the received multi-source talent data to obtain a two-layer talent knowledge graph, a job competency subgraph, and a learning resource subgraph. The time-decay module is used to update the node state and assign weights to the edge decay of the two-layer talent knowledge graph based on the half-life decay function determined by the training interaction timestamp and the preset capability half-life, so as to obtain the time-decayed talent knowledge graph. The meta-capability extraction module is used to decouple cross-layer information bottleneck features and extract meta-capabilities from a time-decaying talent knowledge graph based on a comparative information network containing cross-layer information bottlenecks, so as to obtain a cross-domain meta-capability vector. The gap quantification module is used to perform domain-adaptive space mapping and dynamic quantification of capability gaps on cross-domain meta-capability vectors based on job capability subgraphs to obtain job capability gap matrices. The path generation module is used to perform constrained path search and branch pruning on the job competency gap matrix and learning resource subgraph to output cross-domain learning paths.

[0006] Compared with existing technologies, the talent capability gap diagnosis and recommendation system based on knowledge graphs provided in this application first organizes multi-source talent data scattered across performance ratings, code submissions, training interactions, and job descriptions into a unified two-layer talent knowledge graph. Simultaneously, it establishes job capability subgraphs and learning resource subgraphs, enabling surface-level business skills and underlying meta-capabilities to achieve vertical correlation within the same graph, thus providing a common foothold for cross-domain comparisons. Based on this, a decay function determined by interaction timestamps and capability half-life is introduced to update node states and edge weights temporally, automatically degrading skills that have not been activated for a long time, resulting in a temporal decay graph reflecting the true capability status. Subsequently, a comparison information network containing cross-layer information bottlenecks is used to decouple the two-layer features, filtering out noise components bound to specific business domains and extracting reusable meta-capability vectors across domains. These vectors are then nonlinearly mapped into the feature space of the target job, and the capability differences are transformed into a computable job capability gap matrix through a domain-adaptive distance metric, avoiding the amplification of cross-domain gaps due to the isolation of subgraphs. Finally, using the gap matrix as the state and the learning resource subgraph as the restricted action space, constrained path search and branch pruning are performed. Under the premise of satisfying the prior relationship, the cross-domain learning path with the best global benefit is converged, so that the diagnostic results can be directly transformed into an executable training program. 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 knowledge graph-based talent gap diagnosis and recommendation system according to an embodiment of this application; Figure 2 This is a schematic diagram of data flow in a knowledge graph-based talent gap diagnosis and recommendation system according to an embodiment of this application; Figure 3 This is a block diagram of the meta-capability extraction module in the knowledge graph-based talent capability gap diagnosis and recommendation system according to an embodiment of this application; Figure 4 This is a block diagram of the gap quantification module in the knowledge graph-based talent gap diagnosis and recommendation system according to an embodiment of this application; Figure 5 This is a block diagram of the path generation module in a knowledge graph-based talent gap diagnosis and recommendation system according to an embodiment of this application; Figure 6 This is an interactive modulation pruning block diagram of the causal pruning unit in a knowledge graph-based talent gap diagnosis and recommendation system 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 talent gap diagnosis methods often use flat, single-layer knowledge graphs to organize the association between employees and skills. Skill nodes in different business domains lack topological pathways and belong to non-overlapping local subgraphs. When transferring between domains, the vector distance tends to be extremely low, significantly exaggerating the gap. At the same time, static graphs do not depict the decay caused by long-term inactivity of skills, resulting in overestimation of ability ratings. The learning resources retrieved based on this are also difficult to match actual training needs. Therefore, the technical solution of this application proposes a talent gap diagnosis and recommendation system based on knowledge graphs. First, graph association modeling is performed on the received multi-source talent data to obtain a two-layer talent knowledge graph consisting of a skill layer and a meta-capability layer, as well as a job capability subgraph and a learning resource subgraph, so that the surface-level business skills and the underlying meta-capabilities form a vertical association in the same graph. Then, based on the half-life decay function determined by the training interaction timestamp and the preset capability half-life, the node state of the two-layer talent knowledge graph is updated and the edge decay weight is assigned to obtain a time-decayed talent knowledge graph that reflects the current capability status. Subsequently, based on the comparative information network containing cross-layer information bottlenecks, the cross-layer information bottleneck features are decoupled and meta-capabilities are extracted from the time-decayed talent knowledge graph, and noise components bound to specific business domains are filtered out to obtain cross-domain meta-capability vectors that can be reused across domains. Then, based on the job capability subgraph, the cross-domain meta-capability vectors are mapped to the domain adaptation space and the capability gap is dynamically quantified to obtain the job capability gap matrix in a unified feature space, avoiding the exaggeration of gaps caused by the isolation of subgraphs. Finally, constrained path search and branch pruning are performed on the job capability gap matrix and the learning resource subgraph to output the cross-domain learning path under the premise of satisfying the prior relationship.

[0014] Figure 1 This is a system block diagram of a knowledge graph-based talent gap diagnosis and recommendation system according to an embodiment of this application. Figure 2 This is a schematic diagram illustrating the data flow of a knowledge graph-based talent gap diagnosis and recommendation system according to an embodiment of this application. Figure 1 and Figure 2As shown, the knowledge graph-based talent capability gap diagnosis and recommendation system 100 according to an embodiment of this application includes: a graph construction module 110, used to perform graph association modeling on received multi-source talent data to obtain a two-layer talent knowledge graph, a job capability subgraph, and a learning resource subgraph; a time-series decay module 120, used to update the node state and assign edge decay weights to the two-layer talent knowledge graph based on a half-life decay function determined by training interaction timestamps and preset capability half-life to obtain a time-series decayed talent knowledge graph; a meta-capability extraction module 130, used to decouple cross-layer information bottleneck features and extract meta-capabilities from the time-series decayed talent knowledge graph based on a comparative information network containing cross-layer information bottlenecks to obtain a cross-domain meta-capability vector; a gap quantification module 140, used to perform domain adaptation space mapping and capability gap dynamic quantification on the cross-domain meta-capability vector based on the job capability subgraph to obtain a job capability gap matrix; and a path generation module 150, used to perform constrained path search and branch pruning on the job capability gap matrix and the learning resource subgraph to output a cross-domain learning path.

[0015] In the aforementioned knowledge graph-based talent gap diagnosis and recommendation system 100, the graph construction module 110 is used to perform graph association modeling on the received multi-source talent data to obtain a two-layer talent knowledge graph, a job competency subgraph, and a learning resource subgraph. It should be noted that, because employee competency information is scattered across independent data sources such as performance records, code repositories, and training ledgers, job requirements exist in natural language text form, and the sequential dependencies between learning resources are not structured, there is a lack of a unified association carrier to support cross-domain competency comparison. Based on this, the technical solution of this application first performs graph association modeling on the received multi-source talent data to obtain a two-layer talent knowledge graph, a job competency subgraph, and a learning resource subgraph. Through the above processing, heterogeneous talent data, job requirements, and training resources can be effectively incorporated into the same graph structure, providing a unified data foundation for subsequent temporal decay weighting, meta-competency decoupling, and gap quantification.

[0016] More specifically, in a specific example of this application, the graph construction module includes: a feature tensor quantization unit, used to extract semantic features and perform numerical normalization on employee project performance scores, code submission sequence text, training interaction timestamps, historical learning pre- and post-learning ability assessment values, target job ability description text, course attributes, practical project attributes, and resource prerequisite relationships from multi-source talent data to obtain employee behavior feature tensors, job semantic feature tensors, and training resource feature tensors; a subgraph topology mapping unit, used to instantiate skill layer nodes and meta-ability layer nodes of the employee behavior feature tensor and establish inter-layer vertical connections, establish unidirectional connections from the job center node to the skill leaf node of the job semantic feature tensor, and inject prerequisite connections and learning effect edges into the training resource feature tensor to generate an initial employee two-layer graph, a job ability subgraph, and a learning resource subgraph, respectively; and a representation aggregation unit, used to perform two-layer graph feature encoding and representation aggregation on the initial employee two-layer graph to obtain a two-layer talent knowledge graph. The multi-source talent data includes employee project performance scores, code submission sequence text, training interaction timestamps, historical pre- and post-learning ability assessment values, target job ability description text, course attributes, practical project attributes, and resource prerequisite relationships; the two-layer talent knowledge graph consists of a skill layer and a meta-ability layer; the learning resource subgraph includes learning resource prerequisite edges and learning effect edges.

[0017] Specifically, the system first receives multi-source talent data and splits it according to data source tags, resulting in three data streams: employee-level interaction records, job-level descriptive text, and corporate training resource records. For the employee-level dimension, employee project performance scores are aligned by assessment cycle and subjected to min-max normalization to unify scores from different scoring systems to the 0-1 range. Code submission sequence text is fed into a pre-trained bidirectional Transformer encoder to extract semantic features of the technology stack and engineering practices from submission information, changed file paths, and dependency references, identifying skill evidence such as distributed message middleware calls and service registration and discovery configurations. Training interaction timestamps are converted to standardized values ​​according to a unified time zone and subtracted from historical pre- and post-learning ability assessment values ​​to obtain the ability increment for a single training session. These normalized values ​​and semantic vectors are concatenated according to employee identifiers to form an employee behavior feature tensor. For the job-level dimension, dependency parsing is performed on the target job ability description text to extract triples consisting of ability items and mastery requirements, which are then mapped to a high-dimensional vector space to form a job semantic feature tensor. For the education and training dimension, knowledge extraction is performed on course attributes, practical project attributes, and resource prerequisite relationships. The course length, ability tags, delivery form, and prerequisite resource identifiers are encoded into multi-dimensional features to form a training resource feature tensor.

[0018] After tensorizing the three features mentioned above, topological mapping is performed on them respectively. The employee behavior feature tensor is parsed, and skill nodes corresponding to specific business technology stacks are instantiated at the skill layer. Meta-capability nodes unrelated to business are instantiated at the meta-capability layer. Meta-capability nodes cover distributed system thinking, architectural abstraction capabilities, and understanding of concurrency control, etc. Vertical connections between the two layers are established based on the co-occurrence frequency threshold of skill evidence and meta-capability tags. At the same time, same-layer association edges are established within the skill layer based on the co-occurrence frequency of two nodes in the same code commit or the same delivery project. Same-layer association edges are also established within the meta-capability layer based on the hierarchical and adjacency relationships in the meta-capability ontology. The initial weight of each same-layer edge is taken from the normalized co-occurrence frequency, so that skill nodes belonging to different technology stacks, such as Spring Cloud service governance and Kubernetes orchestration and scheduling, can obtain a through path through the same distributed system thinking node, generating an initial two-layer employee graph. The job semantic feature tensor is analyzed, with the target job entity as the central node and the extracted specific ability requirements as skill leaf nodes. One-way edges are established from the central node to the leaf nodes, and mastery level requirements are attached to these edges, generating a job ability subgraph. The training resource feature tensor is analyzed, and course nodes and practice project nodes are instantiated. Edges representing mandatory prerequisites are injected based on resource pre-learning relationships, and edges representing expected learning outcomes are injected based on the statistical difference in ability assessment values ​​before and after historical learning, generating a learning resource subgraph.

[0019] Based on the initial two-layer graph of employees, it is fed into a graph attention encoder to perform feature encoding and representation aggregation. For any central node, its own features and the features of its first-order neighbors, including those across layers, are first projected onto the same attention computation space using a shared weight matrix. The projection results are then concatenated, mapped to a scalar using a learnable parameter vector, and activated with a leaky modified linear unit. The normalized attention distribution is then solved within the node's neighborhood using the Softmax function. Finally, the projected features of the neighbors are weighted and summed according to this distribution, and then nonlinearly activated to obtain the updated node representation, which is mathematically expressed as:

[0020] in, Indicates the first Layer center node Updated feature representation vector, Represents a non-linear activation function. Indicates the relationship with the central node There exists a set of first-order neighbor nodes connected by edges, including nodes connected across different levels. This represents a modified linear unit activation function with leakage. This represents the transpose of the learnable parameter vector in the attention mechanism. The shared weight matrix represents the feature transformation of nodes across layers. Indicates the first Layer center node The feature representation vector, Indicates the first Layer neighbor nodes The feature representation vector, Indicates the first Layers are used for denominator normalization of neighboring nodes within the same neighborhood. The feature representation vector, This represents a tensor concatenation operation along the feature dimension. This represents an exponential function with the natural constant as its base. The fractional part represents the neighboring nodes. Relative to the central node The normalized attention coefficient represents the weight of the neighbor in the central node's representation update process. Therefore, the contribution of frequently validated skill nodes in the skill layer to their connected meta-capability nodes is higher than that of occasionally occurring skill nodes. After completing multiple rounds of neighborhood aggregation, the representations of the meta-capability layer nodes have incorporated information from skill nodes in different business domains. This graph structure is output as a two-layer talent knowledge graph, which, together with the aforementioned job capability subgraph and learning resource subgraph, serves as the result of graph association modeling.

[0021] In the aforementioned knowledge graph-based talent competency gap diagnosis and recommendation system 100, the time-series decay module 120 is used to update the node states and assign decay weights to the edges of the two-layer talent knowledge graph based on a half-life decay function determined by training interaction timestamps and preset competency half-lives, thereby obtaining a time-series decayed talent knowledge graph. It should be noted that because skill mastery decays over time, the two-layer talent knowledge graph exhibits a static topology after modeling, with skills that have not been used for a long time retaining their initial weights, resulting in an overestimation of competency ratings. Therefore, the technical solution of this application further updates the node states and assigns decay weights to the edges of the two-layer talent knowledge graph based on a half-life decay function determined by training interaction timestamps and preset competency half-lives. Through this processing, the skill states in the graph can be effectively aligned with the actual competency levels currently held by employees, providing time-series corrected input for subsequent meta-competency decoupling.

[0022] More specifically, in a specific example of this application, the time-decay module includes: a timestamp alignment unit, used to extract the absolute time of the most recent interaction from each skill node in the two-layer talent knowledge graph, convert it into a standardized timestamp, and append the standardized timestamp to the node attribute vector to obtain a two-layer graph with timestamp attributes; a depreciation coefficient calculation unit, used to calculate the capability depreciation coefficient based on the half-life function on the two-layer graph with timestamp attributes to obtain a node depreciation coefficient matrix; and an edge weighting update unit, used to dynamically update the initial edge weights in the two-layer graph with timestamp attributes based on the node depreciation coefficient matrix to obtain a time-decayed talent knowledge graph.

[0023] Specifically, the process begins by parsing the graph attributes and traversing all nodes in the skill layer and meta-capability layer of the two-layer talent knowledge graph. For each skill node, the absolute time of its most recent use or verification is retrieved from the underlying interaction records. The retrieval scope includes the project delivery date, the last submission time of the corresponding code module, and the completion time of the relevant training. The latest of these is taken as the most recent interaction time for that node and converted into a standardized timestamp according to a unified time zone. For each meta-capability node, the latest of the most recent interaction times among all skill nodes connected by vertical edges between layers is taken as the most recent interaction time for that meta-capability node. Its half-life constant is determined by the larger value of the meta-capability's sedimentation attribute. It should be noted that the training interaction timestamp mentioned in this application refers to the interaction timestamp of a skill being used or verified, and its value is the latest of the project delivery date, the last submission time of the code module, and the training completion time. The standardized timestamp is appended as a temporal feature attribute to the attribute vector of the corresponding node, allowing the node to retain its original semantic representation while carrying a temporal state that can be used for time difference calculation, thus obtaining a two-layer graph with timestamp attributes. For skill nodes that have not participated in delivery for a long time and have no corresponding code submission, the interval between their timestamp and the current diagnosis time is relatively large, and this interval is the direct basis for subsequent loss calculation.

[0024] After obtaining the two-layer graph with timestamp attributes, the temporal characteristic attributes of each node are read and capability depreciation calculation based on the half-life function is performed. Specifically, the depreciation coefficient calculation unit includes: an idle duration calculation subunit, used to calculate the time difference between the standardized timestamp of each node in the two-layer graph with timestamp attributes and the current diagnosis time to obtain the node idle duration of each node; a decay rate determination subunit, used to perform halving rate calibration on the preset capability half-life corresponding to each node to obtain the node decay rate of each node; and an exponential depreciation mapping subunit, used to perform negative exponential decay mapping on the node idle duration of the corresponding node based on the node decay rate of each node to obtain the node depreciation coefficient matrix.

[0025] Specifically, for any skill node, first calculate the time difference between the current diagnosis time and the most recent interaction timestamp of that node. Then, use the ratio of the natural logarithm of 2 to the preset half-life constant of the skill as the decay rate. After negativeing ​​the product of the rate and the time difference, map it using the natural exponential function to obtain a dynamic capability degradation coefficient greater than 0 and less than or equal to 1. Its mathematical expression is:

[0026] in, Indicates the first Skill Nodes The dynamic capability loss coefficient, This represents an exponential function with the natural constant as its base. Represents the natural logarithm of 2 and is used to calibrate the halving rate. Indicates the first Each skill node has an inherent half-life constant, measured in days. This represents the absolute timestamp of the current execution capability gap diagnosis. This represents the absolute timestamp of the most recent interaction extracted from the node attribute vector. In the formula... The quotient of the half-life constant represents the exponential decay rate of the skill. Multiplying this by the time difference indicates the number of times the skill has halved. Therefore, for every half-life, the decay coefficient is reduced to half its original value. The half-life constant is set according to the skill attributes. Skills that are easily forgotten, such as programming language syntax and framework interfaces, have smaller values, while skills that are more deeply ingrained, such as distributed system design concepts, have larger values. This causes Spring Cloud configuration skills and message middleware maintenance skills, which have been idle for two years, to experience varying degrees of decay. Furthermore, the distributed system thinking and meta-capabilities they support are correspondingly decayed due to the decrease in the weight of inter-layer edges. The decay coefficients calculated for each node are assembled into a node decay coefficient matrix according to the node index.

[0027] Based on the obtained node loss coefficient matrix, all initial edges and their weights between the skill layer and the meta-ability layer, as well as between nodes in the same layer, are extracted from the two-layer graph with timestamp attributes. Dynamic penalty updates are then applied to the edge weights. For any edge, the loss coefficients of its starting node and target node are taken. These two coefficients are then weighted and fused according to a preset direction penalty bias factor to obtain the joint loss weight of the edge. The initial edge weight and the joint loss weight are then multiplied element-wise. The mathematical expression is as follows:

[0028] in, This indicates that the node will be updated. Pointing to node Edge weights, This indicates a two-layer graph with timestamp attributes consisting of nodes. Pointing to node The initial edge weights, This represents the direction penalty bias factor, and its value ranges from 0 to 1. This represents the dynamic capability loss coefficient of the starting node. This represents the dynamic capability degradation coefficient of the target node. The weighted sum within parentheses in the formula is a joint representation of the degradation status of the nodes at both ends of the connection. The bias factor controls the proportion of the starting node and the target node in the penalty calculation, causing the weight of the vertical connection from the skill node to the meta-capability node to decrease when the skill is idle for a long time. As a result, the meta-capability node no longer receives support from the evidence of the expired skill. The topology structure after the weight update is encapsulated and output as a time-decayed talent knowledge graph.

[0029] In the aforementioned knowledge graph-based talent gap diagnosis and recommendation system 100, the meta-capability extraction module 130 is used to decouple cross-layer information bottleneck features and extract meta-capabilities from the time-decaying talent knowledge graph based on a comparative information network containing cross-layer information bottlenecks to obtain a cross-domain meta-capability vector. It should be noted that, given that the skill layer node representations in the time-decaying talent knowledge graph are still bound to specific business technology stacks and carry domain-specific components such as framework interfaces and configuration syntax, nodes belonging to different business domains fall into non-overlapping feature clusters, and employees' truly transferable distributed system thinking and architectural abstraction capabilities are masked by surface skill details. Based on this, the technical solution of this application further decouples cross-layer information bottleneck features and extracts meta-capabilities from the time-decaying talent knowledge graph based on a comparative information network containing cross-layer information bottlenecks, compressing domain-specific components in the surface features while retaining the two-layer mapping relationship to obtain a cross-domain meta-capability vector. Through the above processing, capabilities with different technology stacks but similar underlying cognitive requirements can be effectively compared in the same representation space, providing input unaffected by domain noise for subsequent gap quantification.

[0030] Figure 3 This is a block diagram of the meta-capability extraction module in a knowledge graph-based talent capability gap diagnosis and recommendation system according to an embodiment of this application. Figure 3As shown, the meta-capability extraction module 130 includes: a view separation encoding unit 131, used to separate the time-decayed talent knowledge graph into a surface business skills view and a bottom-level implicit meta-capability view according to node type labels, and to perform initial encoding on them to obtain a surface business feature matrix and a bottom-level implicit feature matrix; a feature decoupling unit 132, used to perform cross-layer information bottleneck feature decoupling on the surface business feature matrix and the bottom-level implicit feature matrix based on a comparative information network containing cross-layer information bottlenecks to obtain decoupling latent variable distribution parameters; and a reparameterization sampling unit 133, used to perform meta-capability feature reparameterization sampling on an auxiliary noise vector randomly drawn from a standard normal distribution based on the decoupling latent variable distribution parameters to obtain a cross-domain meta-capability vector.

[0031] In the aforementioned knowledge graph-based talent capability gap diagnosis and recommendation system 100, the view separation encoding unit 131 is used to separate the time-decaying talent knowledge graph into a surface-level business skills view and a bottom-level implicit meta-capability view according to node type labels, and to perform initial encoding on them to obtain a surface-level business feature matrix and a bottom-level implicit feature matrix. It should be noted that, since skill layer nodes and meta-capability layer nodes in the time-decaying talent knowledge graph are mixed and stored in the same heterogeneous topology, the feature semantics and abstract levels of the two types of nodes are different. Directly inputting them into the comparative information network cannot define the source of positive and negative samples for mutual information constraints. Based on this, the technical solution of this application further separates the time-decaying talent knowledge graph into a surface-level business skills view and a bottom-level implicit meta-capability view according to node type labels, and performs initial encoding on them. Through the above processing, two feature inputs with clear hierarchy and consistent dimensions can be effectively provided for cross-layer decoupling.

[0032] More specifically, in a concrete example of this application, the heterogeneous topology of the time-decaying talent knowledge graph is parsed, and the node type labels written by each node during the modeling phase are read. Nodes labeled as business skills and their corresponding edges are assigned to the surface business skills view, while nodes labeled as meta-capabilities and their corresponding edges are assigned to the underlying implicit meta-capability view. The vertical edges between the skill layer and the meta-capability layer are not included in the neighborhood calculation within the view but are recorded separately as inter-layer correspondences for subsequent determination of positive sample pairings between the surface features and underlying features of the same employee. After view separation, the surface business skills view contains specific technology stack nodes such as Spring Cloud service governance, Kubernetes orchestration and scheduling, and message middleware operation and maintenance, while the underlying implicit meta-capability view contains nodes unrelated to business, such as distributed system thinking, architectural abstraction capabilities, and concurrency control understanding.

[0033] Based on the completed view separation, spatial domain graph convolution operations are performed on both views. For any view, adjacency relationships are first constructed from the edges of that view and then symmetric normalization is performed. The edge weights used in the normalization process are the weights updated after half-life degradation, so that skills that have not been validated for a long time contribute less to neighborhood aggregation. Then, the first-order and second-order neighbor features of each node are aggregated layer by layer. The aggregation results are transformed by a learnable weight matrix and nonlinear activation is applied to transform the originally sparse node attributes into a continuous dense representation. After performing the above operations on the surface business skill view, the representation of each skill node integrates the information of adjacent technology stacks in the same domain, and the output is the surface business feature matrix. After performing the above operations on the bottom implicit meta-capability view, the representation of each meta-capability node integrates the information of similar cognitive dimensions, and the output is the bottom implicit feature matrix. The two matrices are output with the same hidden layer dimension, and the row indexes correspond to the skill node index and the meta-ability node index, respectively. Combined with the inter-layer correspondence of the aforementioned records, the row of the skill node under the same employee name and the row of the meta-ability node connected by the inter-layer vertical edge can be directly retrieved and form a cross-layer positive sample pair.

[0034] In the aforementioned knowledge graph-based talent gap diagnosis and recommendation system 100, the feature decoupling unit 132 is used to perform cross-layer information bottleneck feature decoupling on the surface business feature matrix and the bottom latent feature matrix based on a contrastive information network containing cross-layer information bottlenecks to obtain decoupled latent variable distribution parameters. It should be noted that since the surface business feature matrix simultaneously mixes transferable cognitive components and domain-specific components attached to a specific technology stack, neighborhood aggregation alone cannot separate the two. Therefore, the technical solution of this application further performs cross-layer information bottleneck feature decoupling on the surface business feature matrix and the bottom latent feature matrix based on a contrastive information network containing cross-layer information bottlenecks. Through the above processing, the domain-specific components can be effectively suppressed while preserving the mapping relationship between the two layers, outputting latent variable distribution parameters available for sampling.

[0035] More specifically, in a concrete example of this application, the surface business feature matrix and the underlying latent feature matrix are input together into a pre-constructed contrastive information network for joint optimization. The surface features are mapped to surface representations in the contrastive space via a projection head, and the underlying features are mapped to posterior distribution parameters of latent variables via a variational encoder. Based on the aforementioned recorded inter-layer vertical correspondence, positive sample pairs are determined; that is, the specific skill representation of the same employee and its supporting meta-capability representation constitute positive samples, while the underlying representations of other employees or other meta-capability dimensions constitute negative samples. Based on this, a cross-layer information bottleneck objective function is constructed, mathematically expressed as:

[0036] in, This represents the total loss scalar of the cross-layer information bottleneck network in the current training iteration. This represents the total number of positive sample pairs processed in the current training batch. Represents the first element in the surface business feature matrix. The feature representation vector of each node Represents the underlying latent feature matrix and The underlying feature representation vector that constitutes positive sample pairing. This indicates the first sample drawn from the negative sample queue. The underlying latent feature representation vectors and are... This does not constitute a positive sample pairing. This represents the cosine similarity calculation function, which takes the ratio of the dot product of two vectors to the product of their magnitudes. This indicates a temperature hyperparameter with a value greater than 0, used to adjust the sharpness of the similarity distribution. This represents the total number of negative samples that match each positive sample. This represents the information bottleneck Lagrange multiplier, which has a value greater than 0 and is used to adjust the compression intensity. This represents the Kullback-Leibler divergence calculation function. This indicates that, given the underlying latent features Latent variables inferred under certain conditions The posterior probability distribution is taken as a diagonal Gaussian distribution determined by the mean vector and standard deviation vector output by the variational encoder. This represents a standard Gaussian prior distribution with a mean of zero and a covariance matrix equal to the identity matrix. Let represent the identity matrix consistent with the dimension of the latent variables. The first term in the formula is the InfoNCE contrastive loss, whose numerator is the similarity index term of positive sample pairs, and the denominator is the sum of the similarity index terms of positive samples and all negative samples. Therefore, the value of this fraction is between 0 and 1, and its logarithm is non-negative. The value of this term decreases as the similarity of positive samples increases. Therefore, optimizing this term is equivalent to maximizing the mutual information between the surface skill representation and its corresponding meta-capability representation, so that Spring Cloud service governance and the distributed system thinking it supports are closer to each other in the representation space. The second term is the information bottleneck penalty term, which constrains the posterior distribution of the latent variables to converge to the standard Gaussian prior, which is equivalent to limiting the upper limit of the amount of information that the latent variables can carry. The compression intensity is controlled by the Lagrange multiplier.

[0037] Under the combined effect of the two constraints, the network must prioritize retaining components that can be reproduced across technology stacks within a limited information capacity. Therefore, features such as framework interface names and configuration syntax that only appear in a single technology stack cannot be suppressed by the bottleneck constraint, while service decomposition and fault-tolerant processing cognition that repeatedly appear in multiple technology stacks are retained. Backpropagation and gradient descent are performed on the contrastive information network. After iterating until the loss converges, the mean vector and standard deviation vector output by the variational encoder are extracted, assembled into decoupled latent variable distribution parameters according to the node index, and output. At this point, the surface business feature matrix and the underlying latent feature matrix have been used up in this stage and will no longer participate in subsequent processing.

[0038] In the aforementioned knowledge graph-based talent gap diagnosis and recommendation system 100, the reparameterization sampling unit 133 is used to perform meta-capability feature reparameterization sampling on auxiliary noise vectors randomly drawn from a standard normal distribution based on the decoupled latent variable distribution parameters to obtain cross-domain meta-capability vectors. It should be noted that since the decoupled latent variable distribution parameters are probabilistic descriptions composed of a mean vector and a standard deviation vector, they are not deterministic values ​​that can be directly used in distance metrics, and direct sampling from the posterior distribution would cause the sampling nodes to become non-differentiable, interrupting gradient backpropagation. Therefore, the technical solution of this application further performs meta-capability feature reparameterization sampling on auxiliary noise vectors randomly drawn from a standard normal distribution based on the decoupled latent variable distribution parameters. Through the above processing, the probabilistic latent variables can be effectively converted into differentiable and subsequently mappable deterministic meta-capability representations.

[0039] More specifically, in a concrete example of this application, the latent variable distribution parameters are analyzed and decoupled. The expected mean vector and standard deviation vector corresponding to each meta-capability dimension are separated by node index. The mean vector represents the employee's central position in dimensions such as distributed system thinking, architectural abstraction ability, and understanding of concurrency control. The standard deviation vector represents the uncertainty corresponding to the sufficiency of evidence for that dimension. Meta-capability dimensions that have been cross-validated by multiple technology stacks over a long period have smaller standard deviations, while dimensions supported only by a single project have larger standard deviations. Subsequently, an auxiliary noise vector consistent with the latent variable dimension is independently and randomly drawn from the standard normal distribution. This noise vector is multiplied by the standard deviation vector item by item at corresponding positions. The product is then added to the mean vector to obtain a deterministic numerical representation, mathematically expressed as:

[0040] in, This indicates the output after reparameterized sampling. Cross-domain meta-capability vectors of each node Represents the first term extracted from the distribution parameters of the decoupled latent variables. The posterior distribution mean vector of each node. This represents an auxiliary noise vector that is independently and randomly drawn from a standard normal distribution. This represents the Hadama product, which is the product of terms multiplied in corresponding positions. Represents the first term extracted from the distribution parameters of the decoupled latent variables. The posterior distribution standard deviation vector of each node. In the formula, the product of the noise vector and the standard deviation vector is the random disturbance introduced by the uncertainty of this dimension. After superimposing it on the mean vector, the randomness of the sampling result is completely transferred to the noise vector that is independent of the network parameters. The partial derivatives of the mean vector and the standard deviation vector with respect to the sampling result can be analytically obtained. Therefore, the gradient can be backpropagated to the aforementioned comparison information network through the sampling nodes.

[0041] After completing node-by-node sampling, the deterministic numerical representations of all nodes are integrated into a high-dimensional dense matrix by index and output as a cross-domain meta-capability vector. Each dimension of this vector corresponds to cognitive capabilities independent of the business domain, and no longer carries traces of specific framework interfaces and configuration syntax. Therefore, the distributed system thinking supported by Spring Cloud service governance and Kubernetes orchestration converges to similar values ​​in this vector, allowing the same employee to reuse the same capability representation to participate in subsequent spatial mapping and gap quantification when moving to target positions in different technology stacks.

[0042] In the aforementioned knowledge graph-based talent gap diagnosis and recommendation system 100, the gap quantification module 140 is used to perform domain-adaptive space mapping and dynamic quantification of capability gaps on cross-domain meta-capability vectors based on the job capability subgraph to obtain a job capability gap matrix. It should be noted that, given that the cross-domain meta-capability vectors reside in the decoupled latent variable space, while the capability requirements in the job capability subgraph are expressed in a topological form from the job center node to the skill leaf node and reside in another feature space, their dimensions and distributions are inconsistent, making direct distance comparison impossible to determine the gap between employees and target positions. Therefore, the technical solution of this application further performs domain-adaptive space mapping and dynamic quantification of capability gaps on cross-domain meta-capability vectors based on the job capability subgraph. First, a nonlinear transformation aligns the meta-capability representations to the geometric space where the target position is located. Then, a domain-adaptive distance metric is used to convert the distribution differences between the two into the degree of deficiency in each capability dimension. Through the above processing, the gap between employee capabilities and job requirements can be effectively transformed from incomparable differences in representation into a readable job capability gap matrix that is readable in each dimension. This prevents job transition scenarios with different technical stacks but similar underlying cognitive requirements from being judged as completely mismatched, and provides clear targets for subsequent learning path searches to fill the gaps.

[0043] Figure 4 This is a block diagram of the gap quantification module in a knowledge graph-based talent capability gap diagnosis and recommendation system according to an embodiment of this application. Figure 4As shown, the gap quantification module 140 includes: a job feature encoding unit 141, used to encode the topological features of the job capability subgraph to obtain the target job feature representation; a spatial mapping unit 142, used to perform cross-domain feature nonlinear spatial mapping on the cross-domain meta-capability vector to obtain the mapped capability representation vector; and a distance metric unit 143, used to perform domain-adaptive distance metric between the mapped capability representation vector and the target job feature representation to obtain the job capability gap matrix.

[0044] In the aforementioned knowledge graph-based talent gap diagnosis and recommendation system 100, the job feature encoding unit 141 is used to encode the target job topological features of the job competency subgraph to obtain a target job feature representation. It should be noted that, because the job competency subgraph stores capacity requirements in the form of discrete center nodes, skill leaf nodes, and their edges, it does not form a continuous representation that can participate in numerical measurement. Based on this, the technical solution of this application further encodes the target job topological features of the job competency subgraph. Through the above processing, the ideal competency standard of the job can be effectively converted into a continuous tensor that can be compared with the employee competency representation.

[0045] More specifically, in a concrete example of this application, the graph structure and initial node features of the job competency subgraph are extracted. The initial node features include semantic vectors of competency items extracted from the target job competency description text, as well as the mastery requirements attached to the edges. This graph structure and features are then input into a graph isomorphic network encoder. During encoding, for any node, its own features and those of its first-order neighbors are summed and aggregated using learnable self-loop weights, ensuring that the central node's features retain an independent contribution share in the aggregation result. The aggregation result is then fed into a multilayer perceptron for nonlinear transformation and batch normalization and activation, iterating sequentially according to a preset number of layers. Compared to aggregation methods that only perform neighborhood averaging, this method has the ability to distinguish between differences in the number of competency requirement nodes and connection structures, resulting in different encoding results for jobs requiring both container orchestration and service mesh governance capabilities compared to jobs requiring only one of these capabilities. Therefore, the differences in competency standards between jobs are reflected in the representation.

[0046] After multiple rounds of iterative aggregation, the node representations output from each layer are read out dimensionally. Discrete skill leaf node features are aggregated to the job center node and integrated into a continuous dense tensor. Each dimension of this tensor corresponds to the required competency dimensions for the job, with values ​​reflecting the expected level of mastery in that dimension. Dimensions described as requiring proficiency in distributed system design have higher values, while dimensions described as requiring only a basic understanding have lower values. This tensor is output as the target job feature representation, used for subsequent domain-adaptive distance measurement within the same space as the mapped employee competency representation. Simultaneously, the cross-domain meta-competency vector and the learning resource subgraph remain in their original states for later use.

[0047] In the aforementioned knowledge graph-based talent gap diagnosis and recommendation system 100, the spatial mapping unit 142 is used to perform cross-domain feature nonlinear spatial mapping on the cross-domain meta-ability vector to obtain the mapped capability representation vector. It should be noted that since the cross-domain meta-ability vector is sampled by a variational encoder and located in the latent variable space, its dimension setting and value distribution are inconsistent with the target job feature representation output by the graph isomorphic network, and the two lack a spatial basis for direct distance measurement. Based on this, the technical solution of this application further performs cross-domain feature nonlinear spatial mapping on the cross-domain meta-ability vector. Through the above processing, employee capability representation and job capability standards can be effectively aligned to fall into the same geometric space and achieve manifold alignment.

[0048] More specifically, in a concrete example of this application, a deep mapping network containing two fully connected layers with bias terms is constructed. The output dimension of the first layer is set according to the transition relationship between the latent variable dimension and the job feature dimension, and the output dimension of the second layer is strictly consistent with the dimension of the target job feature representation. The cross-domain meta-capability vector is input node by node into this network. First, a linear transformation is performed on the weight matrix of the first layer, and a bias term is superimposed. Then, a modified linear unit activation is applied to introduce nonlinearity. The activation result is then fed into the second layer to perform the same linear transformation and activation, resulting in an output with the same dimension as the target job feature representation, mathematically expressed as:

[0049] in, Indicates the first The mapped capability representation vectors corresponding to each node. This indicates a modified linear unit activation function. This represents the weight transformation matrix of the first fully connected layer of the deep mapping network. The first sample obtained from the aforementioned sampling Cross-domain meta-capability vectors This represents the bias term vector of the first fully connected layer. This represents the weight transformation matrix of the second fully connected layer. This represents the bias term vector of the second fully connected layer. This represents the matrix-vector multiplication operation. The activation result of the inner layer is the representation of the meta-capability characterization in the intermediate transition space. The outer transformation completes the projection onto the job feature space based on this. The superposition of the two nonlinear transformations allows the nonlinear correspondence between the latent variable space and the job feature space to be fitted, avoiding the situation where a single linear transformation can only express scaling and rotation but cannot characterize the interaction between capability dimensions.

[0050] After completing the node-by-node mapping, all outputs are concatenated along the dimensional direction according to the meta-capability node index and aggregated through attention-weighted pooling to form an employee-level mapped capability representation vector. Its total dimension strictly matches the target job feature representation, and each dimensional block corresponds one-to-one with the skill leaf nodes in the job capability subgraph. Through this mapping, the employee's values ​​in dimensions such as distributed system thinking and architectural abstraction capabilities are projected to coordinate positions corresponding to the service decomposition design, container orchestration, and other capability dimensions required by the job. Therefore, the cognitive capabilities accumulated from the traditional technology stack can be compared dimensionally with the capability standards of the new technology stack job in the same space. Simultaneously, the target job feature representation and the learning resource subgraph maintain their state for subsequent distance measurement and path search invocation.

[0051] Specifically, the weight transformation matrix and bias terms of the aforementioned deep mapping network are determined during the training phase. Training samples are taken from historical successful job transfer records. Specifically, the cross-domain meta-ability vector of employees who have passed the probationary period assessment before the transfer is used as input, and the target job feature representation of the transferred position is used as the regression target. The mean square error between the employee-level mapped ability representation vector obtained by aggregation in the above manner and the target job feature representation is used as the alignment loss, ensuring that employees whose abilities have been confirmed to meet the standards fall into the neighborhood of the target job feature representation after mapping. Simultaneously, the maximum mean difference is used to measure the distance between the distribution of the employee-level mapped ability representation vector and the distribution of the target job feature representation, and this distance is used as the domain adaptation loss, ensuring that the marginal distributions of the two spaces are aligned rather than just fitted on a single sample. The total loss is the weighted sum of the alignment loss and the domain adaptation loss, with weights of 1.0 and 0.3, respectively. The Adam optimizer is used, with an initial learning rate of 1×10⁻⁶. -3 The alignment loss is decayed using cosine annealing with a batch size of 64. Every 20 rounds, the alignment loss is evaluated on the validation set, and training terminates if it fails to decrease for 5 consecutive rounds. After training, this network, along with the aforementioned contrastive information network and variational encoder, undergoes joint fine-tuning. The total loss during fine-tuning is the weighted sum of the cross-layer information bottleneck loss and the aforementioned total loss. After fine-tuning, all parameters are frozen, and only forward mapping is performed during the inference phase without updating the weights. For target positions with insufficient historical transfer records, transfer records from other positions within the same job family are used as transfer samples for training, and their alignment loss weights are reduced to 0.5.

[0052] In the aforementioned knowledge graph-based talent gap diagnosis and recommendation system 100, the distance metric unit 143 is used to perform a domain-adaptive distance metric on the mapped capability representation vector and the target job feature representation to obtain a job capability gap matrix. It should be noted that although the mapped capability representation vector and the target job feature representation fall into the same space, the geometric difference between them remains in vector form and is not converted into a gap metric that can be directly used for path search. Based on this, the technical solution of this application further performs a domain-adaptive distance metric on the mapped capability representation vector and the target job feature representation. Through the above processing, the spatial distribution difference can be effectively converted into a dimensional capability deficiency degree and output as a structured matrix.

[0053] More specifically, in a concrete example of this application, the corresponding components of the mapped capability representation vector and the target job feature representation are extracted dimension by dimension within the manifold-aligned space. First, the L2 norm square of the difference between the two is calculated to measure the absolute geometric difference between the employee's actual capability and the job's expected standard in that dimension. Then, a preset distance scaling factor is introduced to normalize the squared result. Subsequently, the result is negative and mapped to a value range greater than 0 and less than or equal to 1 using the natural exponential function to obtain the capability matching degree for that dimension. Finally, the matching degree is subtracted from 1 and processed using a non-negative truncation function to obtain a scalar of the capability deficiency degree in that dimension, whose mathematical expression is: (Modified as follows:)

[0054] in, Indicates the first A scalar measure of the degree of capability deficiency in each capability dimension. This represents the capability dimension index corresponding to the skill leaf node in the job capability subgraph, with a value ranging from 1 to... , This represents the total number of skill leaf nodes in the job competency subgraph, which is also the total number of competency dimensions. This represents a non-negative truncation function that takes the larger of the two terms within the parentheses. This represents an exponential function with the natural constant as its base. The mapped capability representation vector and the target job feature representation are shown in the first... The squared L2 norm of the differences between the subvectors corresponding to each capability dimension. The mapped capability representation vector at the employee level is related to the first... Sub-vectors corresponding to each capability dimension The target job characteristics are represented in the first... Each capability dimension corresponds to a sub-vector, and two sub-vectors have the same dimension, with their block positions determined by the node order of the job capability subgraph. This represents the distance scaling factor, with a value greater than 0. The exponential term in the formula represents the capability matching degree in this dimension, with a value greater than 0 and less than or equal to 1. It monotonically decreases with increasing geometric distance, approaching 0, and the matching degree is 1 when the distance is 0. The difference between 1 and this matching degree represents the degree of missing data. Non-negative truncation ensures that if this difference is negative due to numerical calculation errors, the gap in this dimension is 0, avoiding physically meaningless negative values. The distance scaling factor is used to adjust the sensitivity of distance differences to the matching degree, preventing the absolute value of distance in high-dimensional space from being too large, causing the exponential term to rapidly approach zero and making gaps in different dimensions indistinguishable.

[0055] After completing the quantification along each dimension, all capability deficiency scalars are arranged into rows according to the node order of the job capability subgraph. The columns are then filled with the expected level of mastery for that dimension, the employee's current mapping value, the deficiency scalar, and the completion priority determined by the weighted terms in the job description. This results in a job capability gap matrix with four columns and the number of rows equal to the total number of capability dimensions. This quantification reveals that employees transitioning from traditional to new technology stacks have gap values ​​close to 0 in dimensions they already possess, such as distributed systems thinking. However, they have higher gap values ​​in dimensions lacking evidence, such as container orchestration practice and service mesh governance. Therefore, the gap distribution aligns with actual training needs. The mapped capability representation vectors and target job feature representations are no longer used in subsequent processing after this stage. The learning resource subgraph retains its state and serves as the action space source for subsequent path search.

[0056] In the aforementioned knowledge graph-based talent competency gap diagnosis and recommendation system 100, the path generation module 150 is used to perform constrained path search and branch pruning on the job competency gap matrix and learning resource subgraph to output a cross-domain learning path. It should be noted that the job competency gap matrix only indicates the degree of employee deficiencies in dimensions such as container orchestration practice and service mesh governance, without providing specific resources and execution order to fill these deficiencies. Furthermore, there are prerequisite topological constraints between courses and practical projects in the learning resource subgraph, and the resource sequences obtained by relevance retrieval are easily influenced by confounding variables such as employee basic competencies and historical course selection preferences, resulting in pseudo-associations. Therefore, the technical solution of this application further performs constrained path search and branch pruning on the job competency gap matrix and learning resource subgraph. Using the gap matrix as the initial state of the search, candidate branches are expanded within the legal range limited by prerequisite edges, and branches with insufficient causal confidence are eliminated based on learning effect edges. Through the above processing, the dimensional gap diagnosis results can be effectively converted into cross-domain learning paths that satisfy prerequisite relationships and have optimal global benefits, while reducing the search overhead caused by invalid branches participating in subsequent traversals.

[0057] Figure 5 This is a block diagram of the path generation module in a knowledge graph-based talent gap diagnosis and recommendation system according to an embodiment of this application. Figure 5 As shown, the path generation module 150 includes: a search space construction unit 151, used to encapsulate the job competency gap matrix into an initial state tensor to generate a reinforcement learning state space, and to perform legality masking on the prerequisite edges of the learning resource subgraph to generate a restricted action search space; a causal pruning unit 152, used to expand the initial candidate branches using a restricted breadth-first search, and to perform causal inference pruning in the restricted action search space to obtain a pruned candidate path set; and a path convergence unit 153, used to perform global reward calculation and optimal path convergence on the pruned candidate path set based on the reinforcement learning state space to obtain a cross-domain learning path.

[0058] In the aforementioned knowledge graph-based talent competency gap diagnosis and recommendation system 100, the search space construction unit 151 is used to encapsulate the job competency gap matrix into an initial state tensor to generate a reinforcement learning state space, and to perform legality masking on the prerequisite connections of the learning resource subgraph to generate a restricted action search space. It should be noted that since the job competency gap matrix and the learning resource subgraph are respectively the quantitative diagnosis result and the resource topology structure, neither possesses the interaction form required for sequential decision-making, and the prerequisite dependencies between resources, if not pre-constrained, will lead to action sequences that violate teaching logic during the search process. Based on this, the technical solution of this application further encapsulates the job competency gap matrix into an initial state tensor to generate a reinforcement learning state space, and performs legality masking on the prerequisite connections of the learning resource subgraph to generate a restricted action search space. Through the above processing, an interactive state benchmark and a legal action range can be effectively established for subsequent branch expansion and pruning.

[0059] More specifically, in a concrete example of this application, a job competency gap matrix is ​​received, and the competency deficiency scalars of each dimension are arranged in the node order of the job competency subgraph to form the initial state tensor of a Markov decision process. The values ​​of dimensions such as container orchestration practice and service mesh governance in the tensor represent the gaps to be filled, while dimensions with values ​​close to 0, such as those related to distributed systems thinking, are marked as satisfied and not used as target dimensions for subsequent state transitions. Simultaneously, two auxiliary components, remaining available learning hours and records of executed actions, are added to this tensor, allowing the state to be updated synchronously with the gap reduction and learning hour consumption after the actions are executed, thereby generating a reinforcement learning state space.

[0060] After constructing the state space, the course nodes and practice project nodes in the learning resource subgraph are parsed and aggregated into an available action set. This action set includes entries such as the Container Orchestration Principles course, the Service Mesh Configuration Practice course, and corresponding deliverable practice projects. Then, the prerequisite edges in the subgraph are read one by one to determine the prerequisite resource identifiers for each action. A topological sort is performed on the available action set, and a validity mask is applied to the action set according to the currently executed action records. Actions whose prerequisite resources are not yet completed are set as unselectable at their corresponding positions in the mask vector, while actions whose prerequisite resources are completed but not yet executed are set as selectable. The mask vector is updated with each action execution step. Through this process, in the state where the Container Orchestration Principles course is not completed, the Service Mesh Configuration Practice course is masked and does not enter the candidate range, thus avoiding the generation of action sequences that bypass the basic steps. The action set that completes the mask constraints is encapsulated into a restricted action search space, which, together with the aforementioned reinforcement learning state space, serves as the input for subsequent branch expansion and causal pruning.

[0061] In the aforementioned knowledge graph-based talent gap diagnosis and recommendation system 100, the causal pruning unit 152 is used to expand the initial candidate branches using a restricted breadth-first search and perform causal inference pruning in the restricted action search space to obtain a pruned candidate path set. It should be noted that, given that although the restricted action search space has excluded actions that violate prior relationships, the number of combinations of legal actions increases rapidly with the increase in gap dimensions, and the correlation between ability improvement and learning actions presented in historical data is affected by confounding variables such as employee basic abilities and historical course selection preferences, retaining branches solely based on relevance will result in resources with limited actual benefits entering candidate paths. Based on this, the technical solution of this application further utilizes a restricted breadth-first search to expand the initial candidate branches and performs causal inference pruning in the restricted action search space. Under the premise of satisfying prior connection constraints, it expands the branches containing sequential actions layer by layer, and estimates the actual processing effect of each learning action on ability gap reduction based on historical verification data to eliminate low-confidence branches. Through the above processing, the resources in the candidate path set and gap filling can be effectively supported at the causal level, while reducing the scale of branches participating in subsequent global reward calculations.

[0062] More specifically, in a concrete example of this application, the initial gap tensor in the reinforcement learning state space is used as the starting point for the search. A restricted breadth-first expansion is performed within the restricted action search space. Each layer takes values ​​only within the range of actions indicated by the mask vector as optional, and the selected actions are required to cover the gap dimensions that are not yet satisfied and not to form a loop with the actions already entered into the branch. Based on this, initial candidate branches containing ordered action sequences are generated layer by layer outward. One branch is represented by executing the container orchestration principle course first and then the service mesh configuration practice course, and the other branch is represented by executing the distributed computing principle course first and then the microservice architecture design course. The learning effect edges of these two courses point to the traffic governance sub-item and the service decomposition design sub-item under the service mesh governance dimension, respectively, and thus also fall within the range of the gap dimensions that are not yet satisfied. For each extended learning action, causal inference pruning is triggered. Employee samples that performed the action and those that did not are extracted from the historical validation dataset. A propensity score is calculated based on the employee's initial ability, past learning records, historical job position, and learning preferences, and its reciprocal is used as the inverse probability weight. After confounding correction for the reduction in the ability gap between the two classes, the expected difference is calculated to obtain the average treatment effect of the action on the reduction of the ability gap. Its mathematical expression is:

[0063] in, Indicates learning action The average treatment effect on reducing capacity gaps This represents the total number of valid samples in the historical validation dataset. This indicates an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. Indicates the first The actual intervention actions that occurred in a historical sample This indicates the target action that is currently being evaluated. Indicates the first The actual reduction in the capability gap after action intervention for each historical sample. In a given confounding matrix of variables The sample selection action under the condition Tendency to divide This represents a confounding variable matrix consisting of employees' initial abilities, past learning records, historical positions, and learning preferences. In the formula, the minuend is the expected ability improvement of samples who performed the action after weight correction, and the subtrahend is the expected ability improvement of samples who did not perform the action after the same correction. The difference between the two is the net effect of the action after removing confounding influences, thus correcting the apparent high gains caused by high-performing employees concentrating on this activity. The average treatment effect of each action is compared with a preset fixed causal confidence threshold. Branches containing actions below this threshold are judged as spurious associations and removed. The remaining branches are written into the pruned candidate path set, and the reinforcement learning state space retains its content and is passed on to subsequent global reward calculations.

[0064] The above treatment can weaken the spurious associations caused by confounding variables such as employees' basic skills and historical course selection preferences. However, its effect estimation granularity remains at the level of a single learning action, assuming that different learning resources are independent of each other, and failing to consider the directional sequential causal interactions between adjacent learning actions. The pre-refinement edges in the learning resource subgraph only indicate whether courses or practical projects can be continued in the topological structure, solving the problem of path legality, but cannot indicate whether the previous learning action enhances or weakens the actual ability improvement effect of the subsequent learning action. In scenarios where learning distributed computing principles is followed immediately by microservice architecture design, the preceding action provides cognitive scaffolding for the following action, and the combined capability gain of the two is higher than the sum of their individual gains. If the average processing effect of a single action in microservice architecture design is lower than a fixed threshold, that branch will still be incorrectly pruned, resulting in the loss of effective paths with synergistic amplification effects. Conversely, in scenarios where only basic linear algebra is completed before skipping machine learning fundamentals and directly entering advanced deep learning, the average processing effects of both actions are high, but the cognitive span between adjacent actions is large, and the combined effect is lower than the sum of their individual effects. Retaining branches solely based on single-action scores cannot identify antagonistic inhibition relationships caused by cognitive overload, knowledge duplication, or connection span. It is evident that single-action evaluation can only answer whether a learning resource is effective when implemented alone, but cannot answer whether the resource remains effective when it is located after a specific preceding action. This leads to the incorrect pruning of low-scoring, high-efficiency sequences and the incorrect retention of high-scoring, low-efficiency sequences, with ineffective branches continuing to participate in subsequent state transitions and global reward calculations. Furthermore, the fixed threshold cannot adapt to the gap span of different employees and the resource density of different positions.

[0065] Therefore, in another specific example of this application, the above-mentioned single action effect pruning is improved into adjacent action sequence causal interaction perception pruning. While retaining the previous edge trimming to determine the legality of the path, the action pair causal interaction coefficient matrix is ​​further introduced to determine whether the legal sequence has actual gain, and the branch selection is completed by replacing the fixed threshold with an adaptive threshold.

[0066] Figure 6This is an interactive modulation pruning diagram of the causal pruning unit in a knowledge graph-based talent gap diagnosis and recommendation system according to an embodiment of this application. Figure 6 As shown, the causal pruning unit 152 includes: an interaction coefficient quantization subunit 1521, used to quantize the sequence causal interaction coefficients of all adjacent action pairs satisfying the prior topological constraints in the restricted action search space to obtain the action pair causal interaction coefficient matrix; a branch expansion subunit 1522, used to expand the candidate branches containing ordered action sequences using restricted breadth-first search under the premise of satisfying the prior edge constraints, and query the action pair causal interaction coefficient matrix in real time to obtain the interaction coefficients of adjacent action pairs; an interaction modulation scoring subunit 1523, used to comprehensively score the marginal effect of each step of each candidate branch plus its interaction gain term with the preceding action based on the interaction coefficients of adjacent action pairs to obtain the interaction modulation candidate branch set; and an adaptive pruning subunit 1524, used to perform adaptive threshold pruning based on the interaction modulation score on the interaction modulation candidate branch set to obtain the pruned candidate path set.

[0067] In the aforementioned knowledge graph-based talent gap diagnosis and recommendation system 100, the interaction coefficient quantization subunit 1521 is used to quantify the sequential causal interaction coefficients of all adjacent action pairs satisfying the pre-condition topological constraints in the restricted action search space to obtain the action pair causal interaction coefficient matrix. It should be noted that, as mentioned earlier, single-action granularity effect estimation assumes that learning resources are independent of each other, and pre-conditioning edges only represent path legality, failing to reflect the directional causal relationship between adjacent actions. Based on this, the technical solution of this application further quantifies the sequential causal interaction coefficients of all adjacent action pairs satisfying the pre-condition topological constraints in the restricted action search space, identifying from the actual learning results whether adjacent actions are collaborative, independent, or mutually inhibiting, to obtain the action pair causal interaction coefficient matrix. Through the above processing, the directional causal relationship between adjacent actions can be effectively converted into a quantifiable basis that can be read by index, providing criteria beyond the single-action granularity for subsequent branch scoring and pruning.

[0068] More specifically, in a specific example of this application, all legal adjacent action pairs are enumerated according to the topological order in the restricted action search space. Action pairs are limited to two learning resources executed sequentially within a preset connection time window. The connection time window is set according to the training cycle. Combinations where the execution interval of two resources exceeds this window are not included in the action pair, thus excluding situations where the preceding and following learning behaviors are out of sync and do not constitute cognitive connection. For each legal action pair, historical samples are extracted from the historical validation dataset, including those executed only as the preceding action, those executed only as the following action, and those executed sequentially as the preceding and following actions. Inverse probability weights are calculated based on the employee's initial ability, past learning records, historical job position, and learning preferences. The ability gap reduction of the three types of samples is then corrected for confounding factors, resulting in the mean independent gain of the preceding action, the mean independent gain of the following action, and the mean joint gain of the action pair. Subsequently, the relative deviation of the joint gain exceeding the sum of the two independent gains represents the sequence causal interaction coefficient, mathematically expressed as:

[0069] in, Indicates front-drive With subsequent actions Sequence causal interaction coefficients when executed in a specified order This represents the predecessor learned action in a legal adjacent action pair. This represents the subsequent learned action in a legal adjacent action pair. This indicates that the operations will be performed sequentially after inverse probability weighting correction. and The average reduction of the joint capability gap This indicates that only the inverse probability weighted correction is performed. The capacity gap is narrowing to the mean. This indicates that only the inverse probability weighted correction is performed. The capacity gap is narrowing to the mean. This represents the directed execution order from the preceding action to the subsequent action. The numerator is the excess of the joint gain relative to the sum of the two independent gains, characterizing the additional capability enhancement generated or lost when adjacent actions are executed in this order. The denominator is the sum of the two independent gains, serving as a normalization benchmark to eliminate differences in gain magnitude between different action pairs, ensuring that the interaction strength of each action pair is on the same comparative scale. An interaction coefficient greater than zero indicates a synergistic amplification effect of the preceding action on the subsequent action; an interaction coefficient less than zero indicates an antagonistic inhibitory effect between adjacent actions; and an interaction coefficient close to zero indicates that the two learning resources are essentially independent.

[0070] Matrix assembly is performed after the coefficients of all legal action pairs have been calculated. It should be noted that... and Since the values ​​do not need to be equal, the action-to-causal interaction coefficient matrix uses a directed matrix structure to store the sequence relationship. The row index of the matrix corresponds to the preceding action, and the column index corresponds to the succeeding action. The average processing effect obtained by inverse probability weighting of each single action is written into the diagonal elements of the matrix, and the interaction coefficient of each adjacent action pair is written into the corresponding off-diagonal elements. This allows subsequent searches to directly read the marginal effects and interaction effects through the row and column indices, avoiding repeated scanning of the historical validation dataset during each branch expansion. For action combinations that do not meet the prior topological constraints or whose execution interval exceeds the connection time window, their corresponding positions are set to null values ​​and treated as having zero interaction coefficients in subsequent queries. That is, only the marginal effect of a single action is included, to ensure that the matrix structure is consistent with the legal range of the restricted action search space.

[0071] In a specific scenario, for employees transitioning from traditional backend development to cloud-native platform roles, their skill gaps are concentrated in two dimensions: container orchestration practice and service mesh governance. The restricted action search space includes courses such as container orchestration principles, service mesh configuration practice, distributed computing principles, and microservice architecture design. Quantification shows that the action pair from distributed computing principles to microservice architecture design has a joint mean gain higher than the sum of the two independent mean gains, and the interaction coefficient is positive, indicating that the former action provides the cognitive foundation for the latter. In another scenario involving an employee transitioning from traditional backend development to machine learning platform roles, their restricted action search space includes courses such as basic linear algebra, machine learning fundamentals, and advanced deep learning. For the action pair from basic linear algebra to advanced deep learning, although the independent mean gains of each resource are high, the joint mean gain of sequentially executed actions is lower than their sum, and the interaction coefficient is negative, indicating that the cognitive leap between adjacent actions inhibits their progress. Meanwhile, the reverse action pair from service mesh configuration practice to container orchestration principles is excluded from the enumeration range due to violating pre-condition topology constraints. The above results, along with the normalized average processing effect of each action, are written into a directed matrix to obtain the action-pair causal interaction coefficient matrix that can be used for subsequent branch expansion and index-based querying.

[0072] In the aforementioned knowledge graph-based talent gap diagnosis and recommendation system 100, the branch expansion subunit 1522 and the interaction modulation scoring subunit 1523 are used to expand candidate branches containing ordered action sequences using a restricted breadth-first search under the premise of satisfying the pre-conditioning edge constraint, and to query the causal interaction coefficient matrix of action pairs in real time to obtain the interaction coefficients of adjacent action pairs; subsequently, based on the interaction coefficients of adjacent action pairs, a comprehensive score is performed on the marginal effect of each action in each candidate branch plus its interaction gain term with the preceding action to obtain the interaction modulation candidate branch set. It should be noted that, as mentioned above, scoring based solely on the average processing effect of a single action cannot reflect the actual effect of a resource when it is located after a specific preceding action. Based on this, the technical solution of this application further utilizes a restricted breadth-first search to expand candidate branches containing ordered action sequences under the premise of satisfying the pre-conditioning edge constraint, and queries the causal interaction coefficient matrix of action pairs in real time to obtain the interaction coefficients of adjacent action pairs, and performs a comprehensive score on the marginal effect of each action in each candidate branch plus its interaction gain term with the preceding action. Through the above processing, the branch score can effectively reflect the benefit of the action itself and the synergistic or antagonistic relationship between adjacent actions, thereby outputting an interactive modulation candidate branch set whose scoring criteria exceed the granularity of a single action.

[0073] More specifically, in a concrete example of this application, starting with the current capability gap tensor in the reinforcement learning state space, a restricted breadth-first expansion is performed within the restricted action search space. Only actions that satisfy the requirements of learning resource edge trimming, no loop formation, and coverage of the current gap dimension are allowed to enter the candidate branch. The coverage determination is based on whether there is an intersection between the capability dimension pointed to by the learning effect edge corresponding to the action and the gap dimension with a value greater than zero in the state tensor. Actions that do not reduce the gap are not included in the expansion range. Loop detection is based on the set of action identifiers already in the branch for repeatability verification. Based on this, candidate branches containing ordered action sequences are generated layer by layer. For each subsequent action expanded, the causal interaction coefficient matrix of the action pair is queried based on the node index of the preceding and subsequent actions. The average treatment effect of the corresponding single action and the causal interaction coefficient of the sequence are read. The diagonal elements provide the marginal benefit of the subsequent action, and the off-diagonal elements provide the interaction strength when the action is located after the current preceding action. The query is completed by direct addressing by row and column index, without the need to backtrack the historical verification dataset during the expansion process. The marginal benefit of a single action is then combined with the interaction correction term of adjacent actions to form an effective causal score for interaction enhancement, the mathematical expression of which is:

[0074] in, Indicates candidate branches Enhanced interaction and effective causal scoring This represents a candidate branch consisting of multiple learning actions in the order of execution. This indicates the number of learning actions contained in the candidate branch. This represents the sequential index of the learning action within the candidate branches. Indicates the first Each learning action, Indicates and Adjacent subsequent learning actions, Indicates learning action The average treatment effect obtained by inverse probability weighting, Indicates adjacent learning actions and The sequence causal interaction coefficient, This indicates taking the smaller of the average treatment effects of two adjacent single actions. The first term in the formula is the sum of the marginal effects of each action within the branch, representing the basic gain of the branch when the relationship between actions is ignored. The second term is the sum of the interaction corrections for all adjacent actions, which is obtained by multiplying the interaction coefficient by the smaller of the average treatment effects of two adjacent actions. The interaction coefficient determines the sign and relative magnitude of the correction, while the smaller value determines its absolute magnitude.

[0075] For example, for employees transitioning from traditional backend development roles to cloud-native platform roles, the search starts with addressing gaps in two key areas: container orchestration practice and service mesh governance. For candidate branches moving from distributed computing principles to microservice architecture design, the average processing effect of subsequent actions is low, but the interaction coefficient obtained from the query matrix is ​​positive. The interaction correction term provides positive compensation based on the lower of the two actions' average processing effects, thus the branch's interaction enhancement effective causal score is higher than the value obtained by accumulating individual actions, allowing it to be retained in subsequent screening. For candidate branches moving directly from basic linear algebra to advanced deep learning, both actions have high average processing effects, but the interaction coefficient is negative. The interaction correction term becomes a deduction, and the branch's score is lowered to below the sum of its marginal effects. For combinations moving from service mesh configuration practice to container orchestration principles, they are excluded during the expansion stage due to violating the pre-conditioning edge constraint and do not enter the scoring process. Encapsulating the above candidate branches and their action sequences with their corresponding scores yields the interaction modulation candidate branch set.

[0076] In the aforementioned knowledge graph-based talent gap diagnosis and recommendation system 100, the adaptive pruning subunit 1524 is used to perform adaptive threshold pruning based on interactive modulation scores on the interactive modulation candidate branch set to obtain a pruned candidate path set. It should be noted that, given the varying sizes of talent gaps, available learning resources, and candidate score dispersion among different employees, using a uniform fixed threshold can easily lead to either complete retention or over-pruning. This results in a large number of similar paths being retained in resource-dense positions, while effective paths are concentratedly pruned when the cross-domain span is large. Therefore, the technical solution of this application further performs adaptive threshold pruning based on interactive modulation scores on the interactive modulation candidate branch set. Through this processing, the pruning boundary can be effectively adjusted according to the current score distribution, thereby outputting a smaller pruned candidate path set while preserving path diversity.

[0077] More specifically, in a concrete example of this application, the interaction modulation candidate branch set is traversed and the effective causal scores of interaction enhancement for all branches are extracted to form a score set. The mean and standard deviation of the effective causal scores of interaction enhancement for all branches in the interaction modulation candidate branch set are calculated, and an adaptive pruning threshold is constructed by subtracting one standard deviation from the mean. Its mathematical expression is as follows:

[0078] in, This represents the adaptive pruning threshold corresponding to the current interactive modulation candidate branch set. This represents the arithmetic mean of all effective causal scores for enhanced interaction. The standard deviation of all interaction-enhanced effective causal scores is represented. This represents the set of scores in the current interactive modulation candidate branch set. The mean term determines the center position of the pruning boundary, representing the overall return level of the current candidate branch, while the standard deviation term determines the downward shift of the boundary relative to the center position, representing the dispersion of the score distribution. Subtracting the two terms allows the threshold to be redefined based on the score distribution obtained with each search, without relying on a pre-set fixed value.

[0079] After obtaining the adaptive pruning threshold, branch selection is performed. The effective causal score of interaction enhancement for each candidate branch is compared with the adaptive pruning threshold. Branches below the threshold are identified as inefficient sequences containing strong antagonistic relationships, weak marginal returns, or spurious associations caused by confounding variables and are deleted. Branches above or equal to the threshold are added to the pruned candidate path set. When the overall quality of candidate branches is high and the scores are concentrated, the threshold is increased with the mean, retaining only paths with higher returns. When the cross-domain span is large and the candidate score distribution is discrete, the standard deviation is increased and the threshold is appropriately lowered to avoid the concentrated removal of a few paths with potential synergistic value. The reinforcement learning state space is synchronously transferred to the subsequent global reward calculation, ensuring that the capability gap state before and after path pruning remains consistent. Since pruning is completed before the branch enters the global reward calculation, the deleted branch no longer participates in state tensor copying and reward inference, and the storage space occupied by candidate paths and the number of branch expansions are reduced accordingly.

[0080] In another scenario involving an employee transitioning from a backend development role to a machine learning platform role, the skill gaps are concentrated in model training and tuning, and feature engineering. Available resources are moderate, and the score distribution is relatively concentrated with a small standard deviation. The threshold increases with the mean. Basic linear algebra branches pointing to advanced deep learning, whose scores are lowered due to negative interaction coefficients, are below the threshold and are deleted. Conversely, basic machine learning branches pointing to advanced deep learning, with positive interaction coefficients and after collaborative compensation, are above the threshold and are retained. If another employee in the same role transitions from a data analysis role and the skill gaps span a wider range, the discrete distribution of candidate scores increases the standard deviation, causing the threshold to decrease. A small number of branches with scores in the low to medium range but possessing collaborative value are retained. These selected branches, along with their action sequences, are written into the pruned candidate path set and output along with the unchanged reinforcement learning state space for global reward calculation and optimal path convergence.

[0081] In the aforementioned knowledge graph-based talent gap diagnosis and recommendation system 100, the path convergence unit 153 is used to perform global reward calculation and optimal path convergence on the pruned candidate path set based on the reinforcement learning state space to obtain a cross-domain learning path. It should be noted that since multiple paths satisfying the prior constraints and possessing causal support still exist in the pruned candidate path set, each path makes trade-offs in terms of gap-filling degree and learning time consumption, and no single deliverable execution plan is determined. Based on this, the technical solution of this application further performs global reward calculation and optimal path convergence on the pruned candidate path set based on the reinforcement learning state space. Through the above processing, a balance can be effectively achieved between gap-filling benefits and learning time costs, and a unique cross-domain learning path can be output.

[0082] Specifically, in this embodiment, the path convergence unit includes: a state transition subunit, used to update the gap tensor and deduct the remaining available learning time for each candidate path in the pruned candidate path set according to the action execution order based on the reinforcement learning state space to obtain a time-step state transition sequence; an immediate reward calculation subunit, used to calculate the weighted difference between the gap reduction gain and the learning resource time cost of the time-step state transition sequence to obtain the immediate reward for each time step; a cumulative reward calculation subunit, used to perform a weighted summation of the immediate rewards for each time step based on the reward discount factor to obtain a cumulative reward score; and an optimal path convergence subunit, used to compare the maximum value of the cumulative reward scores of each candidate path to obtain a cross-domain learning path.

[0083] More specifically, in a concrete example of this application, each valid path in the pruned candidate path set is traversed. Starting from the initial gap tensor in the reinforcement learning state space, the state transition is progressively advanced according to the action execution order. For each learning action executed, the gap tensor is updated according to the capability dimension pointed to by the edge corresponding to the learning effect of that action, and the remaining available learning time is simultaneously deducted to obtain a time-step state transition sequence. For each time step, an immediate reward is calculated, where the positive gain is the reduction in the gap tensor before and after the action execution multiplied by a gain coefficient, and the negative penalty is the inherent time cost of the learning resource multiplied by a penalty coefficient. The immediate reward is then multiplied by a discount factor that decreases with each time step and gradually accumulated to obtain the cumulative reward score for that path, mathematically expressed as:

[0084] in, Indicates the first The cumulative reward score of each candidate path, Denotes the first path in the candidate path set after pruning. Candidate paths, This indicates the total number of learning actions included in the candidate path. This represents the index of the discrete time step in the state transition process. This represents the reward discount factor, which takes values ​​greater than 0 and less than or equal to 1. This represents the positive gain weighting coefficient. Indicates the execution of an action Before the time step The gap tensor at time intervals. Indicates the execution of an action Then transfer to time step The gap tensor at time intervals. This represents the negative penalty weighting coefficient. Indicates action The inherent cost is calculated by taking the number of class hours required for the course or the number of days required to deliver the practical project. The difference between the two gap tensors is the actual gap amount filled by the action. Multiplying this by the gain coefficient results in a positive reward, while the product of the inherent cost and the penalty coefficient results in a negative reward. The discount factor causes the contribution of subsequent actions to the total score to gradually decrease. Therefore, when the gap-filling effect is comparable, the path that places the resources with the larger gap reduction at the beginning will get a higher score.

[0085] After calculating the cumulative reward scores for all candidate paths, optimal path convergence is performed. The cumulative reward scores of each path are compared, and the path with the highest value is selected. This path satisfies the following conditions: the action sequence is acyclic, all prerequisite resources for each action are already arranged, and the path is connected. This path is then encapsulated and output as a cross-domain learning path in execution order. Through this convergence, employees transitioning from traditional backend development roles to cloud-native platform roles receive an execution plan that prioritizes container orchestration principles, followed by service mesh configuration practice, culminating in a deliverable practical project. Their gaps in container orchestration practice and service mesh governance are sequentially filled, while their existing distributed systems thinking is no longer required for repetitive learning, thus controlling learning time consumption. At this point, the pruned candidate path set and reinforcement learning state space are used up in this stage and transformed into the final recommendation result.

[0086] In summary, the knowledge graph-based talent capability gap diagnosis and recommendation system according to the embodiments of this application is explained. First, it performs graph association modeling on the received multi-source talent data to obtain a two-layer talent knowledge graph composed of a skill layer and a meta-capability layer, as well as a job capability subgraph and a learning resource subgraph, enabling a vertical association between surface-level business skills and underlying meta-capabilities. Then, it performs node state updates and edge decay weighting on the two-layer talent knowledge graph using a half-life decay function determined by training interaction timestamps and preset capability half-lives, automatically degrading skills that have not been activated for a long time. Subsequently, it uses a comparative information network containing cross-layer information bottlenecks to decouple features and extract meta-capabilities from the time-decaying talent knowledge graph, filtering out domain-specific components to obtain a cross-domain meta-capability vector. Next, based on the job capability subgraph, it performs domain adaptation space mapping and dynamic quantification of capability gaps on this vector, obtaining a dimensionally readable job capability gap matrix within a unified space, avoiding exaggeration of cross-domain gaps due to the isolation of subgraphs. Finally, it performs constrained path search and branch pruning on the gap matrix and the learning resource subgraph, outputting a cross-domain learning path that satisfies the prior relationship and has the optimal global benefit.

Claims

1. A talent skill gap diagnosis and recommendation system based on knowledge graphs, characterized in that, include: The graph construction module is used to perform graph association modeling on the received multi-source talent data to obtain a two-layer talent knowledge graph, a job competency subgraph, and a learning resource subgraph. The time-decay module is used to update the node state and assign weights to the edge decay of the two-layer talent knowledge graph based on the half-life decay function determined by the training interaction timestamp and the preset capability half-life, so as to obtain the time-decayed talent knowledge graph. The meta-capability extraction module is used to decouple cross-layer information bottleneck features and extract meta-capabilities from a time-decaying talent knowledge graph based on a comparative information network containing cross-layer information bottlenecks, so as to obtain a cross-domain meta-capability vector. The gap quantification module is used to perform domain-adaptive space mapping and dynamic quantification of capability gaps on cross-domain meta-capability vectors based on job capability subgraphs to obtain job capability gap matrices. The path generation module is used to perform constrained path search and branch pruning on the job competency gap matrix and learning resource subgraph to output cross-domain learning paths.

2. The knowledge graph-based talent gap diagnosis and recommendation system according to claim 1, characterized in that, The multi-source talent data includes employee project performance scores, code submission sequence text, training interaction timestamps, historical pre- and post-learning ability assessment values, target job ability description text, course attributes, practical project attributes, and resource prerequisite relationships; the two-layer talent knowledge graph consists of a skill layer and a meta-ability layer; the learning resource subgraph includes learning resource prerequisite edges and learning effect edges.

3. The knowledge graph-based talent gap diagnosis and recommendation system according to claim 2, characterized in that, The map construction module includes: The feature tensor unit is used to extract semantic features and normalize numerical values ​​from employee project performance scores, code submission sequence text, training interaction timestamps, historical learning before and after ability assessment values, target job ability description text, course attributes, practical project attributes and resource prerequisite relationships in multi-source talent data to obtain employee behavior feature tensor, job semantic feature tensor and training resource feature tensor. The subgraph topology mapping unit is used to instantiate skill layer nodes and meta-capability layer nodes for the employee behavior feature tensor and establish vertical connections between layers. It also establishes unidirectional connections from the job center node to the skill leaf node for the job semantic feature tensor and injects prerequisite connection edges and learning effect edges into the training resource feature tensor to generate the initial employee two-layer graph, job capability subgraph and learning resource subgraph respectively. The representation aggregation unit is used to perform two-layer graph feature encoding and representation aggregation on the initial employee two-layer graph to obtain a two-layer talent knowledge graph.

4. The knowledge graph-based talent gap diagnosis and recommendation system according to claim 1, characterized in that, The timing decay module includes: The timestamp alignment unit is used to extract the absolute time of the most recent interaction from each skill node in the two-layer talent knowledge graph, convert it into a standardized timestamp, and append the standardized timestamp to the node attribute vector to obtain a two-layer graph with timestamp attributes. The loss coefficient calculation unit is used to calculate the capability loss coefficient based on the half-life function of the two-layer graph with timestamp attributes to obtain the node loss coefficient matrix; The edge weighting update unit is used to dynamically update the initial edge weights in a two-layer graph with timestamp attributes based on the node depreciation coefficient matrix, thereby obtaining a time-decayed talent knowledge graph.

5. The knowledge graph-based talent gap diagnosis and recommendation system according to claim 4, characterized in that, The loss coefficient calculation unit includes: The idle time calculation subunit is used to calculate the time difference between the standardized timestamp of each node in the two-layer graph with timestamp attribute and the current diagnosis time to obtain the idle time of each node. The decay rate determination subunit is used to perform a halving rate calibration on the preset capability half-life corresponding to each node to obtain the node decay rate of each node. The exponential decay mapping sub-unit is used to perform negative exponential decay mapping on the idle time of the corresponding node based on the node decay rate of each node to obtain the node decay coefficient matrix.

6. The knowledge graph-based talent gap diagnosis and recommendation system according to claim 1, characterized in that, The meta-capability extraction module includes: The view separation encoding unit is used to separate the time-decayed talent knowledge graph into a surface business skills view and a bottom implicit meta-capability view according to the node type label, and to perform initial encoding on them to obtain the surface business feature matrix and the bottom implicit feature matrix. The feature decoupling unit is used to perform cross-layer information bottleneck feature decoupling on the surface business feature matrix and the bottom latent feature matrix based on the comparative information network containing cross-layer information bottlenecks in order to obtain the distribution parameters of the decoupling latent variables. The reparameterized sampling unit is used to perform meta-capability feature reparameterized sampling on auxiliary noise vectors randomly drawn from the standard normal distribution based on the decoupled latent variable distribution parameters to obtain cross-domain meta-capability vectors.

7. The knowledge graph-based talent gap diagnosis and recommendation system according to claim 1, characterized in that, The gap quantization module includes: The job feature encoding unit is used to encode the topological features of the job competency subgraph to obtain the feature representation of the target job. The spatial mapping unit is used to perform cross-domain feature nonlinear spatial mapping on the cross-domain meta-capability vector to obtain the mapped capability representation vector. The distance metric unit is used to perform a domain-adaptive distance metric on the mapped capability representation vector and the target job feature representation to obtain the job capability gap matrix.

8. The knowledge graph-based talent gap diagnosis and recommendation system according to claim 1, characterized in that, The path generation module includes: The search space construction unit is used to encapsulate the job competency gap matrix into an initial state tensor to generate a reinforcement learning state space, and to perform legality masking on the edge connections of the learning resource subgraph to generate a restricted action search space. The causal pruning unit is used to expand the initial candidate branches using a restricted breadth-first search and perform causal inference pruning in a restricted action search space to obtain a set of pruned candidate paths. The path convergence unit is used to perform global reward calculation and optimal path convergence on the pruned candidate path set based on the reinforcement learning state space to obtain cross-domain learning paths.

9. The knowledge graph-based talent gap diagnosis and recommendation system according to claim 8, characterized in that, The causal pruning unit includes: The interaction coefficient quantization subunit is used to perform sequence causal interaction coefficient quantization on all adjacent action pairs that satisfy the prior topological constraints in the restricted action search space to obtain the action pair causal interaction coefficient matrix. The branch expansion subunit is used to expand the candidate branch containing the ordered action sequence by using a restricted breadth-first search under the premise of satisfying the pre-repaired edge constraint, and query the action pair causal interaction coefficient matrix in real time to obtain the interaction coefficient of adjacent action pairs. The interactive modulation scoring subunit is used to comprehensively score the marginal effect of each step in each candidate branch, superimposed with the interaction gain term of the preceding action, based on the interaction coefficient of adjacent action pairs, to obtain the interactive modulation candidate branch set. An adaptive pruning subunit is used to perform adaptive threshold pruning on the interactive modulation candidate branch set based on the interactive modulation score to obtain a pruned candidate path set.

10. The knowledge graph-based talent gap diagnosis and recommendation system according to claim 8, characterized in that, The path convergence unit includes: The state transition subunit is used to update the gap tensor and deduct the remaining available learning time for each candidate path in the pruned candidate path set according to the action execution order, based on the reinforcement learning state space, so as to obtain the time-step state transition sequence. The instant reward calculation subunit is used to perform weighted difference calculation on the gap reduction gain and learning resource time cost of the time-step state transition sequence to obtain the instant reward for each time step. The cumulative return calculation subunit is used to calculate the cumulative return score by weighting the instantaneous returns at each time step based on the reward discount factor. The optimal path convergence subunit is used to compare the maximum value of the cumulative reward scores of each candidate path to obtain the cross-domain learning path.