Multi-input multi-output salary prediction method based on recruitment advertisement data
Through the multi-input and multi-output salary prediction model based on BERT and LSTM, the job advertising data is processed, and the problem of insufficient salary prediction accuracy in the existing technology is solved, and more accurate salary prediction is achieved, helping companies and job seekers to make more scientific decision making.
Patent Information
- Application Number
- CN202411276445.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-12
- Publication Date
- 2024-12-13
AI Technical Summary
It is difficult for the existing technology to accurately predict compensation, especially under a variety of complex factors and nonlinear relationships, which leads to insufficient accuracy of the compensation prediction model, affecting corporate compensation strategies and job seekers' career planning.
Using a multi-input and multi-output salary prediction method based on recruitment advertising data, BERT and LSTM models are used to process different forms of data, and a model with input as recruitment requirements and output as upper and lower limits of salary is constructed to capture various complex factors affecting salary and their non-existent factors. Linear relationships to improve the complexity and flexibility of the model.
It has achieved more accurate salary predictions, helping companies formulate scientific compensation strategies and job seekers to reasonably evaluate salary expectations, and considering various factors such as industry type, enterprise size, and regional differences, improving prediction accuracy and accuracy.
Smart Images
Figure CN119129835A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data processing, and in particular to a multi-input and multi-output salary prediction method based on recruitment advertisement data. Background Art
[0002] In today's increasingly competitive job market, both companies and job seekers are faced with the challenge of how to reasonably evaluate and predict salaries. With technological development and industry changes, job requirements are becoming more and more complex. Therefore, in order to attract and retain high-quality talents, companies need to ensure that their compensation strategies match market trends. For job seekers, understanding their own market value in different work scenarios and reasonably anticipating salary levels are the keys to their career development. However, salary is affected by many factors, such as industry type, company size, regional differences, work experience, educational background, etc. The complex interaction between these factors makes it extremely difficult to accurately predict salary.
[0003] Since salary is affected by many variables and there are complex nonlinear relationships between these variables, simple linear models often fail to capture all the information. Therefore, building an effective salary prediction model can help companies formulate more scientific salary strategies during the recruitment process, ensuring that the salary level can both attract outstanding talents and meet the company's budget; at the same time, the salary prediction model can also help job seekers reasonably evaluate their salary expectations based on their own conditions and market demand, so as to better plan their careers. Summary of the invention
[0004] In order to make up for the shortcomings of the existing technology, the present invention proposes a multi-input and multi-output salary prediction method based on recruitment advertisement data, which captures the various complex factors affecting salary and their nonlinear relationships, but the prediction accuracy still needs to be optimized.
[0005] The technical solution adopted by the present invention to solve its technical problems is: the multi-input and multi-output salary prediction method based on recruitment advertisement data described in the present invention obtains salary data from recruitment notices and recruitment resumes to obtain an initial salary prediction model, strengthens the feature processing and output processing of the salary prediction model, allows different forms of data to be processed in different ways, connects scalars with a fully connected layer, and connects sequence data separately with an LSTM layer, so that the salary prediction model can process multiple input features and predict multiple outputs (such as a salary upper limit and a salary lower limit), increases the complexity and flexibility of the salary prediction model, and captures more subtle relationships in the data.
[0006] Preferably, the multiple input features include a time vector, an industry vector, a position address vector, a number of skills, a company size vector, a welfare vector, a financing stage vector, an education level, and years of work experience, and the multiple outputs include a salary cap (us j) and salary floor (ls j ).
[0007] Preferably, the information of the recruitment market is used to predict the salary. The jth recruitment position sample P j The published job information will contain the jth job market information M j , the jth position level requirement lv j , where the jth market information M j ={p j ,w j ,g j ,I j ,T j ,f j}, the jth position level requirement lv j ={exp j ,b j ,c j},
[0008] where p j is the jth job address, w j is the welfare of the jth position, g j is the company size of the jth position, I j is the industry with the jth skill combination, T j is the posting time of the jth job, f j is the financial stage of the company in the jth position, exp j is the work experience required for the jth position, b j is the educational background required for the jth position, c j is the number of skills required for the jth position.
[0009] Preferably, S j is the salary requirement for the jth position, S j ={ls j ,us j}, where ls j is the lower limit of salary for the jth position, us j is the upper limit of the salary of the jth position. Based on the salary prediction problem in this paper, the expression of the salary prediction model is finally obtained, that is,
[0010] (us j ,ls j )=f(M j ,lv j |Θ)
[0011] Where Θ represents the hyperparameters that can be adjusted in subsequent experiments, such as learning rate and batch size.
[0012] The beneficial effects of the present invention are as follows: In order to achieve accurate and effective salary prediction, the salary change trend is given, and the enterprise is helped to make a reasonable salary strategy and the job seeker is helped to choose the expected job, and the characteristics that affect the salary, such as industry type, enterprise scale, regional differences, work experience, and educational background, are considered. BERT (Bidirectional Encoder Representations from Transformers) is used to process the text data into vectors, and finally LSTM (Long Short-Term Memory) is used as the basic framework to construct a multi-input multi-output salary prediction model with the input as the recruitment requirements and the output as the upper and lower limits of the job salary. By constructing a multi-input multi-output salary prediction model based on BERT and LSTM, the various complex factors affecting the salary and their nonlinear relationships are successfully captured, but the prediction accuracy still needs to be optimized. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] The present invention is further described below in conjunction with the accompanying drawings and implementation modes.
[0014] Figure 1 is a coding unit diagram of the Transformer model of the present invention;
[0015] Figure 2 It is a structural diagram of the BERT model of the present invention.
[0016] Figure 3 It is a structural diagram of the LSTM model of the present invention. DETAILED DESCRIPTION
[0017] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further described below in conjunction with the accompanying drawings and preferred embodiments.
[0018] like Figure 1 As shown in the figure, a multi-input and multi-output salary prediction method based on recruitment advertisement data obtains salary data from recruitment notices and recruitment resumes to obtain an initial salary prediction model, strengthens the feature processing and output processing of the salary prediction model, allows different forms of data to be processed in different ways, connects scalars with a fully connected layer, and connects sequence data separately with an LSTM layer, so that the salary prediction model can process multiple input features and predict multiple outputs (such as salary caps and salary floors), increases the complexity and flexibility of the salary prediction model, and captures more subtle relationships in the data.
[0019] Preferably, the multiple input features include a time vector, an industry vector, a position address vector, a number of skills, a company size vector, a welfare vector, a financing stage vector, an education level, and years of work experience, and the multiple outputs include a salary cap (us j ) and salary floor (ls j ).
[0020] Preferably, the information of the recruitment market is used to predict the salary. The jth recruitment position sample P j The published job information will contain the jth job market information M j , the jth position level requirement lv j , where the jth market information M j ={p j ,w j ,g j ,I j ,T j ,f j}, the jth position level requirement lv j ={exp j ,b j ,c j},
[0021] where p j is the jth job address, w j is the welfare of the jth position, g j is the company size of the jth position, I j is the industry with the jth skill combination, T j is the posting time of the jth job, f j is the financial stage of the company in the jth position, exp j is the work experience required for the jth position, b j is the educational background required for the jth position, c j is the number of skills required for the jth position.
[0022] Preferably, S j is the salary requirement for the jth position, S j ={ls j ,us j}, where ls j is the lower limit of salary for the jth position, us j is the upper limit of the salary of the jth position. Based on the salary prediction problem in this paper, the expression of the salary prediction model is finally obtained, that is,
[0023] (us j ,ls j )=f(M j ,lv j |Θ)
[0024] Where Θ represents the hyperparameters that can be adjusted in subsequent experiments, such as learning rate and batch size.
[0025] We used crawler technology to crawl job data from different industries from 2016 to 2024 from websites such as Boss Direct, Lagou, and Zhaopin. We used BERT (Bidirectional Encoder Representations from Transformers) to process text data into vectors. Finally, we used LSTM (Long Short-Term Memory) as the basic framework to construct a multi-input and multi-output salary prediction model with recruitment requirements as input and upper and lower limits of job salary as output.
[0026] BERT is a pre-trained language model launched by Google. It is widely used in many tasks in the field of natural language processing (NLP). Its internal structure consists of the encoder part of a multi-layer transformer. The structure of the encoder is as follows: Figure 1 As shown in the figure. The Transformer model uses the Self-Attention mechanism. Self-Attention can parallelly calculate the relevance and semantic dependency of all words in a sentence to the current word. Therefore, its feature extraction capability is stronger than that of RNN. The core of this encoding unit is Self-Attention and feedforward neural network. Through these two structures, the input text data can be converted into the corresponding representation vector.
[0027] from Figure 2 It can be seen that the BERT model is composed of multiple layers of Transformer Encoder stacks, where Figure 2 The "Trm" in Figure 1 BERT can learn the deep semantic information of text, thereby generating more representative and expressive word vectors to improve model performance.
[0028] The reason why the present invention chooses LSTM is mainly to take into account the overall impact of the market, because the corresponding data has properties such as periodicity and continuity. The LSTM model, as a convolutional neural network model for time series problems derived from the recurrent neural network, can use gate functions to explore the time series change rules in time series.
[0029] The LSTM model is mainly used to model time series data. Its gating mechanism can effectively handle long-term dependencies in time series data. It is mainly composed of 3 gating units and 1 memory unit to control the flow and storage of information. The input gate determines how much new information is added to the cell, the forget gate determines whether the information will be forgotten at each moment, and the output gate determines whether there is information output at each moment. The basic structure of the LSTM model is as follows: Figure 3 shown.
[0030] Input Gate: Controls the flow of input information.
[0031] Forget Gate: Determines how much previous information should be forgotten.
[0032] Output Gate: Controls how much information is passed from the cell state to the hidden state.
[0033] Cell State: Stores long-term information throughout the entire LSTM network.
[0034] Hidden State: Stores short-term information and passes it to the next time step.
[0035] In order to accurately predict the upper and lower bounds of salary, the model inputs include time vector, industry vector, position address vector, number of skills, company size vector, benefits vector, financing stage vector, education level and years of work experience. These input features represent important factors affecting salary, covering multiple dimensions such as position information, company background, personal skills and economic environment.
[0036] The output of the model includes two target variables: salary cap (us j ) and salary floor (ls j ). Using two separate output layers to predict the two target variables ensures that the model can simultaneously learn the features that affect different salary levels.
[0037] In addition, in order to effectively process various types of input data, we perform specific processing on each input feature. Take the j-th job position sample as an example:
[0038] a. Time (T j )、Industry(I j ) and position location (p j): These inputs are sequence data. In order to convert the relevant text into data that can be used by computers, it is also necessary to convert the text into semantic vectors through a word embedding model. Here, the BERT (Bidirectional Encoder Representations from Transformers) model is selected as the word embedding model. The LSTM (Long Short-Term Memory Network) layer is used for processing, which can capture the time dependency of sequence data and thus learn the temporal change pattern of features.
[0039] b. Number of skills (c j )、Educational background(b j ) and years of working experience (exp j ): These are scalar inputs that are processed through fully connected (Dense) layers to learn the impact of the number of skills, education background, and years of experience on salary.
[0040] c. Company size (g j )、welfare(w j )、Financing stage(f j ): These features are processed by fully connected (Dense) layers to extract the nonlinear relationship between features.
[0041] d. This article also performs batch normalization and dropout on some features to improve the generalization ability of the model and prevent overfitting.
[0042] e. Feature fusion: All input features are concatenated after passing through their respective processing layers to form a comprehensive feature representation that captures the relationship between all input features, providing rich input for subsequent dense layers.
[0043] f. Model output, based on the comprehensive feature representation, this paper adds two fully connected layers (Dense) and uses the Huber loss function (Huber loss) to optimize the model. The Huber loss function can provide a compromise between balancing the mean square error (MSE) and the mean absolute error (MAE), which is very effective when dealing with noisy regression tasks.
[0044] Finally, the two output layers of the model are used to predict the upper and lower bounds of the salary. Through this multi-input and multi-output model architecture design, the model can simultaneously learn the impact of multiple features on different salary levels, thereby improving the accuracy of the prediction.
[0045] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.
Claims
1. A multi-input and multi-output salary prediction method based on recruitment advertisement data, which obtains salary data from recruitment announcements and resumes to obtain an initial salary prediction model. Features: The feature processing and output processing of the salary prediction model have been strengthened, so that different forms of data can be processed in different ways. Scalars are connected with fully connected layers, and sequence data are connected separately with LSTM layers. This enables the salary prediction model to process multiple input features and predict multiple outputs, increasing the complexity and flexibility of the salary prediction model and capturing more subtle relationships in the data.
2. A multi-input multi-output salary prediction method based on recruitment advertisement data according to claim 1, Features: Multiple input features include time vector, industry vector, job address vector, number of skills, company size vector, benefits vector, financing stage vector, education level and years of work experience, and multiple outputs include salary cap (us j ) and salary floor (ls j ).
3. A multi-input multi-output salary prediction method based on recruitment advertisement data according to claim 1, Features: Using the information of the recruitment market to predict the salary, the jth recruitment position sample P j The published job information will contain the jth job market information M j , the jth position level requirement lv j , where the jth market information M j ={p j ,w j ,g j ,I j ,T j ,f j }, the jth position level requirement lv j ={exp j ,b j ,c j }, where p j is the jth job address, w j is the welfare of the jth position, g j is the company size of the jth position, I j is the industry with the jth skill combination, T j is the posting time of the jth job, f j is the financial stage of the company in the jth position, exp j is the work experience required for the jth position, b j is the educational background required for the jth position, c j is the number of skills required for the jth position.
4. A multi-input multi-output salary prediction method based on recruitment advertisement data according to claim 1, Features: S j is the salary requirement for the jth position, S j ={ls j ,us j }, where ls j is the lower limit of salary for the jth position, us j is the upper limit of the salary of the jth position. Based on the salary prediction problem in this paper, the expression of the salary prediction model is finally obtained, that is, (us j ,ls j )=f(M j ,lv j |Θ) Where Θ represents the hyperparameters that can be adjusted in subsequent experiments, such as learning rate and batch size.