A two-way employment recommendation method based on career development trend modeling

Abstract

The application belongs to the technical field of data analysis. The application provides a two-way employment recommendation method based on career development trend modeling. The disclosure embodiment models the career development trend of job seekers by using the DMU-Net architecture. The DMU-Net realizes multi-level feature extraction through a time series feature pyramid network, thereby capturing seasonal changes in the shallow layer and identifying long-term trends in the deep layer. Then, a cross-layer attention mechanism is designed to integrate feature information of different scales, and a dynamic multi-key value memory matrix is introduced to effectively reduce information loss in the down-sampling process. In addition, through the design of a double-view encoder and an adaptive feature fusion layer, two-way matching of job seekers and recruitment demand is realized, and more accurate employment recommendation is provided.

Description

TECHNICAL FIELD

[0001] Embodiments of the present disclosure relate to the technical field of data analysis, and particularly, to a two-way employment recommendation method based on career development trend modeling. BACKGROUND

[0002] The employment recommendation system is a technology that gradually emerges with the development of the Internet and big data technology. It uses intelligent algorithms and a large amount of behavior data of job seekers to recommend suitable job opportunities for users. In the modern job market, the matching between job seekers and employers becomes increasingly complex. The traditional recruitment process not only takes a long time, but also easily misses potential suitable talents. Therefore, how to effectively match suitable talents to suitable positions has become the focus of attention of enterprises and researchers. The employment recommendation system analyzes the information of job seekers' skills, interests, work experience, etc., and combines the specific requirements of the position to help enterprises find candidates who meet the requirements more efficiently, and also provides more personalized position recommendations for job seekers.

[0003] Current employment recommendation algorithms mainly rely on two types of information resources: one is the historical application records of job seekers, and the other is the text information of resumes and job descriptions. For historical application records, the collaborative filtering method is usually used to mine the potential matching degree from the application relationship between job seekers and positions. In the processing of text information, the key features in the job seeker and position description are extracted through natural language processing technology to achieve more accurate matching recommendation. In addition, some advanced hybrid models combine interactive data and text features to further improve the recommendation effect, so that the system can better meet the diverse needs of enterprises and job seekers.

[0004] Although the automated recruitment system has made significant progress in job recommendation, the current recommendation system still faces some problems. They are mainly based on the historical behavior and current ability of job seekers, and the dependence on this static information limits the system's ability to capture the long-term career goals of job seekers. Therefore, the factor of "future career development trend" is ignored, which leads to the fact that the recommendation results cannot fully support the career growth of job seekers.

[0005] Therefore, it is necessary to improve one or more problems in the above related technical solutions.

[0006] It should be noted that this section aims to provide background or context to the technical solutions of the present disclosure stated in the claims. The description herein is not admitted to be prior art merely because it is included in this section. SUMMARY

[0007] The purpose of the embodiments of the present disclosure is to provide a two-way employment recommendation method based on career development trend modeling, thereby at least partially overcoming one or more problems caused by limitations and defects of related technologies.

[0008] According to the embodiments of the present disclosure, a two-way employment recommendation method based on career development trend modeling is provided, which comprises:

[0009] a career development trend model is constructed; wherein the career development trend model comprises a knowledge graph embedding layer, a multi-scale dual-view meta-path encoding module, and an adaptive dual-view feature fusion layer; the multi-scale dual-view meta-path encoding module comprises a DMU-Net unit and a Bi-LSTM unit, the DMU-Net unit comprises a time series feature pyramid network, a cross-layer attention fusion module, and a dynamic multi-key-value memory matrix module; the adaptive dual-view feature fusion layer comprises an adaptive feature fusion layer and a dual-view meta-path aggregation layer;

[0010] an interaction data between a job seeker and a position is constructed to construct a meta-path instance;

[0011] a TransR model in the knowledge graph embedding layer is used to map entities and relationships of the meta-path instance to a low-dimensional vector space to obtain meta-path embedding;

[0012] a time series feature pyramid network is used to extract deep trend information and shallow seasonal features layer by layer through multi-layer pooling operations on the meta-path embedding to generate multi-scale features;

[0013] a cross-layer attention fusion module is used to effectively integrate multi-scale features at different levels to obtain initial cross-layer fusion features at the current level;

[0014] a dynamic multi-key-value memory matrix module is used to update the memory of the initial cross-layer fusion features to obtain the final cross-layer fusion features at the current level;

[0015] a Bi-LSTM unit is used to capture dependency relationships under the perspectives of the candidate and the position respectively to generate dual-view aggregated representations, and the meta-path embedding is input as candidate feature for inversion, and the inverted meta-path embedding is input as position feature, so that the DMU-Net unit can obtain dual-view features; wherein the dual-view aggregated representations comprise candidate aggregated representations and position aggregated representations, and the dual-view features comprise candidate feature representations and position feature representations;

[0016] an adaptive feature fusion layer uses an adaptive attention mechanism to weight features of the dual-view aggregated representations and the dual-view features, effectively fuses multi-dimensional features, and generates candidate comprehensive feature representations and position comprehensive feature representations;

[0017] input the candidate comprehensive feature representation and the position comprehensive feature representation into a dual-view meta-path aggregation layer, set different first attention weights for the candidate and the position respectively, and form a dual-view meta-path aggregation representation; wherein the dual-view meta-path aggregation representation includes a candidate meta-path aggregation representation and a position meta-path aggregation representation;

[0018] calculate the final matching probability of the candidate and the position according to the dual-view meta-path aggregation representation, and obtain a recommendation result according to the final matching probability.

[0019] Further, in the step of constructing the meta-path instance by constructing the interaction data of the job seeker and the position, comprising:

[0020] construct an interaction graph according to the interaction data of the job seeker and the position, and establish a knowledge graph in combination with the attribute information of the job seeker and the attribute information of the position;

[0021] According to the requirements of the recommendation task, define different meta-path types between the candidate and the position;

[0022] In the knowledge graph, use a random walk algorithm to generate instances, and extract meta-path instances that meet the task requirements from the meta-path types.

[0023] Further, in the step of generating multi-scale features by embedding the meta-path into the time series feature pyramid network through multi-layer pooling operation to extract deep trend information and shallow seasonal features layer by layer, comprising:

[0024] embedding the meta-path into the time series feature pyramid network, inputting the meta-path into the time series feature pyramid network, dividing the data into L layers through an average pooling function, and using different depths to extract deep trend information and shallow seasonal features in each layer to generate multi-scale features ; wherein,

[0025]

[0026] wherein, is the output feature of the i-th layer, is the output feature of the 1st l- layer, represents an average pooling operation.

[0027] Further, in the step of effectively integrating the multi-layer input data of different levels by using a cross-layer attention fusion module to obtain the initial cross-layer fusion features of the current level, comprising:

[0028] inputting the multi-scale features into the cross-layer attention fusion module, calculating the correlation between time steps through a multi-head self-attention mechanism, and splicing and integrating after parallel processing of multiple attention heads;

[0029] Apply residual connection and layer normalization operations to the output of multi-head self-attention, and then enrich the feature expression through two linear transformations and ReLU activation in the feedforward network;

[0030] Apply layer normalization and residual connection to obtain the attention feature set ;

[0031] The attention features of the current layer are spliced ​​with the cross-layer fusion features of the previous layer, and the cross-layer attention mechanism is introduced again for fusion. The linear mapping restores the spliced ​​features to the original sequence length to obtain the initial cross-layer fusion features of the current layer. .

[0032] Furthermore, the step of using a dynamic multi-key memory matrix module to update the memory of the initial cross-layer fusion features to obtain the final cross-layer fusion features of the current layer includes:

[0033] During the downsampling process of the DMU-Net unit, each layer Output features Update features are obtained through MLP processing:

[0034]

[0035] Where, is the linear transformation matrix, is the first bias term, To update features, is the sigmoid function;

[0036] Through the global bond matrix corresponding to each layer and output features Calculate the second attention weight associated with the current level feature:

[0037]

[0038] Where, is the second attention weight, For the The first layer of the global bond matrix key vectors, For the The first layer of the global bond matrix key vectors;

[0039] Update the private memory matrix using the second attention weight:

[0040]

[0041] Where, For the i The memory slot is inl The value of the layer, For the i The memory slot is in l- The value of layer 1;

[0042] During the upsampling process of the DMU-Net unit, the initial cross-layer fusion features are multiplied by the corresponding global key matrix to calculate the third attention weight Based on the third attention weight, perform weighted summation on each memory slot of the private memory matrix to obtain the memory preference representation of the current level:

[0043]

[0044] The initial cross-layer fusion features and memory preference representation Add together to get the final cross-layer fusion features of this level .

[0045] Furthermore, the Bi-LSTM unit is used to capture the dependency relationships from the candidate and position perspectives, generating a dual-perspective aggregate representation. The meta-path embedding is then inverted as the candidate feature input, and the inverted meta-path embedding is used as the position feature input. The steps to enable the DMU-Net unit to obtain dual-perspective features include:

[0046] The Bi-LSTM unit is introduced to generate the forward hidden state representation and backward hidden state representation of candidates and positions respectively, and the forward hidden state representation and the backward hidden state representation are averaged to obtain the dual-view aggregate representation of candidates and positions:

[0047]

[0048] Where, For candidate aggregation, is the position aggregation representation, For Bi-LSTM units at time steps The forward hidden state on is represented as, For Bi-LSTM units at time steps The backward hidden state representation on T is the length of the meta-path;

[0049] The meta-path embedding is used as the feature input of the candidate, and the inverted meta-path embedding is used as the feature input of the position, so that the DMU-Net unit can obtain dual-view features:

[0050]

[0051] in, represents the candidate feature representation, representing the position feature representation, representing the processing procedure of the DMU-Net model.

[0052] Further, in the step of effectively fusing the multi-dimensional features by weighting the dual-view aggregated representation and the dual-view features through the adaptive attention mechanism to generate the candidate comprehensive feature representation and the position comprehensive feature representation, the adaptive feature fusion layer comprises:

[0053] stacking the dual-view aggregated representation and the dual-view features extracted by the DMU-Net unit and the Bi-LSTM unit to form the candidate fusion feature representation and the position fusion feature representation .

[0054] calculating the fourth attention weight of the stacked features by using the adaptive attention mechanism:

[0055]

[0056] wherein, is the fourth attention weight corresponding to the candidate, is the first trainable weight matrix, is the candidate fusion feature representation, is the second bias term, is the fourth attention weight corresponding to the position, is the second trainable weight matrix, is the position fusion feature representation, is the third bias term.

[0057] weighting and summing the candidate feature representation and the position feature representation by using the fourth attention weight to obtain the fused candidate comprehensive feature representation and the position comprehensive feature representation :

[0058]

[0059] wherein, is the weight assigned to the kth feature type of the candidate by the attention mechanism, is the feature representation of the kth feature type of the candidate, is the weight assigned to the kth feature type of the position by the attention mechanism, is the feature representation of the kth feature type of the position.

[0060] Further, in the step of inputting the candidate comprehensive feature representation and the position comprehensive feature representation into the dual-view meta-path aggregation layer and setting different first attention weights for the candidate and the position to form the dual-view meta-path aggregated representation, the method comprises:

[0061] assigning first attention weights to the fused candidate comprehensive feature representation and the position comprehensive feature representation;

[0062] performing weighted summation on the meta-path representations using the aggregated attention weights to obtain a dual-view meta-path aggregated representation:

[0063]

[0064] wherein, is a candidate meta-path aggregated representation, is a position meta-path aggregated representation, is a candidate attention weight of the i-th meta-path, is a position attention weight of the i-th meta-path, is a candidate comprehensive feature representation of the i-th meta-path, is a position comprehensive feature representation of the i-th meta-path. l Further, the final matching probability of the candidate and the position is calculated according to the dual-view meta-path aggregated representation, and the recommended result is obtained according to the final matching probability in the step of: l the matching score is normalized by a Sigmoid function to obtain the final matching probability of the candidate and the position, and the process can be represented as:

[0065]

[0066]

[0067]

[0068] wherein, is a first weight of a linear layer, is a second weight of a linear layer, is a fourth bias term, is a fifth bias term, represents the matching degree of the candidate to the position , and represents the matching degree of the position to the candidate ;

[0069] the average value is taken as the final matching probability, and the formula is as follows:

[0070]

[0071] wherein, is the final matching probability.

[0072] Further, the total loss function of the career development trend model is:​​​​

[0073]

[0074]

[0075]

[0076] wherein, is a total loss function, is a bilateral matching loss, is a unilateral matching loss, is a weighting coefficient, is a training data set representation, is a job seeker one-way interaction score, is a job one-way interaction score, is a score difference between representations of one-way matching.

[0077] Embodiments of the present disclosure provide technical solutions that can include the following beneficial effects:

[0078] In embodiments of the present disclosure, the above-mentioned two-way employment recommendation method based on career development trend modeling is used to model the career development trend of job seekers using a dynamic memory enhanced U-Net (DMU-Net) architecture. DMU-Net achieves multi-level feature extraction through a time series feature pyramid network, thereby capturing seasonal changes in the shallow layer and identifying long-term trends in the deep layer. A cross-layer attention mechanism is then designed to fuse feature information of different scales, and a dynamic multi-key-value memory matrix is introduced to effectively reduce information loss during downsampling. By introducing a multi-scale dual-view encoder (i.e., Bi-LSTM) and a cross-layer attention mechanism, the model can capture complex temporal dependency relationships at different scales and identify potential directions of career development for job seekers. This model not only integrates the current abilities and behaviors of job seekers, but also focuses on their future growth paths, making the recommendation results more forward-looking and accurate. By using a dynamic multi-key-value memory matrix, the model effectively retains and retrieves low-level feature information during downsampling and upsampling, reducing the loss of feature information during level conversion, thereby ensuring the balance between high-level abstract features and low-level detailed features. Through dual-view meta-path aggregation, the weights and importance of features are dynamically adjusted from the needs of candidates and employers, improving the understanding of the matching needs of both parties and generating more targeted matching features. BRIEF DESCRIPTION OF DRAWINGS

[0079] The accompanying drawings, which are incorporated herein and form a part of the specification, illustrate embodiments consistent with the present disclosure and, together with the description, further serve to explain the principles of the disclosure. It is to be noted, as evident from the following description, that the drawings need not be to scale as the emphasis instead is on the functional teaching speedily conveyed by the drawings to a person with ordinary skill in the art. While the embodiments will be described in the general context of a method, it should be clear these embodiments can also be implemented in combination with software carried by one or more computers, and that one or more computer programs comprising computer-executable instructions can be used to program computers to perform a process.

[0080] Figure 1 A step diagram illustrating a two-way employment recommendation method based on career development trend modeling in an exemplary embodiment of the present disclosure is shown;

[0081] Figure 2 A framework diagram illustrating a career development trend model in an exemplary embodiment of the present disclosure is shown;

[0082] Figure 3 A framework diagram illustrating a DMU-Net unit in an exemplary embodiment of the present disclosure is shown;

[0083] Figure 4 A memory update operation flowchart diagram of a DMU-Net unit in an exemplary embodiment of the present disclosure is shown;

[0084] Figure 5 A memory read operation flowchart diagram of a DMU-Net unit in an exemplary embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0085] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations can be implemented in any

[0086] In addition, the drawings are to be regarded as being schematic only and therefore are not intended to limit the scope of the disclosure. Like reference numerals are used to designate like parts throughout the several views and the repetitive description of like parts will be omitted for the sake of brevity. Some of the block components shown in the drawings can be functional blocks that do not necessarily have to have physical or logical boundaries in the corresponding circuitry.

[0087] A two-way employment recommendation method based on career development trend modeling is provided in the present example implementation. Referring to FIG. 1, the two-way employment recommendation method based on career development trend modeling can include steps S101-S110. Figure 1

[0088] ​Step S101: constructing a career development trend model; wherein the career development trend model comprises a knowledge graph embedding layer, a multi-scale dual-view meta-path encoding module, and an adaptive dual-view feature fusion layer; the multi-scale dual-view meta-path encoding module comprises a DMU-Net unit and a Bi-LSTM unit, the DMU-Net unit comprises a time series feature pyramid network, a cross-layer attention fusion module, and a dynamic multi-key-value memory matrix module; the adaptive dual-view feature fusion layer comprises an adaptive feature fusion layer and a dual-view meta-path aggregation layer;

[0089] Step S102: constructing an interactive data construction meta-path instance of a job seeker and a position;

[0090] Step S103: mapping entities and relationships of the meta-path instance to a low-dimensional vector space by using a TransR model in the knowledge graph embedding layer to obtain meta-path embedding;

[0091] Step S104: generating multi-scale features by layer-by-layer extracting deep trend information and shallow seasonal characteristics of the meta-path embedding through a multi-layer pooling operation by using a time series feature pyramid network;

[0092] Step S105: effectively integrating multi-scale features at different levels by using a cross-layer attention fusion module to obtain initial cross-layer fusion features at the current level;

[0093] Step S106: updating the initial cross-layer fusion features by using a dynamic multi-key-value memory matrix module to obtain final cross-layer fusion features at the current level;

[0094] Step S107: respectively capturing dependency relationships under the perspectives of the candidate and the position by using a Bi-LSTM unit to generate dual-view aggregated representations, and inputting the meta-path embedding as candidate feature input for inversion, and inputting the inverted meta-path embedding as position feature input, so that the DMU-Net unit can obtain dual-view features; wherein the dual-view aggregated representations comprise candidate aggregated representations and position aggregated representations, and the dual-view features comprise candidate feature representations and position feature representations;

[0095] Step S108: effectively fusing multi-dimensional features by weighting the dual-view aggregated representations and the dual-view features through an adaptive attention mechanism in the adaptive feature fusion layer to generate candidate comprehensive feature representations and position comprehensive feature representations;

[0096] Step S109: inputting the candidate comprehensive feature representations and the position comprehensive feature representations into the dual-view meta-path aggregation layer, setting different first attention weights for the candidate and the position respectively, and forming dual-view meta-path aggregated representations; wherein the dual-view meta-path aggregated representations comprise candidate meta-path aggregated representations and position meta-path aggregated representations;

[0097] Step S110: Calculate the final matching probability between the candidate and the position based on the dual-view meta-path aggregation representation, and obtain the recommendation result based on the final matching probability.

[0098] The aforementioned bidirectional job recommendation method based on career development trend modeling utilizes a dynamic memory-augmented U-Net architecture to model a job seeker's career development trends. DMU-Net utilizes a time series feature pyramid network to achieve multi-level feature extraction, capturing seasonal variations at the shallowest level and identifying long-term trends at the deeper level. A cross-layer attention mechanism is then designed to fuse feature information at different scales, and a dynamic multi-key-value memory matrix is ​​introduced to effectively reduce information loss during downsampling. By introducing a multi-scale dual-view encoder and cross-layer attention mechanism, the model captures complex temporal dependencies at different scales and identifies potential career development paths for job seekers. This model not only integrates job seekers' current abilities and behaviors but also focuses on their future growth paths, resulting in more forward-looking and accurate recommendation results. By utilizing a dynamic multi-key-value memory matrix, the model effectively retains and retrieves low-level feature information during downsampling and upsampling, minimizing feature information loss during hierarchical transformations and ensuring a balance between high-level abstract features and low-level detailed features. Through dual-perspective meta-path aggregation, the weight and importance of features are dynamically adjusted based on the needs of candidates and recruiters respectively, which improves the ability to understand the matching needs of both parties and generates more targeted matching features.

[0099] Below, we will refer to Figures 1 to 5 Each step of the above-mentioned two-way employment recommendation method based on career development trend modeling in this example implementation is described in more detail.

[0100] In step S101, the career development trend model includes a knowledge graph embedding layer, a multi-scale dual-view meta-path encoding module, and an adaptive dual-view feature fusion layer. The multi-scale dual-view meta-path encoding module includes a DMU-Net unit and a Bi-LSTM unit. The DMU-Net unit includes a time series feature pyramid network, a cross-layer attention fusion module, and a dynamic multi-key-value memory matrix module. The adaptive dual-view feature fusion layer includes an adaptive feature fusion layer and a dual-view meta-path aggregation layer. The multi-scale dual-view meta-path encoding module extracts multi-level career development trend features through the DMU-Net and captures complex dependencies from the dual perspectives of candidates and positions. The adaptive dual-view feature fusion layer fuses the features of the DMU-Net and Bi-LSTM through an attention mechanism, assigning higher weights to important features while highlighting the different requirements of candidates and positions on the meta-path.

[0101] like Figure 2 Shown is the framework diagram of the career development trend model.

[0102] like Figure 3 The figure shows the framework diagram of the DMU-Net unit.

[0103] In step S102, an interaction graph is first constructed based on the interaction data between job seekers and employers, and a complete knowledge graph is created by combining the attribute information of both parties. Based on this, different meta-path types between candidates and positions are defined according to the requirements of the recommendation task. For example, meta-path types could include "candidate - possessed skills - required skills - position" or "candidate - possessed education - required education - position." These paths are used to capture the potential relationships between candidates and positions. Within the knowledge graph, a random walk algorithm is used to generate instances, extracting meta-path instances that meet the task requirements from the predefined meta-path types.

[0104] In step S103, the meta-path instance extracted above is input into the knowledge graph embedding layer, and the entities and relationships in the knowledge graph are embedded into the low-dimensional vector space using the TransR model to obtain the meta-path embedding , used for subsequent feature extraction.

[0105] In step S104, multi-scale feature extraction and encoding are performed in the multi-scale dual-view meta-path encoding module, and the meta-path instances of candidates and positions are modeled using DMU-Net. First, the temporal feature pyramid network of DMU-Net is used to extract deep trend information and shallow seasonal features layer by layer through multi-layer pooling operations to form the L-layer output. .

[0106] Specifically, the meta-path is embedded into It is divided into L layers, and each layer extracts trend features of different depths through average pooling. Through this multi-layer processing, FPN captures seasonal changes in the shallow layer and separates long-term trends in the deep layer, ultimately generating multi-level input data. :

[0107]

[0108] in, For the The pooling result of the layer, Represents an average pooling operation.

[0109] In step S105, DMU-Net uses the cross-layer attention mechanism to integrate features from different layers. The cross-layer attention fusion module captures the global dependencies between each time step through the multi-head self-attention mechanism. Then, the obtained current layer attention features are spliced ​​with the cross-layer fusion features of the previous layer, and the cross-layer attention mechanism is used to strengthen the fusion again to obtain the cross-layer fusion features of the current layer. .

[0110] Specifically, the multi-scale feature The correlation between time steps is calculated by the multi-head self-attention mechanism, and multiple attention heads are processed in parallel and then spliced and integrated. Subsequently, the output of the multi-head self-attention is applied to the residual connection and layer normalization operation, and then the feature expression is enriched through two linear transformations and ReLU activation of the feedforward network; layer normalization and residual connection are applied again to obtain the attention feature set . Then, the attention features of the current layer are spliced with the cross-layer fusion features of the previous layer, and the fusion is strengthened again through the cross-layer attention mechanism, so that the finally generated features are consistent with the input feature length, ensuring the integrity of the cross-layer features, and obtaining the cross-layer fusion features of the current level (i.e., the initial cross-layer fusion features).

[0111] In step S106, the downsampling process of the DMU-Net may cause the loss of low-level features, therefore a dynamic multi-key value memory matrix is introduced, including a global and a private memory matrix. The global matrix is used to generate dynamic weights through an attention mechanism, and the private memory matrix independently stores features for each user, thereby preserving user preferences at different levels. The private memory matrix can update the memory during the downsampling process, maintain the short-term and long-term features of the candidates, and read the memory during the upsampling process, thereby avoiding the loss of information caused by downsampling.

[0112] Specifically, memory update, in the downsampling process of the DMU-Net, the output of each layer is first updated as the input sequence of the memory matrix to update the private memory matrix; the input features are processed using a multi-layer perceptron (MLP) to generate updated features that can capture the key information of the layer, as follows:

[0113]

[0114] wherein, is a linear transformation matrix, is a bias term.

[0115] Then, each layer has a corresponding global key matrix , which is used to calculate the attention weight related to the current level feature , as follows:

[0116]

[0117] wherein, denotes the th key vector of the th global key matrix of the th layer.

[0118] Finally, the private memory matrix updates the memory slots layer by layer before the downsampling average pooling of DMU-Net, as follows:

[0119]

[0120] in, Indicates the The memory slot is in The value of the layer.

[0121] Through this weighted update method, the model can dynamically retain key feature information between levels.

[0122] like Figure 4 Figure 2 shows the memory update operation flow chart of the DMU-Net unit.

[0123] Memory retrieval, in the DMU-Net upsampling process, the model first fuses the cross-layer features and the corresponding global bond matrix Multiply and calculate the attention weight Based on these weights, a weighted sum is performed on each memory slot of the private memory matrix to obtain the memory preference representation of the current level, as follows:

[0124]

[0125] in, Represents a private memory matrix Middle i The contents of the slots, and Controls the contribution ratio of this slot in the current level preference representation.

[0126] Next, the cross-layer fusion features and memory preference representation Add together to get the final output of this level （ That is, the final cross-layer fusion feature).

[0127] like Figure 5 The figure shows the memory reading operation flow chart of the DMU-Net unit.

[0128] In step S107, the dual-view meta-path encoding in the multi-scale dual-view meta-path encoding module is performed to increase the bidirectional feature modeling capability for DMU-Net through Bi-LSTM. The forward and backward hidden states of Bi-LSTM capture the dependency relationship under the candidate and position perspectives respectively, generating an aggregated representation of the dual perspective. and Meanwhile, the original meta-path embedding is input as the candidate feature, and the reversed meta-path embedding is input as the position feature, enabling the DMU-Net to obtain double-view features and .

[0129] Specifically, to compensate for the shortcomings of DMU-Net in sequence order modeling, the model introduces Bi-LSTM to capture long-distance dependencies through its bidirectional structure, and generates forward and backward hidden state representations of the candidate and the position respectively. The average of the bidirectional hidden state forms the sequential dependency representation of the candidate and the position:

[0130]

[0131] wherein, and represent the representations of the candidate and the position on the meta-path respectively. represents the length of the meta-path, and represent the forward and backward hidden states of Bi-LSTM at time step .

[0132] The model inputs the original meta-path embedding as the candidate feature and the reversed meta-path embedding as the position feature, enabling the DMU-Net to obtain double-view information. This symmetric design enhances the matching feature expression between the candidate and the position:

[0133]

[0134] wherein, represents the candidate feature representation, represents the position feature representation, represents the processing process of the DMU-Net model.

[0135] In step S108, the adaptive feature fusion module in the adaptive double-view feature fusion layer is performed, and the candidate and position features extracted by DMU-Net and Bi-LSTM are input into the adaptive double-flow fusion layer. Through the adaptive attention mechanism, the features from different sources are weighted, and the multi-dimensional features are effectively fused to generate the comprehensive feature representations of the candidate and the position and This step ensures that the model can integrate feature information from different sources, enhancing the comprehensiveness and accuracy of matching.

[0136] Specifically, the candidate and position features extracted by DMU-Net and Bi-LSTM are stacked respectively to form the fusion feature representations and to ensure the independence of the two features during the fusion process. The adaptive attention mechanism is used to calculate the weights of stacked features and :

[0137]

[0138] where, , is a trainable weight matrix, , is a bias term. is a sigmoid function used to convert attention scores into weights between 0 and 1.

[0139] Then, the two different types of features are weighted and summed using attention weights to obtain the fused candidate feature representation and job feature representation:

[0140]

[0141] In the above formula, and are the weights assigned by the attention mechanism to the kth feature type of the user and the item, respectively, reflecting the relative importance of DMU-Net and Bi-LSTM features in the current task. and represent the feature representation of the kth feature type of the user and the item, respectively.

[0142] In step S109, the dual-view meta-path aggregation module in the adaptive dual-view feature fusion layer is performed, and different attention weights are set for the candidate and the recruiter during the fusion of features, forming dual-view meta-path aggregation representations and . This aggregation layer focuses on capturing the respective needs of candidates and recruiters, highlighting information critical to matching while suppressing irrelevant features.

[0143] Specifically, the aggregation attention weights are assigned to the features of the candidate and the job to reflect the different needs of the candidate and the recruiter in the same meta-path. These weights are calculated based on the candidate and job feature representations to form the aggregated feature representations:

[0144]

[0145] where, and represent the aggregated candidate and job representations, and are the attention weights of the th meta-path, using and as input.

[0146] In step S110, the matching probability is predicted, and the candidate and the job feature after the dual-view aggregation are input into a linear layer to calculate the matching probability. The matching score is normalized by a Sigmoid function to obtain the final matching probability of the candidate and the job. The process can be represented as:

[0147]

[0148] wherein, , is the weight of the linear layer, , is the bias term, represents the matching degree of the candidate to the job , represents the matching degree of the job to the candidate . In order to combine the matching scores of the dual views, the average value thereof is taken as the final matching probability, and the formula is as follows:

[0149]

[0150] In addition, the method further comprises:

[0151] Model optimization, in the model optimization stage, a bilateral quadruple loss is adopted. The loss function contains a bilateral matching loss and a unilateral matching loss, which can model the bidirectional matching relationship between the candidate and the job, and effectively handle the unilateral matching situation.

[0152] On the basis of each positive sample matching , the corresponding negative samples and are generated, wherein and are randomly sampled negative samples. The training set is defined as:

[0153]

[0154] In this formula, represents a quadruple sample, and and represent the matching and non-matching sample sets, respectively.

[0155] The bilateral matching loss is used to evaluate the bidirectional matching relationship between the candidate and the job, and is defined as follows:

[0156]

[0157] To further utilize the one-sided matching information in negative samples, a one-sided matching loss is introduced , which is defined as follows:

[0158]

[0159] wherein, represents the score difference between one-way matching, helping the model better understand the asymmetric relationship existing in negative samples. The final loss function is defined by integrating the bidirectional matching loss and the one-way matching loss as follows:

[0160]

[0161] wherein, is the coefficient for balancing the two parts of the loss. By minimizing this loss function, the model can better optimize the parameters, thereby improving the performance of matching prediction.

[0162] In one specific embodiment, in order to illustrate the effectiveness of the method MUKRR proposed in this embodiment, a large number of experiments are carried out on the data set of the Aliyun human job matching programming competition provided by the large online recruitment platform of China, namely, ZhiLian recruitment. Compared with nine relatively novel methods, the effectiveness of the MUKRR method is verified. The indexes used in the experiment are Recall@5, Precision@5, NDCG@5 and MRR@5 respectively, and the experimental results are shown in Table 1. The experimental results show that the performance of the proposed MUKRR is significantly better than that of other methods, which again proves that the method proposed in this embodiment is effective.

[0163] Table 1

[0164]

[0165] The models compared in this embodiment have been published in top journals or conferences in the field of artificial intelligence, including SIGIR, WWW, RecSys, TMIS, and UAI, etc. The compared models are: BPRMF (Bayesian Personalized Ranking from Implicit Feedback), NCF (Neural Collaborative Filtering), LightGCN (Simplifying and Powering Graph Convolution Network for Recommendation), LFRR (Latent Factor Reciprocal Recommender), PJFNN (Person-Job Fit Neural Network), BPJFNN (Basic Person-Job Fit Neural Network), APJFNN (Ability-aware Person-Job Fit Neural Network), DPGNN (Dual-Perspective Graph Neural Network), KAERR (Knowledge-Aware Explainable Reciprocal Recommender System).

[0166] Through the above-mentioned two-way employment recommendation method based on career development trend modeling, the career development trend of the job seeker is modeled by using the dynamic memory enhanced U-Net architecture. DMU-Net realizes multi-level feature extraction through a time series feature pyramid network, thereby capturing seasonal changes in the shallow layer and identifying long-term trends in the deep layer. Then, a cross-layer attention mechanism is designed to fuse feature information of different scales, and a dynamic multi-key-value memory matrix is introduced to effectively reduce information loss in the down-sampling process. By introducing a multi-scale dual-view encoder and a cross-layer attention mechanism, the model can capture complex temporal dependencies at different scales and identify the potential direction of the job seeker's career development. This model not only integrates the current ability and behavior of the job seeker, but also focuses on the future growth path, making the recommendation results more forward-looking and accurate. By using a dynamic multi-key-value memory matrix, the model effectively retains and retrieves low-level feature information during the down-sampling and up-sampling processes, reducing the loss of feature information during level conversion, thereby ensuring the balance between high-level abstract features and low-level detailed features. Through dual-view meta-path aggregation, the weight and importance of the features are dynamically adjusted from the needs of the candidates and the employers, respectively, improving the understanding of the matching needs of both parties and generating more targeted matching features.

[0167] In addition, the terms "first", "second", etc. are used herein only to describe different instances, and do not imply or suggest relative importance or a number of the indicated technical features. Thus, the features defined with "first", "second" can include one or more of the features explicitly or implicitly. In the description of the embodiments of the present disclosure, the meaning of "a plurality of" is two or more, unless otherwise explicitly and specifically limited.

[0168] In the description of the present disclosure, the description referring to the terms "one embodiment", "some embodiments", "an example", "a specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present disclosure. In the present disclosure, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Also, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in the present disclosure.

[0169] Other embodiments of the present disclosure will be apparent to those skilled in the art upon consideration of the specification and practice of the application disclosed herein. The present application is intended to cover any variations, uses or adaptive changes of the present disclosure following the general principles of the present disclosure and including common knowledge or conventional techniques in the art not disclosed in the present disclosure. The specification and examples are only considered to be exemplary, and the true scope and spirit of the present disclosure are indicated by the appended claims.

Claims

1. A two-way employment recommendation method based on career development trend modeling, characterized by: The method includes: Construct a career development trend model; the career development trend model includes a knowledge graph embedding layer, a multi-scale dual-view meta-path encoding module, and an adaptive dual-view feature fusion layer; the multi-scale dual-view meta-path encoding module includes a DMU-Net unit and a Bi-LSTM unit; the DMU-Net unit includes a time series feature pyramid network, a cross-layer attention fusion module, and a dynamic multi-key-value memory matrix module; the adaptive dual-view feature fusion layer includes an adaptive feature fusion layer and a dual-view meta-path aggregation layer; Constructing the interaction data between job seekers and positions to build meta-path instances; The TransR model in the knowledge graph embedding layer is used to map the entities and relations of the meta-path instances into a low-dimensional vector space to obtain the meta-path embedding. The time series feature pyramid network is used to embed meta-paths and extract deep trend information and shallow seasonal features layer by layer through multi-layer pooling operations to generate multi-scale features; The cross-layer attention fusion module is used to effectively integrate the multi-scale features of different layers to obtain the initial cross-layer fusion features of the current layer; The dynamic multi-key memory matrix module is used to update the memory of the initial cross-layer fusion features to obtain the final cross-layer fusion features of the current layer; Bi-LSTM units are used to capture the dependencies between the candidate and position perspectives, generating a dual-perspective aggregate representation. Meta-path embeddings are then inverted as candidate feature inputs, and the inverted meta-path embeddings are used as position feature inputs, enabling the DMU-Net unit to obtain dual-perspective features. The dual-perspective aggregate representation includes candidate aggregate representation and position aggregate representation, and the dual-perspective features include candidate feature representation and position feature representation. The adaptive feature fusion layer uses an adaptive attention mechanism to weight the dual-view aggregate representation and dual-view features, effectively fusing multi-dimensional features to generate comprehensive feature representations of candidates and positions. Inputting the candidate comprehensive feature representation and the position comprehensive feature representation into the dual-view meta-path aggregation layer, setting different first attention weights for the candidate and the position respectively, to form a dual-view meta-path aggregation representation; wherein the dual-view meta-path aggregation representation includes the candidate meta-path aggregation representation and the position meta-path aggregation representation; The final matching probability between candidates and positions is calculated based on the dual-view meta-path aggregation representation, and the recommendation results are obtained based on the final matching probability.

2. The two-way employment recommendation method based on career development trend modeling according to claim 1 is characterized in that: The steps of constructing a meta-path instance of interaction data between job seekers and positions include: Build an interaction graph based on the interaction data between job seekers and positions, and build a knowledge graph by combining the job seekers' attribute information and the position's attribute information; Define different meta-path types between candidates and positions based on the requirements of the recommendation task; In the knowledge graph, the random walk algorithm is used to generate instances, and meta-path instances that meet the task requirements are extracted from the meta-path types.

3. The two-way employment recommendation method based on career development trend modeling according to claim 1 is characterized in that: The time series feature pyramid network is used to embed the meta-path into the step of extracting deep trend information and shallow seasonal features layer by layer through multi-layer pooling operations to generate multi-scale features, including: Embedding meta paths Input into the time series feature pyramid network, divide the data into L layers through the average pooling function, and use different depths in each layer to extract deep trend information and shallow seasonal features to generate multi-scale features. ;in, in, For the The output features of the layer, For the The output features of the layer, Represents an average pooling operation.

4. The two-way employment recommendation method based on career development trend modeling according to claim 3 is characterized in that: The steps of using the cross-layer attention fusion module to effectively integrate multi-layer input data at different levels to obtain the initial cross-layer fusion features of the current level include: The multi-scale features are input into the cross-layer attention fusion module. The correlation between time steps is calculated through the multi-head self-attention mechanism. Multiple attention heads are processed in parallel and then spliced ​​and integrated. Apply residual connection and layer normalization operations to the output of multi-head self-attention, and then enrich the feature expression through two linear transformations and ReLU activation in the feedforward network; Apply layer normalization and residual connection to obtain the attention feature set ; The attention features of the current layer are spliced ​​with the cross-layer fusion features of the previous layer, and the cross-layer attention mechanism is introduced again for fusion. The linear mapping restores the spliced ​​features to the original sequence length to obtain the initial cross-layer fusion features of the current layer. .

5. The two-way employment recommendation method based on career development trend modeling according to claim 4 is characterized in that: The step of using a dynamic multi-key memory matrix module to update the memory of the initial cross-layer fusion features to obtain the final cross-layer fusion features of the current layer includes: During the downsampling process of the DMU-Net unit, each layer Output features Update features are obtained through MLP processing: Where, is the linear transformation matrix, is the first bias term, To update features, is the sigmoid function; Through the global bond matrix corresponding to each layer and output features Calculate the second attention weight associated with the current level feature: Where, is the second attention weight, For the The first layer of the global bond matrix key vectors, For the The first layer of the global bond matrix key vectors; Update the private memory matrix using the second attention weight: Where, For the i The memory slot is in The value of the layer, For the i The memory slot is in The value of the layer; During the upsampling process of the DMU-Net unit, the initial cross-layer fusion features are multiplied by the corresponding global key matrix to calculate the third attention weight Based on the third attention weight, perform weighted summation on each memory slot of the private memory matrix to obtain the memory preference representation of the current level: The initial cross-layer fusion features and memory preference representation Add together to get the final cross-layer fusion features of this level .

6. The two-way employment recommendation method based on career development trend modeling according to claim 5 is characterized in that: The steps of using Bi-LSTM units to capture the dependencies between the candidate and position perspectives, generating a dual-perspective aggregate representation, and inverting the meta-path embedding as the candidate feature input. The inverted meta-path embedding is then used as the position feature input to enable the DMU-Net unit to obtain dual-perspective features, including: The Bi-LSTM unit is introduced to generate the forward hidden state representation and backward hidden state representation of candidates and positions respectively, and the forward hidden state representation and the backward hidden state representation are averaged to obtain the dual-view aggregate representation of candidates and positions: Where, For candidate aggregation, is the position aggregation representation, For Bi-LSTM units at time steps The forward hidden state on is represented as, For Bi-LSTM units at time steps The backward hidden state representation on T is the length of the meta-path; The meta-path embedding is used as the feature input of the candidate, and the inverted meta-path embedding is used as the feature input of the position, so that the DMU-Net unit can obtain dual-view features: in, represents the candidate feature representation, Indicates job characteristics. express The model processing process, is the reversed meta-path embedding.

7. The two-way employment recommendation method based on career development trend modeling according to claim 6 is characterized in that: The adaptive feature fusion layer uses an adaptive attention mechanism to weight the dual-view aggregate representation and dual-view features, effectively fusing multi-dimensional features to generate comprehensive candidate feature representations and position feature representations. The steps include: The dual-view aggregate representation and dual-view features extracted by the DMU-Net unit and the Bi-LSTM unit are stacked to form a candidate fusion feature representation and position fusion feature representation ; The fourth attention weight of the stacked features is calculated using the adaptive attention mechanism: in, is the fourth attention weight corresponding to the candidate, is the first trainable weight matrix, Fusion feature representation for the candidate, is the second bias term, is the fourth attention weight corresponding to the position, is the second trainable weight matrix, is the position fusion feature representation, is the third bias term; Use the fourth attention weight to perform weighted summation on the candidate feature representation and the position feature representation to obtain the fused candidate comprehensive feature representation and comprehensive characteristics of positions : Where, The weight assigned by the attention mechanism to the k-th feature type of the candidate, is the feature representation of the candidate’s k-th feature type, The weight assigned by the attention mechanism to the k-th feature type of the position, is the feature representation of the k-th feature type of the position.

8. The two-way employment recommendation method based on career development trend modeling according to claim 7 is characterized in that: The step of inputting the candidate comprehensive feature representation and the position comprehensive feature representation into the dual-view meta-path aggregation layer, setting different first attention weights for the candidate and the position respectively, and forming the dual-view meta-path aggregation representation includes: Assign the first attention weight to the fused candidate comprehensive feature representation and position comprehensive feature representation; The meta-path representation is weighted summed using the aggregated attention weights to obtain the dual-view meta-path aggregate representation: in, is the candidate meta-path aggregation representation, is the position meta-path aggregation representation, For the The candidate attention weights of each meta-path, For the The position attention weight of each meta-path, For the The candidate comprehensive feature representation of the element path, For the Comprehensive feature representation of positions of meta-paths.

9. The two-way employment recommendation method based on career development trend modeling according to claim 8 is characterized in that: The steps of calculating the final matching probability between the candidate and the position based on the dual-view meta-path aggregation representation and obtaining the recommendation result based on the final matching probability include: The matching scores are normalized by the Sigmoid function to obtain the final matching probability between the candidate and the position. The process can be expressed as: in, is the first weight of the linear layer, is the second weight of the linear layer, is the fourth bias term, is the fifth bias term, Indicates candidate About the position The matching degree, Indicate position To the candidates degree of matching; The average value is used as the final matching probability, and the formula is as follows: Where, is the final matching probability.

10. The two-way employment recommendation method based on career development trend modeling according to claim 9 is characterized in that: The total loss function of the career development trend model is: in, is the total loss function, is the bilateral matching loss, is the one-sided matching loss, is the trade-off coefficient, is the training data set representation, Score one-way interactions with job seekers, Score one-way interactions for work, represents the score difference between one-way matches.

Images