Employment recommendation method, system, device and medium based on ensemble learning model

Abstract

The application discloses an employment recommendation method, system, device and medium based on an ensemble learning model, comprising the following steps: obtaining a user u and each recommendation item c, and performing vectorization coding on the features of the user u and the features of each recommendation item c; inputting the feature vector of the user u, the interaction record vector of the user u, and the feature vector of the user u and each recommendation item c into a trained ensemble learning model to obtain the scores of each recommendation item, and completing employment recommendation based on the ensemble learning model, wherein the ensemble learning model is composed of a similarity information component, a global interaction information component and a feature matching information component, and the method, system, device and medium can improve the generalization ability and scalability of the model, thereby improving the accuracy of the employment recommendation.

Description

Technical Field

[0001] This invention belongs to the field of graduate employment information recommendation technology, and relates to an employment recommendation method, system, device and medium based on an ensemble learning model. Background Technology

[0002] In this era of information overload, college graduates often spend a significant amount of time sifting through suitable job information, frequently becoming overwhelmed and confused. Traditional graduate career guidance services often suffer from limited resources of academic advisors, who are typically only available at specific times and locations, leading to delays and neglect of individual student needs. In the context of the information age, leveraging recommendation technology to provide job recommendations for graduates, facilitating intelligent career guidance, and helping users efficiently access information of interest are of great importance in providing more timely and personalized services.

[0003] Interaction records between users and employers in employment information are typically sparse, making traditional collaborative filtering algorithms based on these records less effective for job recommendations. Furthermore, data scarcity hinders the training and transfer of deep learning models widely used in e-commerce recommendation systems. Therefore, current graduate job recommendations largely employ a hybrid approach combining content-based and collaborative filtering methods. Jin Lianxu et al. proposed the Sensitive Personal Rank algorithm, which constructs a user-item bipartite graph using historical graduate data. A random walk strategy is used to first calculate the employer's interest sensitivity to current graduates, then incorporates this sensitivity into the graduate similarity calculation, finally recommending employers of past graduates similar to the current user. Liu et al. introduced the concept of student profiles, generating profiles based on the current user's on-campus data, calculating similarity between profiles, and then recommending employers based on already graduated and employed students. Wang Long et al. identified similar students by comparing their marginal features, then filled in missing interaction data using a weighted sum of the ratings of these similar students. Zhou et al. used a clustering algorithm to divide students into different groups, generated an interaction matrix using the historical employment records of the groups, and calculated the final recommendation score by combining feature preferences and preferences for geographical location. Their experiments showed that their method outperformed traditional collaborative filtering in terms of hit ratio and mean reciprocal rank, indicating that their method overcomes the difficulties caused by the sparsity of interaction records to some extent.

[0004] Existing research on graduate employment recommendation solutions shares the common advantages of simple algorithms, ease of engineering, and alignment with accumulated experience in general employment guidance. However, its shortcomings lie in the fact that most existing works build models based on a single perspective, making it difficult to adapt to the diverse distribution characteristics of data in specific application scenarios. This results in poor generalization and scalability of the models, and consequently, lower accuracy in employment recommendations. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a job recommendation method, system, device, and medium based on an ensemble learning model. This method, system, device, and medium can improve the generalization ability and scalability of the model, thereby improving the accuracy of job recommendations.

[0006] To achieve the above objectives, this invention discloses a job recommendation method based on an ensemble learning model, comprising:

[0007] Obtain user u and each recommended item c, and vectorize the features of user u and each recommended item c.

[0008] The feature vector of user u, the existing interaction record vector of user u, and the feature vectors of user u and each recommended item c are input into the trained ensemble learning model to obtain the score of each recommended item, thus completing the job recommendation based on the ensemble learning model. The ensemble learning model consists of a similarity information component, a global interaction information component, and a feature matching information component.

[0009] The specific process of inputting the feature vector of user u, the existing interaction record vector of user u, and the feature vectors of user u and each recommended item c into the trained ensemble learning model to obtain the score of each recommended item is as follows:

[0010] The feature vector of user u is input into the trained similarity information component to obtain the recommendation score S(u,c);

[0011] Input the existing interaction record vector of user u into the trained global interaction information component to obtain the recommendation score G(u,c);

[0012] Input the feature vectors of user u and each recommended item c into the trained feature matching information component to obtain the recommendation score M(u,c);

[0013] The recommendation score for item c for user u is calculated as follows:

[0014]

[0015] Where, α i(i = 1, 2, 3) represent the weights of the similarity information component, the global interaction information component, and the feature matching information component, respectively.

[0016] The specific process of vectorizing the features of user u and the features of each recommended item c is as follows:

[0017] Let user u have m-dimensional features, and user u be represented as... in, Let represent the value of the i-th feature of user u. Suppose that the recommended item c has n-dimensional features, and the recommended item c is represented as... Construct the pooled user vector u of recommendation item c c .

[0018] The pooled user vector u for constructing recommendation item c is... c The specific process is as follows:

[0019] For recommended item c, let U be the set of users whose interaction records contain the recommended item. c Then the pooled user vector u of recommended project c is... c for:

[0020]

[0021] For user u i The recommended score S(u) for project c i c) is:

[0022]

[0023] in, For user u j The collection of interaction records, sim(u i ,u j ) for user u i With user u j Similarity, sim(u i ,u c ) for user u i With user u c The similarity is calculated using ε1 and ε2 as weighting coefficients.

[0024] The optimization objective during the training process of the global interactive information component is:

[0025]

[0026] Where U is the set of all users, I u Let be the vector representation of user u's interaction records, θ be the trainable parameters of the autoencoder, and h(I) be the vector representation of user u's interaction records. u;θ) is the output of the autoencoder, λ is the regularization parameter, V is the transformation matrix from the input layer to the hidden layer, and W is the transformation matrix from the hidden layer to the output layer;

[0027] The loss function during the training of the global interactive information component is:

[0028]

[0029] Among them, I v This represents the set of all interaction record vectors used as the validation set. This represents the actual value of the interaction record between user u and project c in this interaction record. This is the predicted value for this interaction.

[0030] The feature matching information component includes an input layer, an embedding layer, an attention scoring layer, a connection layer, and a scoring layer connected in sequence.

[0031] Existing job application records are described as tuples of (User ID, Job ID), and are considered positive examples, thus being labeled as positive examples. Negative example tuples are generated from jobs the user has not applied for, through random negative sampling, and are also labeled as negative examples. The negative sampling ratio is set to 40% of the number of positive examples. The user's representation vector is then concatenated within each tuple. and project vectors Used as positive and negative samples for training;

[0032] The input layer takes positive and negative samples as input, the scoring layer outputs a score representing the degree of matching between the user and the job, i.e., the recommendation score, the embedding layer is used to transform the encoded high-dimensional sparse feature vector into a low-dimensional dense vector, and the attention scoring layer uses a cross-attention mechanism to give attention scores on the user side and the recommendation item side respectively.

[0033] This invention discloses a job recommendation system based on an ensemble learning model, comprising:

[0034] The encoding module is used to obtain user u and each recommended item c, and to vectorize and encode the features of user u and each recommended item c.

[0035] The recommendation module is used to input the feature vector of user u, the existing interaction record vector of user u, and the feature vectors of user u and each recommendation item c into the trained ensemble learning model to obtain the score of each recommendation item and complete the job recommendation based on the ensemble learning model. The ensemble learning model consists of a similarity information component, a global interaction information component, and a feature matching information component.

[0036] The present invention discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the job recommendation method based on an ensemble learning model.

[0037] The present invention discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the job recommendation method based on an ensemble learning model.

[0038] The present invention has the following beneficial effects:

[0039] In practical operation, the job recommendation method, system, device, and medium based on the ensemble learning model described in this invention perform job recommendations based on the trained ensemble learning model. The ensemble learning model consists of a similarity information component, a global interaction information component, and a feature matching information component to improve the model's generalization ability and scalability, thereby improving the accuracy of job recommendations.

[0040] Furthermore, the similarity information component takes into account the similarity between the target user and the user group, which can alleviate the problem of noise interference to some extent.

[0041] Furthermore, during the training of the global interaction information component, this invention prioritizes inputting denser user group interaction vectors for model training. This not only improves the model's performance but also effectively accelerates the model's training, enabling the overall method to better adapt to scenarios with sparse interaction records. Attached Figure Description

[0042] Figure 1 This is a structural diagram of the ensemble learning model in this invention;

[0043] Figure 2 This is a structural diagram of the global interactive information component;

[0044] Figure 3 A structural diagram of the feature matching information component; Detailed Implementation

[0045] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, not all embodiments, and are not intended to limit the scope of the present invention. Furthermore, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion regarding the concepts disclosed in the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0046] The accompanying drawings show structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not drawn to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0047] Example 1

[0048] refer to Figures 1 to 3 The job recommendation method based on an ensemble learning model of the present invention includes the following steps:

[0049] 1) Construct pooled user vectors

[0050] The features of users and recommended items are vectorized and encoded so that any user or recommended item can be represented by a vector.

[0051] Here, let user u have m-dimensional features, and user u be represented as... in, Let represent the value of the i-th feature of user u. Similarly, if the recommended item c has n-dimensional features, then the recommended item c is represented as... Simultaneously, a pooled user vector is constructed, and the pooled user vector of recommended item c is denoted as u. c .

[0052] For a specific recommended item c, let U be the set of users whose interaction records contain this item. c Set the pooled user vector u of the recommended item c. c For all user vectors u∈U c After summing, divide by the number of users in the user set (|U c The vector obtained by |) is:

[0053]

[0054] A pooled user vector does not represent a specific user, but it is formally identical to a user vector. Geometrically, a pooled user represents the set of users U. c The center of a spherical cluster in Euclidean space.

[0055] 2) Training of integrated learning model components

[0056] The ensemble learning model includes a similarity information component, a global interaction information component, and a feature matching information component.

[0057] 2.1) The training process for the similarity information component is as follows:

[0058] Calculate the Euclidean distance (denoted as E_distance in the formula) between the corresponding user vectors, and then calculate the similarity between users according to Formula 2.

[0059]

[0060] To address the challenges of distinguishing user similarity and determining nearest neighbors in traditional collaborative filtering algorithms for scenarios with sparse interaction records, this invention introduces the similarity between the target user and a specific user group (i.e., pooled user vectors) to further determine the recommendation score of recommended items. Let the set of all interaction records be I, and user u... i The set of interaction records is Its vector representation is The set of interaction records containing the recommended item c is denoted as I. c In quantifying user u i When giving a recommendation score to item c, in addition to considering user u i Similarity to individual historical users also needs to be considered in addition to user u i The degree of similarity with the user group containing item c in all interaction records, i.e., the user vector u. i The pooled user vector u corresponding to project c c The similarity, therefore, assuming the recommendation list does not contain duplicate recommended items, user u i The final recommendation score for item c will be determined by satisfying the following criteria: All users and those connected to user u i The users with the highest similarity jointly determine the recommendation score S(u) for item c. i c) is:

[0061]

[0062] in, For user u j The collection of interaction records, sim(u i ,u j) for user u i With user u j Similarity, sim(u i ,u c ) for user u i With user u c similarity, These are the weighting parameters for the two parts, both defaulting to 1. You can adjust these weighting parameters to adjust the emphasis on each part's score based on the actual situation. This is because user u is taken into consideration. i The similarity between the user group corresponding to the recommended item c and sim(u) i ,u c Therefore, it can be further derived from S(u) i The size of c) directly determines the position of each recommended item c in the final output recommendation list. Meanwhile, due to the pooling of users u... c It is composed of user group U c This is a joint decision, so adding a user group similarity metric can alleviate the noise interference problem to some extent.

[0063] The recommendation process based on similarity information components is as follows:

[0064] For the current user u i Its recommended list rec_list(u i Initially empty.

[0065] Project set C, For each recommended item c, use formula (1) to construct the pooled user vector u. c .

[0066] Call formula (3) to select S(u) i The largest c (c∈C) except ), add it to rec_list(u i In ), update C except =C except ∪{c}.

[0067] Repeat the previous step until... Or rec_list(u i The algorithm process is shown in Table 1, which meets the specified requirements.

[0068] 2.2) The training process for the global interactive information component is as follows:

[0069] Constructing user group interaction vectors

[0070] For a given item c, c∈I will satisfy iThe interaction record vectors of all users in the user cluster are weighted and pooled, with the weights being the current user u and the pooled user u corresponding to that user cluster. c The similarity between them, i.e.

[0071]

[0072] Result I of the construction c It is a relatively dense vector with the same dimension as the interaction record vector of a single user, representing the interaction record of the user group corresponding to a certain recommended item c.

[0073] The training process for the global interactive information component is as follows:

[0074] The global interactive information component is implemented using a three-layer neural network autoencoder with a single hidden layer. Its loss function is MSE, and an L2 regularization term is added to mitigate overfitting. The final optimization objective of the model is:

[0075]

[0076] Where U is the set of all users, I u Let be the vector representation of user u's interaction records, θ be the trainable parameters of the autoencoder, i.e., the transformation matrix between the input layer and the hidden layer and between the hidden layer and the output layer, h(I u ;θ) is the output of the autoencoder, i.e., I u The predicted values ​​are given by λ, which is the regularization parameter, V, which is the transformation matrix from the input layer to the hidden layer, and W, which is the transformation matrix from the hidden layer to the output layer.

[0077] To address the issues of insufficient autoencoder training, slow convergence, and difficulty in updating model parameters caused by sparse interaction record vectors during the training phase, the global interaction information component, in addition to inputting the interaction record vector I of a single user, also incorporates other inputs during training. u In addition, user group interaction vectors I will be constructed based on user groups. c It is also used as the training input for the autoencoder. This is because the vector is denser and its meaning does not deviate from the original training objective. Simultaneously, the user group interaction vector I is given priority as input. c Model training is performed because I c The method incorporates dense user group interaction vectors, allowing the model to rapidly receive a large amount of information, thus enabling faster and more thorough convergence during training. Therefore, adding user group interaction vectors not only improves model performance but also effectively accelerates training, allowing the model to better adapt to scenarios with sparse interaction records.

[0078] Furthermore, this invention also specifies the stopping conditions for the current global interactive information component. Generally, model training uses the elbow rule for early stopping. However, in this invention, due to the sparsity of the input vector, the loss function value often approaches 0, resulting in insignificant changes during iteration, which is not conducive to monitoring and determining the timing of stopping training. Considering that the model's ultimate purpose is to predict the values ​​of unknown interactions, let cell_loss be used as the loss function for monitoring the model's training progress on the validation set, specifically:

[0079]

[0080] Among them, I v This represents the set of all interaction record vectors used as the validation set. This represents the actual value of the interaction record between user u and project c in this interaction record. This is the model's predicted value for this interaction. This loss function specifically verifies the interaction data missing from the training set, rather than the overall difference between output vectors, which is more in line with the model's task objective and can therefore assist in model training.

[0081] After training the autoencoder using interactive data, the existing interaction record vectors of the current user are input into the autoencoder. The scores of each item output by the autoencoder are arranged from high to low, and the corresponding recommendation list is output, which is the recommendation result of the global interactive information component.

[0082] 2.3) The training process for the feature matching information component is as follows:

[0083] Existing job application records are described as tuples of (User ID, Job ID), and are considered positive examples, thus being labeled as positive examples. Negative example tuples are generated from jobs the user has not applied for, through random negative sampling, and are also labeled as negative examples. The negative sampling ratio is set to 40% of the number of positive examples. The user's representation vector is then concatenated within each tuple. and project vector Used as positive and negative samples for training.

[0084] The entire network takes the aforementioned positive and negative samples as input and outputs a score representing the degree of user-job matching as a recommendation score. The embedding layer transforms the encoded high-dimensional sparse feature vectors into low-dimensional dense vectors; it is a fully connected layer. The attention scoring layer employs a cross-attention mechanism, also a fully connected layer, providing attention scores for both the user and recommendation item sides. Specifically, the embedding vectors of all user-side features are used as input to provide the attention score for each feature on the recommendation item side, and vice versa. The concatenation layer multiplies the user and recommendation item attention scores with their corresponding feature embedding vectors and concatenates them as input to the scoring layer. The scoring layer outputs the matching degree recommendation score, and the final output is obtained using a Logistic Regression (LR) model. The mathematical form of LR is:

[0085]

[0086] The activation function of the scoring layer is the Sigmoid function, and the activation function of the output layer of the attention score generation network is the Tanh function.

[0087]

[0088] All other activation functions uniformly use the ReLU function, specifically:

[0089] Relu(x) = max(0,x) (9)

[0090] The model parameters are trained using the Adam optimizer until convergence, with the optimization objective being to minimize the loss function.

[0091] This component takes a target user and all recommended items as input, and outputs a recommendation list for the target user and the corresponding recommendation score according to the scores of each recommended item.

[0092] 3) Job information recommendation based on ensemble learning models

[0093] For a target user u and a recommended item c, the recommendation score for the user is obtained by weighted summation of the recommendation scores from the three components. The similarity information component gives the recommendation degree score of recommended item c for target user u by judging the similarity between users. The recommendation score S(u,c) is obtained by inputting the feature vector of target user u.

[0094] The global interaction information component provides a recommendation score for the recommended item c to the target user u from the perspective of global attention. It takes the target user u’s existing interaction record vector as input and calculates the recommendation score for the target user u for each recommended item c through forward propagation of the model to obtain G(u,c).

[0095] The feature matching information component provides a recommendation score based on the degree of bidirectional feature matching between the user and the recommended item. The input feature vectors of the target user u and the recommended item c are used to perform forward propagation calculation by a trained neural network to obtain the recommendation score M(u,c).

[0096] The final recommendation score for item c for target user u is:

[0097]

[0098] Where, α i (i = 1, 2, 3) represent the weights of each component, with a default value of 1. The recommendation scores for all recommended items c for the target user u are calculated according to the above process. The recommended items are sorted in descending order of their recommendation scores to obtain the final recommendation list.

[0099] Example 2

[0100] The method described in this invention was tested using data recorded at an online campus job fair. This dataset contains three parts: "student characteristics," "job characteristics," and "job application records." The specific features and related descriptions of the student characteristics data are shown in Table 2, and the specific features and related descriptions of the job characteristics data are shown in Table 3. Job application records are described as tuples of (user ID, job ID). A total of 8951 students registered for the event, but only 1840 students had interaction records, resulting in 5086 valid applications for 841 jobs. The experiment compared the recommendation performance of the method described in this invention (EIGRM) with classic recommendation methods including Deep Crossing, GPRM, DIN, AutoRec, CF, and FCF. Recall, the modified mean average precision (mAP), and the F1-score were used as evaluation metrics for the recommendation model's recommendation results. The experimental results are shown in Table 4, demonstrating that this invention achieves optimal recommendation results.

[0101] Table 1

[0102]

[0103] Table 2

[0104]

[0105] Table 3

[0106]

[0107]

[0108] Table 4

[0109]

[0110] Example 3

[0111] The job recommendation system based on an ensemble learning model described in this invention includes:

[0112] The encoding module is used to obtain the target user u and each recommended item c, and to vectorize and encode the features of the target user u and the features of each recommended item c.

[0113] The recommendation module is used to input the feature vector of the target user u, the existing interaction record vector of the target user u, and the feature vectors of the target user u and each recommendation item c into the trained ensemble learning model to obtain the score of each recommendation item and complete the job recommendation based on the ensemble learning model. The ensemble learning model consists of a similarity information component, a global interaction information component, and a feature matching information component.

[0114] Recommended modules include:

[0115] The first calculation module is used to input the feature vector of the target user u into the trained similarity information component to obtain the recommendation score S(u,c);

[0116] The second calculation module is used to input the existing interaction record vector of the target user u into the trained global interaction information component to obtain the recommendation score G(u,c);

[0117] The third calculation module is used to input the feature vectors of the target user u and each recommended item c into the trained feature matching information component to obtain the recommendation score M(u,c);

[0118] The fourth calculation module is used to calculate the recommendation score of recommended item c for the target user u:

[0119]

[0120] Where, α i (i = 1, 2, 3) represent the weights of the similarity information component, the global interaction information component, and the feature matching information component, respectively.

[0121] Example 4

[0122] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the job recommendation method based on an ensemble learning model. The memory may include main memory, such as high-speed random access memory (RAM), and may also include non-volatile memory, such as at least one disk storage device. The processor, network interface, and memory are interconnected via an internal bus, which may be an industry-standard architecture bus, a peripheral component interconnection standard bus, an extended industry-standard architecture bus, etc. The bus may be categorized as an address bus, a data bus, a control bus, etc. The memory stores the program; specifically, the program may include program code, which includes computer operation instructions. The memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0123] Example 5

[0124] A computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the job recommendation method based on an ensemble learning model. Specifically, the computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. The volatile memory may include random access memory (RAM) and / or cache memory, etc. The non-volatile memory may include read-only memory (ROM), hard disk, flash memory, optical disk, magnetic disk, etc.

[0125] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0126] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0127] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0128] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An employment recommendation method based on an ensemble learning model, characterized by, The method comprises the steps of: obtaining a user u and each recommendation item c, and vectorizing and encoding the features of the user u and the features of each recommendation item c; inputting the feature vector of the user u, the interaction record vector of the user u and the feature vector of the user u and each recommendation item c into the trained integrated learning model to obtain the score of each recommendation item, and completing the employment recommendation based on the integrated learning model, wherein the integrated learning model is composed of a similarity information component, a global interaction information component and a feature matching information component; the specific process of inputting the feature vector of the user u, the interaction record vector of the user u and the feature vector of the user u and each recommendation item c into the trained integrated learning model to obtain the score of each recommendation item is as follows: inputting the feature vector of the user u into the trained similarity information component to obtain a recommendation score S(u, c); inputting the interaction record vector of the user u into the trained global interaction information component to obtain a recommendation score G(u, c); inputting the feature vector of the user u and each recommendation item c into the trained feature matching information component to obtain a recommendation score M(u, c); the recommendation score of the recommendation item c for the user u is calculated as: wherein, α i (i = 1, 2, 3) represent the weights of the similarity information component, the global interaction information component and the feature matching information component, respectively. For a user u i , the recommendation score S(u i , c) of an item c is: wherein, is a set of interaction records of the user u j , sim(u i , u j ) is a similarity of the user u i to the user u j , sim(u i , u c ) is a similarity of the user u i to the user u c , and ε1 and ε2 are weight coefficients. the optimization target in the training process of the global interaction information component is: where U is the set of all users, I u is the vector representation of the interaction record of user u, θ is the trainable parameter of the autoencoder, h(I u ; θ) is the output result of the autoencoder, λ is the regularization parameter, V is the conversion matrix between the input layer and the hidden layer, and W is the conversion matrix between the hidden layer and the output layer. the loss function in the training process of the global interaction information component is: where I v denotes the set of all interaction record vectors used as validation set, denotes the true value of the interaction of user u with item c in the interaction record, is the predicted value for this interaction. 2.The employment recommendation method based on the ensemble learning model according to claim 1, characterized in that, the specific process of vectorizing and encoding the features of the user u and the features of each recommendation item c is as follows: Let user u have m-dimensional features, user u is represented as wherein, represents the value of the i-th dimensional feature of user u, let recommended item c have n-dimensional features, recommended item c is represented as Construct a pooled user vector u of recommended item c c . 3.The employment recommendation method based on the ensemble learning model according to claim 2, characterized in that, The pooling user vector u of the constructed recommendation item c c The specific process is as follows: For the recommended item c, the set of users in the interaction record containing the recommended item is denoted as U c The pooled user vector u for the recommended item c is then c is: 4.The employment recommendation method based on the ensemble learning model according to claim 1, characterized in that, the feature matching information component comprises an input layer, an embedding layer, an attention score layer, a connection layer and a scoring layer connected in sequence; The existing post delivery record is described as a tuple in the form of (user ID, post ID), regarded as a positive example, given a positive example label, and negative example tuples are generated from posts that the user has not delivered through random negative sampling, given a negative example label, and the negative sampling ratio is set to 40% of the number of positive examples. The user's representation vector in each tuple is spliced and the project vector Training as positive and negative samples; the input of the input layer is positive and negative samples, the output of the scoring layer is a score representing the matching degree of the user and the post, i.e., a recommendation score, the embedding layer is used to convert the encoded high-dimensional sparse feature vector into a low-dimensional dense vector, and the attention score layer adopts a cross attention mechanism to give the attention scores of the user side and the recommendation item side respectively.

5. An employment recommendation system based on an ensemble learning model, characterized by, The method comprises the steps of: an encoding module is configured to obtain a user u and each recommendation item c, and vectorize and encode the features of the user u and the features of each recommendation item c; a recommendation module is configured to input the feature vector of the user u, the interaction record vector of the user u and the feature vector of the user u and each recommendation item c into the trained integrated learning model to obtain the score of each recommendation item, and complete the employment recommendation based on the integrated learning model, wherein the integrated learning model is composed of a similarity information component, a global interaction information component and a feature matching information component; the specific process of inputting the feature vector of the user u, the interaction record vector of the user u and the feature vector of the user u and each recommendation item c into the trained integrated learning model to obtain the score of each recommendation item is as follows: inputting the feature vector of the user u into the trained similarity information component to obtain a recommendation score S(u, c); inputting the interaction record vector of the user u into the trained global interaction information component to obtain a recommendation score G(u, c); inputting the feature vector of the user u and each recommendation item c into the trained feature matching information component to obtain a recommendation score M(u, c); the recommendation score of the recommendation item c for the user u is calculated as: wherein a i (i = 1, 2, 3) represent the weights of the similarity information component, the global interaction information component, and the feature matching information component, respectively. For a user u i , the recommendation score S(u i , c) for an item c is: wherein, is a set of interaction records of the user u j , sim(u i , u j ) is a similarity of the user u i to the user u j , sim(u i , u c ) is a similarity of the user u i to the user u c , and ε1 and ε2 are weight coefficients. the optimization target in the training process of the global interaction information component is: where U is the set of all users, I u is the vector representation of the interaction record of user u, θ is the trainable parameter of the autoencoder, h(I u ; θ) is the output result of the autoencoder, λ is the regularization parameter, V is the conversion matrix between the input layer and the hidden layer, and W is the conversion matrix between the hidden layer and the output layer. The loss function in the global interaction information component training process is: where I v denotes the set of all interaction record vectors used as validation set, denotes the true value of the interaction of user u with item c in the interaction record, is the predicted value for this interaction.

6. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor implements the steps of the employment recommendation method based on the ensemble learning model according to any one of claims 1-4 when executing the computer program.

7. A computer-readable storage medium storing a computer program, wherein the computer program comprises the following steps of: receiving a request for a resource from a client; determining whether the client is authorized to access the resource; and if the client is authorized to access the resource, providing the resource to the client. The computer program is executed by the processor to implement the steps of the employment recommendation method based on the ensemble learning model according to any one of claims 1-4.

Images