Job staff recommendation explanation method and device based on gradient features and community structure

Abstract

The application relates to a post staff recommendation explanation method and device based on gradient features and community structures. The method comprises the following steps: acquiring a pre-trained convergent graph neural network and original embedding features, and simultaneously performing subsequent operations according to a knowledge graph storing entities such as personnel, skills, posts and businesses, taking the graph neural network as an on-duty staff recommendation network, combining the original embedding features and gradient vectors obtained by the graph neural network through a back propagation algorithm to form gradient features. The knowledge graph nodes are divided into communities, and node community labels are generated by maximizing the module degree. Finally, the gradient features, node community labels and the graph neural network model are input into an explanation framework, the graph mask and the feature mask are optimized to maximize the information quantization index, and a post staff recommendation result containing a key subgraph and a feature mask is generated. The method can continuously improve the explainability of on-duty staff recommendation in different scenarios.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and apparatus for interpreting job recommendations based on gradient features and community structure. Background Technology

[0002] In enterprise human resource management, on-the-job personnel recommendation is a key step in optimizing human resource allocation and improving organizational efficiency. Graph Neural Networks (GNNs), with their ability to deeply model the relationships and topological structures of entities such as personnel, skills, positions, and business scenarios, have become a core technological direction for solving the on-the-job personnel recommendation problem. However, current on-the-job personnel recommendation systems based on GNNs generally face the core pain point of low recommendation interpretability. Recommendation results often exhibit problems such as skill matching bias and insufficient scenario adaptation, leading to wasted human resources and hindered business progress. It is necessary to address the shortcomings of traditional technologies in terms of accurate matching logic and scenario adaptability to provide efficient and reliable personnel allocation support for organizational decision-making.

[0003] Current mainstream GNN-based employee recommendation technologies primarily model the association between employees and positions through path matching or feature attention. However, these methods have significant limitations in terms of matching comprehensiveness, feature effectiveness, and scenario synergy, directly hindering the improvement of recommendation interpretability. While path reasoning explains recommendation logic by finding semantic paths between entities, it still struggles to avoid the influence of subjective settings in the comprehensiveness of path selection and the quantification of path weights, and it doesn't adequately consider the synergistic effect of node features and the overall topology in the graph structure. Attention mechanisms are limited by the model's dynamic range of feature sensitivity capture, and may neglect global structural information due to excessive focus on local features, leading to one-sided interpretation results that fail to fully reflect the combined influence of features and structure in recommendation decisions. Therefore, there is an urgent need for an innovative interpretation method to support the diverse interpretability requirements in job recommendation, continuously improve the interpretability of employee recommendations in different scenarios, and provide reliable technical support for the optimal allocation of enterprise human resources. Summary of the Invention

[0004] Therefore, it is necessary to provide a method and apparatus for interpreting on-the-job personnel recommendations based on gradient features and community structure, which can continuously improve the interpretability of on-the-job personnel recommendations in different scenarios, in order to address the above-mentioned technical problems.

[0005] A job recommendation explanation method based on gradient features and community structure, the method comprising:

[0006] Obtain a pre-trained, converged graph neural network and original embedding features; the original embedding features include personnel-related features, job-related features, and business domain-related features; store a set of entities related to job and personnel recommendations based on the knowledge graph, the entity set includes personnel, skills, job titles, and business; the nodes in the knowledge graph are entities;

[0007] The graph neural network is used as the on-the-job personnel recommendation network, and the original embedded features are combined with the gradient vectors obtained by the backpropagation algorithm in the graph neural network to form gradient features.

[0008] The nodes of the knowledge graph are divided into communities, and node community tags are generated by maximizing the modularity.

[0009] Gradient features, node community labels, and a pre-trained converged graph neural network model are input into the graph neural network interpretation framework. By optimizing the graph mask and feature mask to maximize the information quantification index, interpretation results containing key subgraphs and feature masks are generated, which are the job personnel recommendation results.

[0010] A job recommendation and interpretation device based on gradient features and community structure, the device comprising:

[0011] The data preprocessing module is used to obtain the pre-trained and converged graph neural network and the original embedding features. The original embedding features include personnel-related features, job-related features, and business domain-related features. The knowledge graph stores the entity set related to job and personnel recommendations, which includes personnel, skills, job titles, and business functions. The nodes in the knowledge graph are entities.

[0012] The gradient feature generation module is used to combine the original embedded features obtained by the graph neural network with the gradient vector obtained by the backpropagation algorithm in the graph neural network to form gradient features.

[0013] The community segmentation module is used to segment nodes of the knowledge graph into communities, generating node community tags by maximizing the modularity.

[0014] The recommendation explanation module is used to input gradient features, node community labels, and a pre-trained converged graph neural network model into the graph neural network explanation framework. By optimizing the graph mask and feature mask to maximize the information quantification index, it generates an explanation result containing key subgraphs and feature masks, i.e., the job personnel recommendation result.

[0015] The aforementioned method and apparatus for interpreting job recommendations based on gradient features and community structure firstly utilizes a knowledge graph as a foundation, integrating entities and relationships such as personnel, skills, positions, and business operations to provide rich structured data support for recommendations, avoiding matching biases caused by a single information source. Gradient features are formed by combining the original embedded features with gradient vectors obtained through backpropagation, quantifying the causal relationship between node features and prediction results, accurately identifying features that play a key role in recommendations, and solving the problem of traditional attention mechanisms overemphasizing local features, ensuring feature effectiveness. Community division is performed on knowledge graph nodes, and community labels are generated. By maximizing modularity to capture the global structural effect, personnel and positions are clustered according to business scenarios, capability requirements, etc., achieving scenario-based matching at the structural level and compensating for the deficiency of path reasoning in ignoring the global topology. Finally, using a graph neural network interpretation framework, combining gradient features, community labels, and the GNN model, graph masks and feature masks are optimized to maximize information quantification indicators, mining key subgraphs and feature dimensions that influence recommendations, and clarifying the importance ranking of each element. This ensures that recommendations focus on key features while conforming to the global structure, achieving accurate matching in diverse scenarios, thereby continuously improving the interpretability of recommendations. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a job recommendation and interpretation method based on gradient features and community structure in one embodiment.

[0017] Figure 2 A block diagram of a job recommendation and interpretation device based on gradient features and community structure in one embodiment;

[0018] Figure 3 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] In one embodiment, such as Figure 1 As shown, a method for interpreting job recommendations based on gradient features and community structure is provided, including the following steps:

[0021] Step 102: Obtain the pre-trained converged graph neural network and the original embedding features; the original embedding features include personnel-related features, job-related features and business domain-related features; the knowledge graph stores the entity set related to job personnel recommendation, the entity set includes personnel, skills, job positions and business; the nodes in the knowledge graph are entities.

[0022] Step 104: Use the graph neural network as the on-the-job personnel recommendation network, and combine the acquired original embedded features with the gradient vector obtained by the backpropagation algorithm in the graph neural network to form gradient features.

[0023] The original embedded features, including personnel-related features such as skill level and work experience; job-related features such as job skill requirements and business scenarios; and business domain-related features such as professional knowledge requirements in business domains such as finance and the Internet, are combined with the gradient vectors obtained by the graph neural network through the backpropagation algorithm to form gradient features.

[0024] By concatenating the original embedded features with gradient vectors, the resulting gradient features not only retain the attribute information of the original features but also incorporate the causal relationship between features and predictions. This fusion allows the model to clearly identify which features are the key drivers of recommendation decisions during subsequent interpretation, rather than relying on subjective human judgment. For example, in the recommendation scenario for senior Java developers, gradient features can quantify the high sensitivity of "Java architecture design experience" (gradient value 0.92) and "microservice project implementation capability" (gradient value 0.88), while showing low sensitivity for "university degree" (gradient value 0.21). This result not only meets the core requirement of seniority for the position but also provides an objective basis for feature importance in recommendation interpretation, completely overcoming the limitations of traditional methods that rely on experience to determine feature importance.

[0025] Step 106: Divide the nodes of the knowledge graph into communities and generate node community tags by maximizing the modularity.

[0026] By maximizing modularity to divide knowledge graph nodes into communities, we can capture the global structural relationships between nodes from a graph learning perspective, transforming fragmented topological information into structured knowledge such as community tags, and providing key support for explaining how structure affects recommendation decisions.

[0027] First, all nodes in the knowledge graph—personnel, skills, positions, and business functions—are initially clustered: nodes with close connections and similar attributes are grouped into the same community, and the quality of the community division is evaluated by calculating modularity. A higher modularity indicates closer connections within a community and sparser connections between communities, reflecting the true topological structure of the graph. The core value of the node community tags generated through this process lies in transforming global structural information into interpretable structural attributes. For example, when recommending cross-departmental collaboration positions, community tags can reveal that the position belongs to a 'cross-departmental collaboration community,' and a candidate might be classified into this community because of "high-frequency collaborative connections with core personnel in three business departments." This structural association becomes a crucial basis for recommendation decisions, something traditional methods simply cannot capture. Simultaneously, community tags can also help explain "why some seemingly matching candidates were not recommended": for example, a candidate might possess "Python skills" (feature matching), but because they belong to an "independent development community" (a significant topological difference from the "team collaboration community" to which the position belongs), their structural fit is insufficient, hence they were not recommended. This dual explanatory dimension of features and structure completely solves the problem of fragmented structural information in traditional methods.

[0028] To address the complex entity relationships and diverse feature dimensions in job recommendations, this paper effectively integrates multiple types of entities and relationships using knowledge graphs and graph neural networks, and utilizes gradient features and community partitioning to handle complex data, thereby improving the accuracy and relevance of recommendations.

[0029] Step 108: Input gradient features, node community labels and pre-trained converged graph neural network model into the graph neural network interpretation framework. Optimize graph mask and feature mask to maximize information quantification index and generate interpretation results containing key subgraphs and feature masks, i.e., job personnel recommendation results.

[0030] Gradient features, node community labels, and a pre-trained, converged graph neural network model are input into the graph neural network interpretation framework. Gradient features are combined to quantify causal relationships between features and to reveal global topological roles through community structure. This approach considers both the impact of local features on recommendations and the matching patterns at the global structural level, resulting in more comprehensive and accurate recommendations. Then, by optimizing the graph mask, key nodes and edges are selected, along with feature masks and key node features, to maximize information quantification. Finally, an interpretation result is generated, containing a key subgraph—the core association subgraph of job positions, skills, and personnel—and its feature mask, representing the job and personnel recommendation results. This recommendation result not only provides matched personnel but also explains the basis for the recommendations, enhancing interpretability.

[0031] The aforementioned job recommendation explanation method based on gradient features and community structure firstly utilizes a knowledge graph as its foundation, integrating entities and relationships such as personnel, skills, positions, and business operations to provide rich structured data support for recommendations, avoiding matching biases caused by a single information source. Gradient features are formed by combining the original embedded features with gradient vectors obtained through backpropagation, quantifying the causal relationship between node features and prediction results, accurately identifying features that play a key role in recommendations, and addressing the problem of traditional attention mechanisms overemphasizing local features, ensuring feature effectiveness. Community division is performed on knowledge graph nodes, generating community labels. By maximizing modularity to capture global structural effects, personnel and positions are clustered according to business scenarios, capability requirements, etc., achieving scenario-based matching at the structural level and compensating for the deficiency of path reasoning in ignoring the global topology. Finally, using a graph neural network explanation framework, combining gradient features, community labels, and the GNN model, graph masks and feature masks are optimized to maximize information quantification indicators, uncovering key subgraphs and feature dimensions that influence recommendations, and clarifying the importance ranking of each element. This ensures that recommendations focus on key features while conforming to the global structure, achieving accurate matching in diverse scenarios, thereby continuously improving the interpretability of recommendations.

[0032] In one embodiment, the acquired original embedding features are combined with the gradient vectors obtained by the backpropagation algorithm in the graph neural network to form gradient features, including:

[0033] The acquired original embedded features are combined with the gradient vectors obtained from the backpropagation algorithm in the graph neural network to form the gradient features.

[0034] ;

[0035] in, This indicates a tensor splicing operation. Represents the original embedded features. This represents the gradient vector.

[0036] In one embodiment, the Louvain community detection algorithm is used to divide the nodes of the knowledge graph into communities, generating node community tags by maximizing the modularity of the nodes, including:

[0037] Each node in the knowledge graph is treated as an independent community. For each node, it is sequentially attempted to be assigned to the community of its neighboring nodes. The assignment is determined based on changes in modularity. This process is repeated until the node's community remains constant. Then, the community is compressed into a new node, and the edge weights are transformed. This process is repeated until the modularity of the entire graph no longer changes and reaches its maximum value, generating the node's community label. Through these algorithmic steps, a knowledge graph network of job personnel can be obtained. All nodes Community category tags.

[0038] In a specific embodiment, the knowledge graph View it as a network containing a set of entities For the set of nodes and the set of relations in a network For the edge set of the network, Adjacency matrix elements (representing nodes) With nodes The connection relationship, ,in ), For nodes The degree, The total number of sides, For nodes The associated community tag. For ease of calculation, an indicator function is defined. (when If the value is 1, then the value is 0.

[0039] In one embodiment, the modularity is

[0040] ;

[0041] in, This represents the sum of the weights of the edges within community c. Indicates community The sum of the weights of the edges connecting the nodes within the same node. Represents the elements of the adjacency matrix. This represents the total number of edges in the knowledge graph. Represents a node The degree, Represents a node The degree, Indicates an indicator function, when The value is 1 if the condition is met, and 0 otherwise. For nodes Community tags associated with For nodes The community tags to which it belongs.

[0042] In one embodiment, gradient features, node community labels, and a pre-trained converged graph neural network model are jointly input into the graph neural network interpretation framework. By optimizing the graph mask and feature mask to maximize the information quantification index, an interpretation result containing key subgraphs and feature masks is generated, including:

[0043] Gradient features, node community labels, and a pre-trained converged graph neural network model are input into the graph neural network interpretation framework to maximize the information quantification index to construct the optimization objective. The parameters of the graph mask and feature mask are iteratively updated through the gradient descent algorithm to generate interpretation results containing key subgraphs and feature masks.

[0044] In one embodiment, the information quantification index is:

[0045] ;

[0046] in, This indicates important subgraph structures. Represents a feature mask. Represents a random variable. Represents a subset of key features. Indicates through feature mask Selected key feature subset Represents random variables The entropy.

[0047] In one embodiment, the key subgraph is the core association subgraph of job positions, skills, and personnel.

[0048] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0049] In one embodiment, such as Figure 2 As shown, a job recommendation and interpretation device based on gradient features and community structure is provided, including: a data preprocessing module 202, a gradient feature generation module 204, a community partitioning module 206, and a recommendation interpretation module 208, wherein:

[0050] The data preprocessing module 202 is used to obtain the pre-trained and converged graph neural network and the original embedding features; the original embedding features include personnel-related features, job-related features and business domain-related features; the knowledge graph stores the entity set related to job personnel recommendation, the entity set includes personnel, skills, job positions and business; the nodes in the knowledge graph are entities;

[0051] The gradient feature generation module 204 is used to combine the obtained original embedded features with the gradient vector obtained by the backpropagation algorithm in the graph neural network to form gradient features by using the graph neural network as the on-the-job personnel recommendation network.

[0052] Community segmentation module 206 is used to segment nodes of the knowledge graph into communities, and generates node community tags by maximizing the modularity.

[0053] The recommendation explanation module 208 is used to input gradient features, node community labels and a pre-trained converged graph neural network model into the graph neural network explanation framework. By optimizing the graph mask and feature mask to maximize the information quantification index, it generates an explanation result containing key subgraphs and feature masks, i.e., the job personnel recommendation result.

[0054] Specific limitations regarding the job recommendation and interpretation device based on gradient features and community structure can be found in the limitations of the job recommendation and interpretation method based on gradient features and community structure mentioned above, and will not be repeated here. Each module in the aforementioned job recommendation and interpretation device based on gradient features and community structure can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0055] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a job recommendation interpretation method based on gradient features and community structure. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0056] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0057] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0058] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0059] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A job recommendation and interpretation method based on gradient features and community structure, characterized in that, The method includes: Obtain a pre-trained converged graph neural network and original embedding features; the original embedding features include personnel-related features, job-related features, and business domain-related features; store a set of entities related to job and personnel recommendations based on a knowledge graph, the set of entities including personnel, skills, jobs, and business; the nodes in the knowledge graph are entities; The graph neural network is used as an on-the-job personnel recommendation network, and the acquired original embedded features are combined with the gradient vector obtained by the backpropagation algorithm in the graph neural network to form gradient features; The nodes of the knowledge graph are divided into communities, and node community tags are generated by maximizing the modularity. The gradient features, node community labels, and a pre-trained converged graph neural network model are input into the graph neural network interpretation framework. By optimizing the graph mask and feature mask to maximize the information quantification index, interpretation results containing key subgraphs and feature masks are generated, which are the job personnel recommendation results. The acquired raw embedding features are combined with the gradient vectors obtained from the backpropagation algorithm in the graph neural network to form gradient features, including: The acquired original embedded features are combined with the gradient vector obtained by the backpropagation algorithm in the graph neural network to form the gradient features: in, This indicates a tensor splicing operation. Represents the original embedded features. Represents the gradient vector; The Louvain community detection algorithm is used to divide nodes in a knowledge graph into communities. Node community tags are generated by maximizing modularity, including: Treat each node in the knowledge graph as an independent community, and try to assign each node to the community of its neighboring nodes in turn. Decide whether to assign based on the change in modularity. Repeat this process until the community to which the node belongs remains unchanged. Then compress the community into a new node and transform the edge weights. Repeat this process until the modularity of the entire graph no longer changes and reaches its maximum value, and generate the node community label of the node.

2. The method according to claim 1, characterized in that, The modularity is: in, This represents the sum of the weights of the edges within community c. Indicates community The sum of the weights of the edges connecting the nodes within the same node. Represents the elements of the adjacency matrix. This represents the total number of edges in the knowledge graph. Represents a node The degree, Represents a node The degree, Indicator function when The value is 1 if the condition is met, and 0 otherwise. For nodes Community tags associated with For nodes The community tags to which it belongs.

3. The method according to claim 1, characterized in that, Gradient features, node community labels, and a pre-trained, converged graph neural network model are input into the graph neural network interpretation framework. By optimizing the graph mask and feature mask to maximize the information quantification index, interpretation results containing key subgraphs and feature masks are generated, including: Gradient features, node community labels, and a pre-trained converged graph neural network model are input into the graph neural network interpretation framework to maximize the information quantification index to construct the optimization objective. The parameters of the graph mask and feature mask are iteratively updated through the gradient descent algorithm to generate interpretation results containing key subgraphs and feature masks.

4. The method according to claim 1, characterized in that, The information quantification index is: in, This indicates important subgraph structures. Represents a feature mask. Represents a random variable. Represents a subset of key features. Indicates through feature mask Selected key feature subset Let Y represent the entropy of the random variable Y.

5. The method according to claim 1, characterized in that, The key subgraph is the core relational subgraph of job positions, skills, and personnel.

6. A job recommendation and interpretation device based on gradient features and community structure, characterized in that, The device includes: The data preprocessing module is used to obtain a pre-trained and converged graph neural network and raw embedding features; the raw embedding features include personnel-related features, job-related features, and business domain-related features; a set of entities related to job and personnel recommendations is stored according to the knowledge graph, the set of entities includes personnel, skills, jobs, and business; the nodes in the knowledge graph are entities; A gradient feature generation module is used to combine the graph neural network as an on-the-job personnel recommendation network with the acquired original embedded features and the gradient vectors obtained by the backpropagation algorithm in the graph neural network to form gradient features, including: The acquired original embedded features are combined with the gradient vector obtained by the backpropagation algorithm in the graph neural network to form the gradient features: in, This indicates a tensor splicing operation. Represents the original embedded features. Represents the gradient vector; The community segmentation module is used to segment nodes of the knowledge graph into communities. It generates node community tags by maximizing module degree, including: Treat each node in the knowledge graph as an independent community, and try to assign each node to the community of its neighboring nodes in turn. Decide whether to assign based on the change in modularity. Repeat this process until the community to which the node belongs remains unchanged. Then compress the community into a new node and transform the edge weights. Repeat this process until the modularity of the entire graph no longer changes and reaches its maximum value, and generate the node community label of the node. The recommendation explanation module is used to input gradient features, node community labels, and a pre-trained converged graph neural network model into the graph neural network explanation framework. By optimizing the graph mask and feature mask to maximize the information quantification index, it generates an explanation result containing key subgraphs and feature masks, i.e., the job personnel recommendation result.

Images