Bidirectional employment recommendation method based on occupational development trend modeling

Abstract

The invention belongs to the technical field of data analysis. The invention provides a two-way employment recommendation method based on occupational development trend modeling. According to the embodiment of the invention, the DMU-Net architecture is used for modeling the vocational development trend of the job seeker. DMU-Net realizes multi-level feature extraction through a time sequence feature pyramid network, so that seasonal changes are captured in a shallow layer, a long-term trend is deeply identified, then a cross-layer attention mechanism is designed to fuse feature information of different scales, a dynamic multi-key value memory matrix is introduced, and information loss in a down-sampling process is effectively reduced. In addition, through the design of a double-view-angle encoder and an adaptive feature fusion layer, bidirectional matching of demands of job seekers and recruiters is realized, and more accurate employment recommendation is provided.

Description

Technical Field

[0001] The disclosed embodiments relate to the technical field of data analysis, and in particular to a two-way employment recommendation method based on career development trend modeling. Background Art

[0002] The employment recommendation system is a technology that has gradually emerged with the development of the Internet and big data technology. It uses intelligent algorithms and a large amount of behavioral data of job seekers to recommend suitable job opportunities to users. In the modern job market, the matching between job seekers and recruiters has become increasingly complex. The traditional recruitment process is not only time-consuming, but also prone to missing potential suitable talents. Therefore, how to effectively match the right talents to the right positions has become the focus of attention of companies and researchers. The employment recommendation system helps companies find qualified candidates more efficiently by analyzing job seekers' skills, interests, work experience and other information, combined with the specific requirements of the position, while also providing job seekers with more personalized job recommendations.

[0003] Current job recommendation algorithms mainly rely on two types of information resources: one is the applicant's historical application records, and the other is the text information of resumes and job descriptions. For historical application records, collaborative filtering methods are usually used to mine potential matches from the application relationship between applicants and positions. In the processing of text information, natural language processing technology is mainly used to extract key features from applicants and job descriptions to achieve more accurate matching recommendations. In addition, some advanced hybrid models also combine interactive data with 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 automated recruitment systems have made significant progress in job recommendation, current recommendation systems still face some problems. They are mainly based on job seekers' historical behaviors and current abilities. This reliance on static information limits the system's ability to capture job seekers' long-term career goals. As a result, the factor of "future career development trends" is ignored, resulting in recommendation results that cannot fully support job seekers' career growth.

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

[0006] It should be noted that this section is intended to provide background or context for the technical solutions of the present disclosure stated in the claims. The description herein is not admitted to be prior art by virtue of being included in this section. Summary of the invention

[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 overcoming one or more problems caused by the limitations and defects of related technologies at least to a certain extent.

[0008] According to an embodiment of the present disclosure, a two-way employment recommendation method based on career development trend modeling is provided, the method comprising: Construct a career development trend model; wherein 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, and 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; Construct the interaction data between job seekers and positions to construct a meta-path instance; 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 the meta-path into a multi-layer pooling operation to extract deep trend information and shallow seasonal features layer by layer to generate multi-scale features. The cross-layer attention fusion module is used to effectively integrate the multi-scale features of different levels to obtain the initial cross-layer fusion features of the current level; 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 level; Bi-LSTM units are used to capture the dependencies under the perspectives of candidates and positions respectively, and a dual-perspective aggregate representation is generated. The meta-path embedding is used as the candidate feature input for inversion, and the inverted meta-path embedding is used as the position feature input, so that the DMU-Net unit can obtain dual-perspective features; wherein the dual-perspective aggregate representation includes the candidate aggregate representation and the position aggregate representation, and the dual-perspective features include the candidate feature representation and the position feature representation; The adaptive feature fusion layer uses the adaptive attention mechanism to weight the dual-view aggregate representation and the dual-view features, effectively fusing the multi-dimensional features to generate the candidate comprehensive feature representation and the position comprehensive feature representation. 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.

[0009] Furthermore, the step of constructing a meta-path instance by constructing the interaction data between job seekers and positions includes: Build an interaction graph based on the interaction data between job seekers and positions, and build a knowledge graph by combining the attribute information of job seekers and the attribute information of positions; Define different meta-path types between candidates and positions according to 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.

[0010] Furthermore, 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 the meta path 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,

[0011] in, For the The output features of the layer, For the l- The output features of layer 1, Represents an average pooling operation.

[0012] Furthermore, the step of effectively integrating multi-layer input data of different levels by using a cross-layer attention fusion module to obtain an initial cross-layer fusion feature of the current level includes: The multi-scale features are input into the cross-layer attention fusion module, and 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; Residual connections and layer normalization operations are applied to the output of multi-head self-attention, and then the feature expression is enriched through two linear transformations and ReLU activation of the feedforward network; Apply layer normalization and residual connection to obtain the attention feature set ; The attention features of the current layer are concatenated with the cross-layer fusion features of the previous layer, and the cross-layer attention mechanism is introduced again for fusion. The concatenated features are linearly mapped to restore the original sequence length to obtain the initial cross-layer fusion features of the current layer. .

[0013] Further, the step of using a dynamic multi-key value memory matrix module to memorize and update the initial cross-layer fusion features to obtain the final cross-layer fusion features of the current layer includes: In the process of downsampling the DMU-Net unit, each layer Output characteristics The updated features are obtained through MLP processing:

[0014] In the formula, is the linear transformation matrix, is the first bias term, To update the features, is the sigmoid function; Through the global key matrix corresponding to each layer and output features Calculate the second attention weight associated with the current level feature:

[0015] In the formula, is the second attention weight, For the The global key matrix of the layer key vectors, For the The global key matrix of the layer key vectors; Update the private memory matrix using the second attention weight:

[0016] In the formula, For the i The memory slot is l The value of the layer, For the i The memory slot is l- The value of layer 1; 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, each memory slot of the private memory matrix is ​​weighted and summed to obtain the memory preference representation of the current level:

[0017] The initial cross-layer fusion features and memory preference representation Add together to get the final cross-layer fusion features of this level .

[0018] Furthermore, the Bi-LSTM unit is used to capture the dependency relationship under the perspective of candidates and positions respectively, and a dual-perspective aggregate representation is generated. The meta-path embedding is used as the candidate feature input for inversion, and the inverted meta-path embedding is used as the position feature input, so that the DMU-Net unit can obtain the dual-perspective feature. The steps include: 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:

[0019] In the formula, is the candidate aggregation representation, is the position aggregation representation, For the Bi-LSTM unit at time step The forward hidden state on is represented by For the Bi-LSTM unit at time step The backward hidden state on represents, 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 obtains dual-view features:

[0020] in, represents the candidate characteristics representation, Indicates job characteristics. Represents the processing process of the DMU-Net model.

[0021] Furthermore, the adaptive feature fusion layer weights the dual-view aggregate representation and the dual-view features through an adaptive attention mechanism, and effectively fuses the multi-dimensional features to generate the candidate comprehensive feature representation and the position comprehensive feature representation, including: 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:

[0022] in, is the fourth attention weight corresponding to the candidate, is the first trainable weight matrix, Fusion feature representation for candidates, 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 :

[0023] In the formula, The weight assigned by the attention mechanism to the kth 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 kth feature type of the position, is the feature representation of the kth feature type of the position.

[0024] Furthermore, 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:

[0025] 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 l The candidate comprehensive feature representation of the element path, For the l The comprehensive characteristics of the position of the meta-path are represented.

[0026] Furthermore, the step of calculating the final matching probability between the candidate and the position according to the dual-view meta-path aggregation representation and obtaining the recommendation result according to the final matching probability includes: 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:

[0027] 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 the candidate About Positions The matching degree, Indicate job title For candidates The matching degree; The average value is used as the final matching probability, and the formula is as follows:

[0028] In the formula, is the final matching probability.

[0029] Furthermore, the total loss function of the career development trend model is:

[0030]

[0031]

[0032] 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 for job seekers, Score one-way interactions for work, represents the score difference between one-way matches.

[0033] The technical solution provided by the embodiments of the present disclosure may have the following beneficial effects: In the embodiment 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 the dynamic memory enhanced U-Net (DMU-Net) architecture. DMU-Net realizes multi-level feature extraction through the time series feature pyramid network, thereby capturing seasonal changes at the shallow level and identifying long-term trends at the deep level. Then, a cross-layer attention mechanism is designed to fuse feature information of different scales, and a dynamic multi-key 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 dependencies at different scales and identify the potential direction of job seekers' career development. The model not only integrates the current abilities and behaviors of job seekers, but also pays more attention to their future growth path, so that the recommendation results are more forward-looking and accurate. By using a dynamic multi-key memory matrix, the model effectively retains and retrieves low-level feature information during downsampling and upsampling, reducing the loss of feature information in hierarchical conversion, thereby ensuring the balance between high-level abstract features and underlying detailed features. Through dual-perspective meta-path aggregation, we dynamically adjust the weights and importance of features based on the needs of candidates and recruiters, improve our ability to understand the matching needs of both parties, and generate more targeted matching features. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings herein are incorporated into the specification and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the specification are used to explain the principles of the present disclosure. Obviously, the accompanying drawings described below are only some embodiments of the present disclosure, and for ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without creative work.

[0035] Figure 1 A step diagram showing a two-way employment recommendation method based on career development trend modeling in an exemplary embodiment of the present disclosure; Figure 2 A framework diagram showing a career development trend model in an exemplary embodiment of the present disclosure; Figure 3 A framework diagram of a DMU-Net unit in an exemplary embodiment of the present disclosure is shown; Figure 4 A flow chart showing a memory update operation of a DMU-Net unit in an exemplary embodiment of the present disclosure is shown; Figure 5 A flow chart showing a memory reading operation of a DMU-Net unit in an exemplary embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0036] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the disclosure will be more comprehensive and complete and to fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0037] In addition, the accompanying drawings are only schematic illustrations of the embodiments of the present disclosure and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated descriptions will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0038] This example embodiment provides a two-way employment recommendation method based on career development trend modeling. Figure 1 As shown in , the two-way employment recommendation method based on career development trend modeling may include: steps S101 to S110.

[0039] Step S101: constructing a career development trend model; wherein 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, and 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; Step S102: constructing a meta-path instance of interaction data between job seekers and positions; Step S103: Mapping entities and relations of meta-path instances to low-dimensional vector space using the TransR model in the knowledge graph embedding layer to obtain meta-path embedding; Step S104: using the time series feature pyramid network to embed the meta-path into a multi-layer pooling operation to extract deep trend information and shallow seasonal features layer by layer, and generate multi-scale features; Step S105: using a cross-layer attention fusion module to effectively integrate multi-scale features at different levels to obtain initial cross-layer fusion features at the current level; Step S106: using a dynamic multi-key value memory matrix module to perform memory update on the initial cross-layer fusion features to obtain the final cross-layer fusion features of the current layer; Step S107: using Bi-LSTM units to capture the dependencies under the perspectives of candidates and positions respectively, generating dual-perspective aggregate representations, and inverting the meta-path embedding as candidate feature input, and using the inverted meta-path embedding as position feature input, so that the DMU-Net unit can obtain dual-perspective features; wherein 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; Step S108: The adaptive feature fusion layer weights the dual-view aggregate representation and the dual-view features through an adaptive attention mechanism, effectively fusing the multi-dimensional features to generate a candidate comprehensive feature representation and a position comprehensive feature representation; Step S109: 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 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; Step S110: Calculate the final matching probability between the candidate and the position according to the dual-view meta-path aggregation representation, and obtain the recommendation result according to the final matching probability.

[0040] Through the above-mentioned two-way employment recommendation method based on career development trend modeling, the career development trend of job seekers is modeled using the dynamic memory enhanced U-Net architecture. DMU-Net realizes multi-level feature extraction through the time series feature pyramid network, thereby capturing seasonal changes at the shallow level and identifying long-term trends at the deep level. Then a cross-layer attention mechanism is designed to fuse feature information of different scales, and a dynamic multi-key memory matrix is ​​introduced to effectively reduce information loss during downsampling. 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 job seekers' career development. The model not only integrates the current abilities and behaviors of job seekers, but also pays more attention to their future growth path, making the recommendation results more forward-looking and accurate. By utilizing the dynamic multi-key memory matrix, the model effectively retains and retrieves low-level feature information during downsampling and upsampling, reducing the loss of feature information in hierarchical conversion, thereby ensuring the balance between high-level abstract features and underlying detailed features. Through dual-perspective meta-path aggregation, we dynamically adjust the weights and importance of features based on the needs of candidates and recruiters, improve our ability to understand the matching needs of both parties, and generate more targeted matching features.

[0041] Next, we will refer to Figures 1 to 5 The various steps of the above-mentioned two-way employment recommendation method based on career development trend modeling in this example implementation are described in more detail.

[0042] 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, and the DMU-Net unit includes a time series feature pyramid network, a cross-layer attention fusion module and a dynamic multi-key memory matrix module; the adaptive dual-view feature fusion layer includes an adaptive feature fusion layer and a dual-view meta-path aggregation layer. Among them, the multi-scale dual-view meta-path encoding module: extracts multi-level career development trend features through 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 DMU-Net and Bi-LSTM through the attention mechanism, giving higher weights to important features, and highlighting the different needs of candidates and positions on the meta-path.

[0043] like Figure 2 Shown is the framework diagram of the career development trend model.

[0044] like Figure 3 The figure shows the framework diagram of the DMU-Net unit.

[0045] In step S102, first, an interaction graph is constructed based on the interaction data between job seekers and recruiters, and a complete knowledge graph is established by combining the attribute information of both parties. On this basis, different meta-path types between candidates and positions are defined according to the requirements of the recommendation task. For example, meta-path types may include "candidate-possessing skills-position required skills-position" or "candidate-possessing education-required education-position", etc. These paths are used to capture the potential relationship between candidates and positions. In the knowledge graph, instances are generated using a random walk algorithm to extract meta-path instances that meet the task requirements from predefined meta-path types.

[0046] 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.

[0047] 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 .

[0048] 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, separates long-term trends in the deep layer, and finally generates multi-level input data :

[0049] in, For the The pooling result of the layer, Represents an average pooling operation.

[0050] In step S105, DMU-Net uses the cross-layer attention mechanism to integrate features of different levels. The cross-layer attention fusion module captures the global dependencies between each time step through a multi-head self-attention mechanism. Then, the obtained attention features of the current layer are concatenated with the cross-layer fusion features of the previous layer, and the fusion is strengthened again through the cross-layer attention mechanism to obtain the cross-layer fusion features of the current layer. .

[0051] Specific, multi-scale features The correlation between time steps is calculated through the multi-head self-attention mechanism, and multiple attention heads are processed in parallel and then spliced ​​and integrated. Subsequently, residual connection and layer normalization operations are applied to the output of the multi-head self-attention, 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 Next, the attention features of the current layer are concatenated 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 final 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 layer. (i.e., the initial cross-layer fusion features).

[0052] In step S106, the downsampling process of DMU-Net may lead to the loss of low-level features, so a dynamic multi-key memory matrix is ​​introduced, including two parts: global and private memory matrices. The global matrix is ​​used to generate dynamic weights through the attention mechanism, and the private memory matrix stores features independently for each user, thereby retaining user preferences at different levels. The private memory matrix can perform memory updates during the downsampling process, maintain the short-term and long-term features of the candidates, and perform memory reading during the upsampling process, avoiding information loss caused by downsampling.

[0053] Specifically, in the memory update process of DMU-Net downsampling, the downsampled output of each layer is first used 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), and the generated update features can capture the key information of the layer, as follows:

[0054] in, is the linear transformation matrix, is the bias term.

[0055] Next, each layer has a corresponding global bond matrix , through which the matrix is ​​calculated and the current level feature The relevant attention weights are as follows:

[0056] in, Indicates The global key matrix of the layer A key vector.

[0057] Finally, the private memory matrix updates the memory slots layer by layer before the downsampling average pooling of DMU-Net as follows:

[0058] in, Indicates The memory slot is The value of the layer.

[0059] Through this weighted update method, the model can dynamically retain key feature information between levels.

[0060] like Figure 4 The figure shows the memory update operation flow chart of the DMU-Net unit.

[0061] Memory retrieval, in the DMU-Net upsampling process, the model first fuses the cross-layer features The corresponding global key matrix Multiply and calculate the attention weight ; Based on these weights, each memory slot of the private memory matrix is ​​weighted and summed to obtain the memory preference representation of the current level, as follows:

[0062] in, Represents a private memory matrix Middle i The contents of the slots, and Controls the contribution of this slot in the current level preference representation.

[0063] 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).

[0064] like Figure 5 The figure shows the memory reading operation flow chart of the DMU-Net unit.

[0065] In step S107, 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 perspectives of candidates and positions respectively, and generate an aggregated representation of the dual perspectives. and At the same time, the original meta-path embedding is used as the candidate feature input, and the inverted meta-path embedding is used as the position feature input, so that DMU-Net can obtain dual-view features and .

[0066] Specifically, to make up for the shortcomings of DMU-Net in sequence order modeling, the model introduces Bi-LSTM, which captures long-distance dependencies through its bidirectional structure and generates forward and backward hidden state representations of candidates and positions respectively. The average of the bidirectional hidden states forms the sequential dependency representation of candidates and positions:

[0067] in, and Respectively represent candidates and Position Representation on a meta-path. represents the length of the meta-path, and Respectively represent Bi-LSTM at time step The forward and backward hidden states on .

[0068] The model uses the original meta-path embedding as the candidate feature input and the inverted meta-path embedding as the position feature input, enabling DMU-Net to obtain dual-view information. This symmetrical design enhances the matching feature expression between candidates and positions:

[0069] in, represents the candidate characteristics representation, Indicates job characteristics. Represents the processing process of the DMU-Net model.

[0070] In step S108, the adaptive feature fusion module in the adaptive dual-view feature fusion layer is used to input the candidate and position features extracted by DMU-Net and Bi-LSTM into the adaptive dual-stream fusion layer. The adaptive attention mechanism is used to weight features from different sources and effectively fuse multi-dimensional features to generate comprehensive feature representations of candidates and positions. and This step ensures that the model can integrate feature information from different sources and enhance the comprehensiveness and accuracy of the matching.

[0071] Specifically, the candidate and position features extracted by DMU-Net and Bi-LSTM are stacked separately to form a fused feature representation and , ensuring that the two features remain independent during the fusion process. The weights of the stacked features are calculated using the adaptive attention mechanism and :

[0072] in, , is a trainable weight matrix, , is the bias term. is the sigmoid function, which is used to convert the attention score into a weight between 0 and 1.

[0073] Then, the attention weights are used to perform a weighted summation of the two different types of features to obtain the fused candidate feature representation and position feature representation:

[0074] In the above formula, and are the weights assigned by the attention mechanism to the k-th feature type of users and items, respectively, reflecting the relative importance of DMU-Net and Bi-LSTM features in the current task. and They represent the feature representation of the kth feature type of users and items respectively.

[0075] In step S109, the dual-view meta-path aggregation module in the adaptive dual-view feature fusion layer is used to set different attention weights for the candidate and the recruiter respectively when fusing features to form a dual-view meta-path aggregation representation. and This aggregation layer focuses on capturing the respective needs of candidates and recruiters, highlighting the information that is critical to matching while suppressing irrelevant features.

[0076] Specifically, aggregate attention weights are assigned to the features of candidates and positions to reflect the different needs of candidates and recruiters on the same meta-path. These weights are calculated based on the candidate and position feature representations to form an aggregate feature representation:

[0077] in, and Respectively represent the candidate and position representation after aggregating the two perspectives, and For the The attention weight of each meta-path is and as input.

[0078] In step S110, matching probability prediction is to input the candidate and position features after dual-view aggregation into the linear layer to calculate their matching probability. The matching score is normalized by the Sigmoid function to obtain the final matching probability of the candidate and the position. The process can be expressed as:

[0079] in, , are the weights of the linear layer, , is the bias term, Indicates the candidate About Positions The matching degree, Indicate job title For candidates In order to combine the matching scores of the two views, their average value is taken as the final matching probability, and the formula is as follows:

[0080] Additionally, the method further comprises: Model Optimization,In the model optimization stage, bilateral quadruple loss is used.,This loss function contains bilateral matching loss and unilateral matching,loss, which can simultaneously model the bidirectional matching relationship,between candidates and positions and effectively handle unilateral matching,cases.

[0081] In each positive sample matching Based on this, generate the corresponding negative samples and ,in and is a randomly sampled negative sample. The training set is defined as:

[0082] In this formula, represents a four-tuple sample, and and Represent the sets of matching and non-matching samples respectively.

[0083] Bilateral matching loss It is used to evaluate the two-way matching relationship between candidates and positions, and its definition is as follows:

[0084] In order to further utilize the one-sided matching information in negative samples, the one-sided matching loss is introduced , defined as follows:

[0085] in, Represents the score difference between one-way matches, helping the model better understand the asymmetric relationship in negative samples. The final loss function is defined by combining the two-way matching loss and the one-way matching loss:

[0086] in, is a coefficient that weighs the two losses. By minimizing this loss function, the model can better optimize the parameters and thus improve the performance of matching prediction.

[0087] In a specific embodiment, in order to illustrate the effectiveness of the method MUKRR proposed in this embodiment, this embodiment conducted a large number of experiments on the dataset of the Alibaba Cloud Job Matching Programming Competition, which is provided by Zhaopin, a large online recruitment platform in China. It was compared with 9 relatively new methods to verify the effectiveness of the MUKRR method. The indicators used in the experiment are Recall@5, Precision@5, NDCG@5 and MRR@5, and the experimental results are shown in Table 1. The experimental results show that the performance of the proposed MUKRR is significantly better than other methods, which once again proves that the method proposed in this embodiment is effective.

[0088] Table 1

[0089] 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. Comparative models: 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-JobFit Neural Network), DPGNN (Dual-Perspective Graph Neural Network), KAERR (Knowledge-Aware Explainable Reciprocal Recommender System).

[0090] Through the above-mentioned two-way employment recommendation method based on career development trend modeling, the career development trend of job seekers is modeled using the dynamic memory enhanced U-Net architecture. DMU-Net realizes multi-level feature extraction through the time series feature pyramid network, thereby capturing seasonal changes at the shallow level and identifying long-term trends at the deep level. Then a cross-layer attention mechanism is designed to fuse feature information of different scales, and a dynamic multi-key memory matrix is ​​introduced to effectively reduce information loss during downsampling. 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 job seekers' career development. The model not only integrates the current abilities and behaviors of job seekers, but also pays more attention to their future growth path, making the recommendation results more forward-looking and accurate. By utilizing the dynamic multi-key memory matrix, the model effectively retains and retrieves low-level feature information during downsampling and upsampling, reducing the loss of feature information in hierarchical conversion, thereby ensuring the balance between high-level abstract features and underlying detailed features. Through dual-perspective meta-path aggregation, we dynamically adjust the weights and importance of features based on the needs of candidates and recruiters, improve our ability to understand the matching needs of both parties, and generate more targeted matching features.

[0091] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present disclosure, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0092] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.

[0093] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any modification, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, 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 in that: The method includes: Construct a career development trend model; wherein 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, and 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; Construct the interaction data between job seekers and positions to construct a meta-path instance; 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 the meta-path into a multi-layer pooling operation to extract deep trend information and shallow seasonal features layer by layer to generate multi-scale features. The cross-layer attention fusion module is used to effectively integrate the multi-scale features of different levels to obtain the initial cross-layer fusion features of the current level; 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 level; Bi-LSTM units are used to capture the dependencies under the perspectives of candidates and positions respectively, and a dual-perspective aggregate representation is generated. The meta-path embedding is used as the candidate feature input for inversion, and the inverted meta-path embedding is used as the position feature input, so that the DMU-Net unit can obtain dual-perspective features; wherein the dual-perspective aggregate representation includes the candidate aggregate representation and the position aggregate representation, and the dual-perspective features include the candidate feature representation and the position feature representation; The adaptive feature fusion layer uses the adaptive attention mechanism to weight the dual-view aggregate representation and the dual-view features, effectively fusing the multi-dimensional features to generate the candidate comprehensive feature representation and the position comprehensive feature representation. 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 attribute information of job seekers and the attribute information of positions; Define different meta-path types between candidates and positions according to 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 the meta path 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 l- The output features of layer 1, 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 the multi-layer input data of 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, and 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; Residual connections and layer normalization operations are applied to the output of multi-head self-attention, and then the feature expression is enriched through two linear transformations and ReLU activation of the feedforward network; Apply layer normalization and residual connection to obtain the attention feature set ; The attention features of the current layer are concatenated with the cross-layer fusion features of the previous layer, and the cross-layer attention mechanism is introduced again for fusion. The concatenated features are linearly mapped to restore 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 value memory matrix module to update the initial cross-layer fusion feature memory to obtain the final cross-layer fusion feature of the current level includes: In the process of downsampling the DMU-Net unit, each layer Output characteristics The updated features are obtained through MLP processing: In the formula, is the linear transformation matrix, is the first bias term, To update the features, is the sigmoid function; Through the global key matrix corresponding to each layer and output features Calculate the second attention weight associated with the current level feature: In the formula, is the second attention weight, For the The global key matrix of the layer key vectors, For the The global key matrix of the layer key vector; Update the private memory matrix using the second attention weight: In the formula, For the i The memory slot is l The value of the layer, For the i The memory slot is l- The value of layer 1; 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, each memory slot of the private memory matrix is ​​weighted and summed 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 feature 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 under the perspectives of candidates and positions respectively, generating dual-perspective aggregate representations, and inverting the meta-path embedding as the candidate feature input, and using the inverted meta-path embedding as the position feature input so that the DMU-Net unit can obtain dual-perspective features include: 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: In the formula, is the candidate aggregation representation, is the position aggregation representation, For the Bi-LSTM unit at time step The forward hidden state on is represented as, For the Bi-LSTM unit at time step The backward hidden state on is represented by 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 obtains dual-view features: in, represents the candidate characteristics representation, Indicates job characteristics. Represents the processing process of the DMU-Net model.

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 the adaptive attention mechanism to weight the dual-view aggregate representation and the dual-view features, effectively fusing the multi-dimensional features to generate the candidate comprehensive feature representation and the position comprehensive feature representation, including: 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 candidates, 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 : In the formula, The weight assigned by the attention mechanism to the kth 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 kth feature type of the position, is the feature representation of the kth 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 l The candidate comprehensive feature representation of the element path, For the l The comprehensive characteristics of the position of the meta-path are represented.

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 according to the dual-view meta-path aggregation representation and obtaining the recommendation result according to 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 the candidate About Positions The matching degree, Indicate job title For candidates The matching degree; The average value is used as the final matching probability, and the formula is as follows: In the formula, 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 for job seekers, Score one-way interactions for work, represents the score difference between one-way matches.

Images