KR102759478B1 - Job search matching method and system - Google Patents

Abstract

The present invention relates to a matching method and system capable of providing a high-quality corporate recommendation service to a user based on a job posting on a job search site, and relates to a method and system capable of matching and providing a job posting that meets a user's requirements based on a user's selection, the method including: a step of receiving at least one desired corporate condition and user specification from the user; a step of learning and calculating a correlation between companies satisfying the corporate condition and corporate preference according to at least one of the user specifications; a step of determining a ranking of companies satisfying the corporate condition according to the correlation; and a step of presenting a list of companies to the user according to the corporate ranking according to the correlation.

Description

{JOB SEARCH MATCHING METHOD AND SYSTEM}

The present invention relates to a matching method and system capable of providing a quality company recommendation service to a user based on a job posting on a job search site, and more particularly, to a method and system capable of matching and providing a job posting that meets the user's requirements based on the user's selection.

In the past, you could get information about job hunting through the job posting section of a newspaper or through publications, but today, you can check the job postings of various companies through the Internet or app services, and choose and apply to companies that meet your qualifications and desired conditions.

Existing job search sites recommend companies in the following order: company list by category, reviews, benefits, and salary. Therefore, users have to check and select information such as company list by category, reviews, benefits, and salary in order to select the company they want to apply to. However, in order to select a company among numerous companies, they have to check the information of each company one by one to see if it meets their desired conditions, which consumes a considerable amount of time.

Therefore, if machine learning is used to recommend similar companies for any job posting that a user wants, it will be easier for the user to use than having to input multiple conditions directly into the system. A system that provides satisfactory recommendation results to the user while having the most concise and intuitive usage method is needed.

Accordingly, the present invention is intended to solve the above-mentioned problems and to provide a method and system capable of listing and recommending companies that suit the user's preferences.

The present invention seeks to provide a method and system capable of finding attributes that affect the achievement of a user's purpose, minimizing the amount of computation, thereby reducing the burden on the system and increasing accuracy.

The present invention seeks to provide a method and system that enables a user to understand the characteristics of a corporate job posting using only visualization data without necessarily reading the text, by utilizing various visualization methods.

According to an embodiment of the present invention for solving the above-described problem, a job search matching method may include a step of receiving input of at least one desired corporate condition and user specification from a user; a step of learning and calculating a correlation between companies satisfying the corporate condition and corporate preference according to at least one of the user specifications; a step of determining a ranking of companies satisfying the corporate condition according to the correlation; and a step of presenting a list of companies to the user according to the corporate ranking according to the correlation.

According to one embodiment of the present invention, the correlation calculation step may include: a step of matching at least one user specification and a corporate preference; a step of learning by defining a correlation between a change in the user specification and a change in the corporate preference as a probability; and a step of calculating a correlation between the corporate condition and the corporate preference according to the at least one user specification according to the probability.

According to one embodiment of the present invention, in the correlation calculation step, the correlation can be calculated by the following mathematical expression 1.

[Mathematical formula 1]

(Here, C(S, P) is the correlation, E(S) is the analysis relevance value, S is the user specification, and P is the corporate preference.)

According to one embodiment of the present invention, the user specifications may include at least one of school, major, gender, age, region, and MBTI.

According to one embodiment of the present invention, the step of determining the ranking of the companies determines a ranking including a company preference according to a combination of a plurality of the user specifications, and the user presentation step can provide a result according to the ranking.

According to an embodiment of the present invention for solving the above-described problem, a job search matching system may include a data processing unit for receiving at least one desired corporate condition and user specification from a user; a learning unit for learning and calculating a correlation between companies satisfying the corporate condition and corporate preference according to at least one of the user specifications; a ranking unit for determining a ranking of companies satisfying the corporate condition according to the correlation; and a visualization unit for presenting a list of companies to the user according to the corporate ranking according to the correlation.

According to one embodiment of the present invention, the learning unit can match at least one user specification with a corporate preference, learn by defining a correlation between a change in the user specification and a change in the corporate preference as a probability, and calculate a correlation between the corporate condition and the corporate preference according to at least one user specification according to the probability.

According to one embodiment of the present invention, the learning unit can calculate the correlation by the following mathematical expression 1.

[Mathematical formula 1]

(Here, C(S, P) is the correlation, E(S) is the analysis relevance value, S is the user specification, and P is the corporate preference.)

According to one embodiment of the present invention, the user specifications may include at least one of school, major, gender, age, region, and MBTI.

According to one embodiment of the present invention, the learning unit determines a ranking including corporate preferences according to a combination of a plurality of the user specifications, and the visualization unit can provide a result according to the ranking.

According to the present invention, companies that match the user's preferences can be listed and recommended in order of rank.

In addition, it is possible to find attributes that affect the user's goal achievement, minimize the amount of computation, and increase the accuracy of prediction while reducing the burden on the system.

In addition, by using various visualization methods, users can understand the characteristics of corporate job postings through visualization data alone without having to read the text.

Meanwhile, the effects of the present invention are not limited to the effects mentioned above, and various effects may be included within a range obvious to those skilled in the art from the contents described below.

Figure 1 is a block diagram of a job search matching system according to one embodiment of the present invention.
Figure 2 is a screen during analysis of a matching page that applies a job search matching method according to one embodiment of the present invention.
Figure 3 is a matching result screen that applies a job search matching method according to one embodiment of the present invention.
Figure 4 is a visualization result screen that applies a job search matching method according to an embodiment of the present invention.
FIG. 5 is a result screen when there is no desired information on a matching page that applies a job search matching method according to an embodiment of the present invention.
Figure 6 is a result screen when there is no matching company as a result of applying a job search matching method according to one embodiment of the present invention.
Figure 7 is a matching result screen that applies a job search matching method according to one embodiment of the present invention.
Figure 8 is a flow chart of a job search matching method according to one embodiment of the present invention.

Hereinafter, the 'job search matching method and system' according to the present invention will be described in detail with reference to the attached drawings. The described embodiments are provided so that those skilled in the art can easily understand the technical idea of the present invention, but the present invention is not limited thereby. In addition, the matters expressed in the attached drawings are schematic drawings for easily explaining the embodiments of the present invention, and may differ from the form actually implemented.

Meanwhile, each component expressed below is only an example for implementing the present invention. Therefore, in other implementations of the present invention, other components may be used without departing from the spirit and scope of the present invention.

In addition, each component may be implemented purely by hardware or software configurations, but may also be implemented by a combination of various hardware and software configurations that perform the same function. In addition, two or more components may be implemented together by a single hardware or software.

Additionally, the expression 'including' certain components is an 'open' expression, simply indicating the presence of those components, and should not be understood as excluding additional components.

Figure 1 is a block diagram of a job search matching system according to one embodiment of the present invention.

Referring to FIG. 1, a job search matching system (1) according to one embodiment of the present invention may include a data processing unit (100), a learning unit (200), and a visualization unit (300).

The data processing unit (100) above can receive at least one desired corporate condition and user specification from the user. The data processing unit (100) can collect corporate job search information through job search and recruitment sites, etc. The data processing unit (100) can crawl and collect corporate job search and recruitment information from job search and recruitment sites. The data processing unit (100) can receive at least one desired corporate condition from the user. The data processing unit (100) can receive at least one specification from the user.

The above data processing unit (100) can crawl using a platform of a job search site (e.g., Jobplanet) that has data of standardized job postings among various job search sites in order to prepare standardized data on job search and recruitment information. The data processing unit (100) can perform crawling appropriate to the situation by making exceptions for job postings that have insufficient information, although the information provided by Jobplanet is relatively standardized compared to other job search and recruitment platforms. The data processing unit (100) can perform data crawling using Selenium, a browser automation tool, and can collect all job postings for each job over a period of 2 to 3 days in order to collect a large amount of data.

The above data processing unit (100) may have a unified format, but may correct different speech styles and unnecessary information (stop word processing) for each job posting, and may have a data tokenization step to extract meaningful data and increase the accuracy of learning. According to one embodiment of the present invention, the data processing unit (100) may directly compare and analyze konlpy (Mecab, Kkma, Okt, etc.), a popular Korean morphological analyzer, and Kiwi tokenizer, a state-of-the-art technology, to find a tokenization method suitable for the project, and may then adopt Kiwi tokenizer, which utilizes more diverse part-of-speech tagging methods. In conclusion, an organic tokenization method was used by mixing Kiwi and mecab tokenizers, and Mecab tokenizer was used for Korean and English combinations that the kiwi tokenizer cannot recognize. The data processing unit (100) may store tokenized data as data for each of welfare and qualification requirements. In the future, the learning model can be set up in the learning unit (200) so that the user can select one of two criteria for similar jobs and receive recommendations for similar companies.

The above data processing unit (100) can perform vectorization of documents for similarity learning of documents having company names and job titles, if the company name and job title are considered as tags in the company's job posting and the job information is considered as a corpus. To this end, the data processing unit (100) can use Doc2Vec, which is a document embedding technology extended from the existing word embedding technology Word2Vec. In order to apply Doc2Vec to data, gensim, a Python library that provides most of the convenience functions required to convert natural language into vectors, can be used.

The above data processing unit (100) refines the accuracy of data preprocessing and can develop a system with the Flask framework so that users can conveniently use the model (vector model representing the similarity for each company's job posting) learned through Doc2Vec. Users can check well-organized job postings, visual data, and job posting similarity that can confirm the reliability of data through the web as a result implemented through the above visualization unit (300).

According to one embodiment of the present invention, the data processing unit (100) crawls all job postings registered in Jobplanet and may use Selenium's web driver to crawl dynamically implemented parts such as JavaScript of the Jobplanet site. Although Jobplanet provides relatively standardized information compared to other job search platforms, the number of types of information provided by each company is different, so data can be collected using exception handling.

The above data processing unit (100) can re-crawl the relevant page when an exceptional situation occurs. After crawling the data, the data can be divided into company name, job posting title, closing date, job position, career, employment type, salary, skills, company introduction, main duties, qualifications, preferential treatment, hiring process, and welfare benefits and saved as a CSV file.

The above data processing unit (100) can perform batch preprocessing (stop word processing + merge data with similar characteristics into one column) on the saved CSV file. For example, required skills, qualifications, and preferred items can be grouped into the 'qualifications' column because they have similar characteristics. For example, job posting title, job title, and main tasks can be grouped into the 'task' column because they have similar characteristics.

Since the above data processing unit (100) recognizes and stores the same job postings with different deadlines as different job postings, it is possible to exclude duplicate data when the 'company name + job posting title' is the same by using the drop_duplicates of the pandas module.

Since the above data processing unit (100) will also provide the user with raw data of the job posting when processing stop words, the job posting in its original state can be stored in a separate column.

The above data processing unit (100) can automatically replace \r with \n when a job posting contains the character "\r" when saving as a CSV, thereby causing a situation where the line is automatically turned over.

The above data processing unit (100) can compare tokenization techniques using multiple tokenizers and then use the tokenizer that tokenizes best (extracts important words from sentences well).

As a result of comparing Kiwi, Mecab, Kkma, and Okt, the Kiwi and Mecab tokenizers divide part-of-speech tags most diversely and are suitable for learning.

The above data processing unit (100) can tokenize stopword-processed data into ‘common nouns, proper nouns, English words, and roots’ using the kiwi tokenizer.

Kiwi tokenizer will raise an exception when it encounters a combination of Korean and English characters that it does not understand. For example, an exception can be raised when rigging TA (often shown in job postings for game developers).

The above data processing unit (100) can tokenize data that generates an exception in kiwi using the Mecab tokenizer.

The above data processing unit (100) can create two CSV files using the developed tokenizer. For example, a CSV file that tokenizes the welfare and task columns and a CSV file that tokenizes the qualifications and task columns can be created.

The reason why the above data processing unit (100) additionally tokenizes the task column when creating each CSV file is because two learning models will be created. For example, a model that has learned welfare/qualification requirements + a model that has learned tasks can be created.

The reason why the above data processing unit (100) uses two learning models is that it will recommend companies that have similar criteria among the 'welfare/qualification requirements' and the company entered by the user. However, if the 'welfare/qualification requirements' are similar but the actual job positions are different, the recommendation is meaningless to the user.

Users try to find jobs in their field of expertise, and there are usually limited types of jobs available in any given field.

The above data processing unit (100) performs learning on tokenized data using a Doc2Vec model, and at this time, the hyperparameters of the Doc2Vec model can be adjusted to adopt a value with the best accuracy.

The reason Doc2Vec was used as a machine learning method at this time is because, compared to the document embedding method using TensorFlow, it was judged that Doc2Vec would be more advantageous for learning in that it is faster and can be learned by adjusting multiple hyperparameters.

The above data processing unit (100) can preferentially extract the index of companies that are 70% or more similar to the company input by the user from the model that learned the task among the two learned models.

The reason why the above data processing unit (100) set the lower limit of job similarity to 70% is because the amount of data used when learning the job model (the model that learned the task column) was large (data on 'job posting title, job, and main tasks' were pushed into the task column), and even if the similarity was 70%, it was similar when the actual data was compared.

The above data processing unit (100) extracts similar companies from the companies extracted from the job model using the 'welfare/qualification requirements' model and outputs them in order of high similarity. For example, companies whose jobs are 70% or more similar and whose 'welfare/qualification requirements' are also similar can be extracted.

The learning unit (200) above can learn and calculate the correlation between companies satisfying the above business conditions and the corporate preference according to at least one of the above user specifications. The learning unit (200) can select and list companies satisfying the above business conditions. The learning unit (200) can list companies that the user is likely to prefer in order of the correlation between the corporate preference according to the user specifications from the list of companies satisfying the above business conditions.

The learning unit (200) above can match at least one user specification and a corporate preference. The learning unit (200) can learn by defining the correlation between the change in the user specification and the change in the corporate preference as a probability. The learning unit (200) can calculate the correlation between the corporate condition and the corporate preference according to at least one user specification according to the probability.

The above learning unit (200) can learn the correlation between the change in the weather condition attribute and the change in the condition of the target by defining it as a probability, and measure the correlation.

The above learning unit (200) can calculate the influence of corporate preference according to at least one of the above user specifications and the companies satisfying the above corporate conditions.

The above influence is a value that varies depending on how much the change in the user specifications affects the above corporate preference, and can be calculated by the following mathematical expression 1.

[Mathematical formula 1]

Here, p(x) is the probability that a change in preference x will occur. The closer the probability of occurrence is to 1, the closer the influence is to 0, so learning from fewer cases is required, which reduces the amount of computation. On the other hand, the closer the probability of occurrence is to 0, the greater the influence is, so learning from many cases is required, which can increase the amount of computation.

In the above learning unit (200), the computational load represents the expected value (average) of the influence. A larger computational load means a larger average influence. As the uncertainty increases, classification becomes more difficult, so the computational amount can be reduced by placing the smallest computational load at the upper decision-making node. The computational load can be obtained by calculating the following mathematical expression 2.

[Mathematical formula 2]

Here, S represents the collection of all events that have already occurred, and c can represent the number of events.

In the above learning unit (200), the correlation can be obtained by calculating the following mathematical expression 3.

[Mathematical Formula 3]

Here, C(S, P) is the correlation, E(S) is the analysis relevance value, S is the user specification, and P is the corporate preference. Here, P refers to a specific specification among the user specifications, and it can be used to determine which specification has the smallest computational load (i.e., the largest amount of information obtained) when classified.

The learning unit (200) can determine the ranking of the user specifications according to the correlation. The learning unit (200) can determine the ranking in order of decreasing correlation. The learning unit (200) can place the user specifications with low correlation in the recommendation priority.

The above user specifications may include at least one of school, major, gender, age, region, and MBTI.

The above learning unit (200) can determine a ranking including corporate preferences based on a combination of multiple user specifications.

The above visualization unit (300) can present a list of companies to the user according to the company ranking according to the correlation.

The above visualization unit (300) can provide a result according to the ranking of the correlation of corporate preferences according to a plurality of the above user specifications. The above visualization unit (300) can provide a result according to the ranking of the correlation of corporate preferences according to a plurality of combinations of the above user specifications.

The visualization unit (300) above can provide data visualized to the user so that the user can easily check which parts are visually similar when the user receives recommendations for similar companies and checks the information of the recommended company announcements. The visualization unit (300) can use a word cloud method that visualizes key words that frequently appear in the content of each company announcement in various ways. The visualization unit (300) can also provide the user with a function to search for similar companies in an intuitive and interesting way by showing similar companies on an additional web page on a two-dimensional coordinate plane using TSNE (a technique for reducing the dimension of high-dimensional vector data).

The above visualization part (300) can output the system developed up to the 'data embedding' stage by loading it on a web server using Flask. Data transmission between the Backend and the Frontend can be performed by transmitting data to the Frontend as a parameter when calling the render_template function, which shows an html file to the client, in the Backend. The data received from the Backend can be displayed in components placed on the html using JavaScript and jQuery in the Frontend. Data transmission from the Frontend to the web server can be implemented by using the get method of the form tag, or by including data in the URL called from the Frontend and transmitting it, and then processing and using it in the Backend.

The UI of the web can be composed intuitively and concisely using various Bootstrap elements (ex. progressbar, radio button, table, card).

If the user needs to wait for the system to perform some operation, such as when the system is learning something new or generating a visualization, you can use a progressbar to let the user know what operation is in progress.

The composition of HTML elements can be designed in terms of layout and size using CSS grammar in the style tag of <head>.

In the case of html files that display visualization information, since newly changed information must be provided, not data cached in the web browser, caching can be prevented using Cache-Control, and when calling a visualization image, a random number is added to the URL and transmitted to the server, so that even if an image file with the same name is overwritten with other visualization information, the changed image can be provided to the user.

An example web page might look like this:

Start page: hello.html,

Job posting link input page: apply.html,

Loading page: index.html,

Analysis results page: result.html,

Company job posting details page: linktoimg.html,

Page that reports that there is no 'welfare' information: nobokri.html,

A page that reports that there are no similar businesses: nolist.html,

Page showing visualization information: display_plot.html

Users enter the Jobplanet job posting link in apply.html, and then enter the criteria (welfare or qualification requirements) by which they would like to receive recommendations for similar companies.

The above visualization section (300) extracts similar companies from the learned model based on the links and criteria entered by the user on the web server and provides the results to the user in a table format through result.html.

The above visualization section (300) provides the user with a WordCloud result that analyzes keywords of the job posting of the company by clicking [View Details] of each company in result.html, and the raw job posting content that has not been preprocessed.

In the case where the above visualization section (300) orders a recommendation for welfare benefits when there is no information about ‘welfare benefits’ in the link entered by the user, the system requests the user to re-enter the link by showing nobokri.html.

In the above visualization section (300), if there is no company similar to the company in the link entered by the user (if there is no company with similar job and criteria), the system displays nolist.html to inform the user that there is no similar company, and recommends other companies that are not similar in 'input criteria' but have similar 'jobs'.

In the above visualization section (300), the user can move from result.html to display_plot.html, which shows visualization information.

The above visualization part (300) uses Plotly.js to draw a scatter plot on HTML using TSNE coordinate data received from the server in display_plot.

The above visualization section (300) shows two plots at this time: a scatter plot visualizing companies with similar user input companies and ‘user-selected criteria’ and a scatter plot visualizing companies with similar ‘jobs’.

By clicking on the marker of each plot, the user is directed to linktoimg.html and provided with the job posting information of the company corresponding to the marker.

The above visualization section (300) provides not only the WordCloud results and raw job posting content in linktoimg.html, but also ‘similarity’ according to each criterion so that the user can directly check the reliability of the data recommended by the system.

In addition, linktoimg.html provides information on which marker the user clicks on the 'Visualization Results' page. If the user clicks on the red marker, it informs them that it is the job posting they entered, and if the user clicks on the blue marker, it provides information on 'what criteria' the job posting was analyzed to be similar to the user-entered job posting.

The embedding vector of each job posting is a 200-dimensional high-dimensional data.

The above visualization unit (300) reduces data to two dimensions using the TSNE module and then provides the data by visualizing it on a two-dimensional plane in order to provide visualized information to the user.

The above visualization unit (300) does not visualize information on all companies at this time, but only visualizes and provides information on companies (the most similar companies) that are a certain distance away from the user input company.

The above visualization unit (300) performs visualization by slightly changing the position of one of the two markers so that they do not overlap, in order to prevent a phenomenon in which the positional difference between any two markers among the visualized information is so minimal that they appear as one marker.

The above visualization section (300) provides a WordCloud module by creating tokenized welfare/qualification data using keywords.

At this time, if the user selects ‘welfare benefits’ as the criterion for judging similarity, the WordCloud result created with ‘welfare benefits’ tokenized data is shown, and if the user selects ‘qualification requirements’, the WordCloud result created with ‘qualification requirements’ tokenized data is shown.

FIG. 2 is a screen during analysis of a matching page to which a job search matching method according to an embodiment of the present invention is applied, FIG. 3 is a matching result screen to which a job search matching method according to an embodiment of the present invention is applied, FIG. 4 is a visualization result screen to which a job search matching method according to an embodiment of the present invention is applied, FIG. 5 is a result screen when there is no desired information on a matching page to which a job search matching method according to an embodiment of the present invention is applied, FIG. 6 is a result screen when there is no matching company as a result of which a job search matching method according to an embodiment of the present invention is applied, and FIG. 7 is a matching result screen to which a job search matching method according to an embodiment of the present invention is applied.

Referring to FIGS. 2 to 7, according to one embodiment of the present invention, the data processing unit (100) copies the URL of a job posting of a desired company from a job search site (e.g., Jobplanet) and inputs it into the system, and then when the user selects an item he or she likes among the welfare and qualification requirements of the posting, the job posting of the URL input by the user is collected from the job search site, and the visualization unit (300) analyzes the company and the job posting and outputs similar companies in a table format on a web page. While performing calculations through the model, a loading screen can be displayed to the user, and the analysis results can be provided after the calculations are completed.

By clicking [View Details] in the 'Details' column of the companies shown in the table in the visualization section (300) above, the user can be provided with the actual job posting content of the company and the WordCloud analysis results that analyze the content by keyword.

If you click [Go to visualization data] at the bottom of the analysis results in the table format above, you can create a two-dimensional visualization of companies with similar [jobs] to the company entered by the user and companies with similar [criteria entered by the user] and provide the user with the visualization data.

In the visualization data of the above visualization section (300), the red dots may represent the company whose URL the user entered, and the blue dots may represent companies analyzed as being similar to the company of the red dots.

The above visualization section (300) can provide the user with [View Details] information about the company corresponding to the point even when the points in the above visualization data are clicked.

The above visualization section (300) can display the actual content of the job posting, the keyword WordCloud analysis results, and the percentage of similarity with the company entered by the user on the page providing detailed information on the company job posting.

If the job posting in the URL entered by the user does not contain ‘welfare’ information and the user requests the results analyzed based on ‘welfare’, the visualization unit (300) may notify that there is no welfare information in the link and request the user to re-enter the URL.

The above visualization unit (300) can adjust the loading time and show the user what operation is being performed when the URL entered by the user is new data that the model has not learned, because it takes more time to perform additional learning.

The visualization unit (300) above can inform the user if there is no similar company in the analysis result (if there is no company with similar job and user input criteria). The visualization unit (300) can then provide a link at the bottom right to check companies with similar jobs. If the user clicks on the link, a list of similar companies in table format is output, but [companies with similar jobs] is specified so that the user can recognize the criteria by which the analysis was conducted.

Figure 8 is a flow chart of a job search matching method according to one embodiment of the present invention.

Referring to FIG. 8, a job search matching method according to an embodiment of the present invention may include a step (S110) of receiving input of at least one desired company condition and user specification from a user.

In step S110, the data processing unit (100) can receive at least one desired corporate condition and user specification from the user. The data processing unit (100) can collect corporate job search information through job search and recruitment sites, etc. The data processing unit (100) can crawl and collect corporate job search and recruitment information from job search and recruitment sites. The data processing unit (100) can receive at least one desired corporate condition from the user. The data processing unit (100) can receive at least one specification from the user.

In step S110, the data processing unit (100) can crawl using a platform of a job search site (e.g., Jobplanet) that has data of standardized job postings among various job search sites in order to prepare standardized data on job search information. The data processing unit (100) can perform crawling appropriate to the situation by making exceptions for job postings that have insufficient information, although the information provided by Jobplanet is relatively standardized compared to other job search platforms. The data processing unit (100) can perform data crawling using Selenium, a browser automation tool, and can collect all job postings for each job over a period of 2 to 3 days in order to collect a large amount of data.

In step S110, the data processing unit (100) may have a data tokenization step so that although the format is unified, different speech styles and unnecessary information for each job posting can be corrected (stopword processing) and meaningful data can be extracted and the accuracy of learning can be increased. According to one embodiment of the present invention, the data processing unit (100) may directly compare and analyze konlpy (Mecab, Kkma, Okt, etc.), a popular Korean morphological analyzer, and Kiwi tokenizer, a state-of-the-art technology, to find a tokenization method suitable for the project, and then adopt Kiwi tokenizer, which uses more diverse part-of-speech tagging methods. In conclusion, an organic tokenization method was used by mixing Kiwi and mecab tokenizers, and Mecab tokenizer was used for Korean and English combinations that cannot be recognized by Kiwi tokenizer. The data processing unit (100) may store tokenized data as data for welfare benefits and qualification requirements, respectively. In the future, the learning model can be set up in the learning unit (200) so that the user can select one of two criteria for similar jobs and receive recommendations for similar companies.

In step S110, the data processing unit (100) may perform vectorization of documents for learning the similarity of documents having the company name and job titles, if the company name and job title are considered as tags in the company's job posting and the job information is considered as a corpus. To this end, the data processing unit (100) may use Doc2Vec, which is a document embedding technology extended from the existing word embedding technology Word2Vec. In order to apply Doc2Vec to our data, gensim, a Python library that provides most of the convenience functions required to convert natural language into vectors, may be used.

In step S110, the data processing unit (100) refines the accuracy of data preprocessing and develops a system using the Flask framework so that users can conveniently use the model (vector model representing the similarity for each company's job posting) learned through Doc2Vec. Users can check well-organized job postings, visual data, and job posting similarity that can confirm the reliability of data through the web as a result implemented through the visualization unit (300).

In step S110, according to one embodiment of the present invention, the data processing unit (100) crawls all job postings registered in Jobplanet and may use Selenium's web driver to crawl dynamically implemented parts such as JavaScript of the Jobplanet site. Although Jobplanet provides relatively standardized information compared to other job search platforms, the number of types of information provided by each company is different, so data can be collected using exception handling.

In step S110, the data processing unit (100) can re-crawl the page when an exceptional situation occurs. After crawling the data, the data can be divided into company name, job posting title, closing date, job position, career, employment type, salary, skills, company introduction, main job, qualification requirements, preferential treatment, hiring process, and welfare benefits and saved as a CSV file.

In step S110, the data processing unit (100) can perform batch preprocessing (stop word processing + merge data with similar characteristics into one column) on the saved CSV file. For example, required skills, qualifications, and preferential conditions can be grouped into the 'qualifications' column because they have similar characteristics. For example, job posting title, job title, and main tasks can be grouped into the 'task' column because they have similar characteristics.

In step S110, the data processing unit (100) recognizes and stores the same job postings with different deadlines as different job postings, so the drop_duplicates function of the pandas module can be used to exclude duplicate data when the 'company name + job posting title' is the same.

In step S110, since the data processing unit (100) will also provide the user with raw data of the job posting when processing stop words, the original job posting can be stored in a separate column.

At step S110, the data processing unit (100) can automatically replace \r with \n when the job posting contains the character "\r" when saving as a CSV file, thereby causing a situation in which a line is automatically turned over.

In step S110, the data processing unit (100) can compare tokenization techniques using multiple tokenizers and then use the tokenizer that tokenizes best (extracts important words from sentences well).

Comparing Kiwi, Mecab, Kkma, and Okt, we found that Kiwi and Mecab tokenizers divide part-of-speech tags most diversely and are suitable for learning.

In step S110, the data processing unit (100) can tokenize stopword-processed data into ‘common nouns, proper nouns, English words, and roots’ using the kiwi tokenizer.

Kiwi tokenizer will raise an exception when it encounters a combination of Korean and English characters that it does not understand. For example, an exception can be raised when rigging TA (often shown in job postings for game developers).

At step S110, the data processing unit (100) can tokenize data that generates an exception in kiwi using the Mecab tokenizer.

In step S110, the data processing unit (100) can create two CSV files using the developed tokenizer. For example, a CSV file that tokenizes the welfare and task columns and a CSV file that tokenizes the qualifications and task columns can be created.

In step S110, the data processing unit (100) additionally tokenizes the task column when creating each CSV file because two learning models will be created. For example, a model that has learned welfare/qualification requirements + a model that has learned tasks can be created.

In step S110, the reason why the data processing unit (100) uses two learning models is that it will recommend companies that have similar criteria among the 'welfare/qualification requirements' and the company input by the user. However, if the 'welfare/qualification requirements' are similar but the actual job positions are different, the recommendation is meaningless to the user.

Users try to find jobs in their field of expertise, and there are usually limited types of jobs available in any given field.

The above data processing unit (100) performs learning on tokenized data using a Doc2Vec model, and at this time, the hyperparameters of the Doc2Vec model can be adjusted to adopt a value with the best accuracy.

The reason Doc2Vec was used as a machine learning method at this time is because, compared to the document embedding method using TensorFlow, it is faster and it is possible to learn by adjusting multiple hyperparameters, so Doc2Vec was judged to be more advantageous for learning.

At step S110, the data processing unit (100) can preferentially extract the index of companies that are 70% or more similar to the company input by the user from the model that learned the task among the two learned models.

In step S110, the reason why the data processing unit (100) set the lower limit of the similarity of the job to 70% is because the amount of data used when learning the job model (the model that learned the task column) was large (data on 'job posting title, job, and main tasks' were pushed into the task column), and even if the similarity was 70%, it was similar when compared to the actual data.

In step S110, the data processing unit (100) extracts similar companies from the companies extracted from the job model using the 'welfare/qualification requirements' model and outputs them in order of high similarity. For example, companies whose jobs are 70% or more similar and whose 'welfare/qualification requirements' are also similar can be extracted.

A job search matching method according to one embodiment of the present invention may include a step (S120) of learning and calculating a correlation between companies satisfying the above-mentioned company conditions and company preferences according to at least one of the above-mentioned user specifications.

In step S120, the learning unit (200) can learn and calculate the correlation between companies satisfying the company conditions and the company preferences according to at least one of the user specifications. The learning unit (200) can select and list companies satisfying the company conditions. The learning unit (200) can list companies that the user may prefer in order of the correlation between the company preferences according to the user specifications from the list of companies satisfying the company conditions.

In step S120, the learning unit (200) can match at least one user specification with a corporate preference. The learning unit (200) can learn by defining the correlation between the change in the user specification and the change in the corporate preference as a probability. The learning unit (200) can calculate the correlation between the corporate condition and the corporate preference according to at least one user specification according to the probability.

In step S120, the learning unit (200) can learn the correlation between the change in the weather condition attribute and the change in the condition of the target by defining the correlation as a probability and measure the correlation.

At step S120, the learning unit (200) can calculate the influence of corporate preference according to at least one of the user specifications and the corporate conditions that satisfy the corporate conditions.

At step S120, the learning unit (200) calculates the influence as a value that varies depending on how much the change in the user specifications affects the corporate preference, using the following mathematical expression 1.

[Mathematical formula 1]

Here, p(x) is the probability that a change in preference x will occur. The closer the probability of occurrence is to 1, the closer the influence is to 0, so learning from fewer cases is required, which reduces the amount of computation. On the other hand, the closer the probability of occurrence is to 0, the greater the influence is, so learning from many cases is required, which can increase the amount of computation.

In step S120, the computational load in the learning unit (200) represents the expected value (average) of the influence. A larger computational load means a larger average influence, and the greater the uncertainty, the more difficult it is to classify. Therefore, the computational amount can be reduced by placing the smallest computational load at the upper decision-making node. The computational load can be obtained by calculating the following mathematical expression 2.

[Mathematical formula 2]

Here, S represents the collection of all events that have already occurred, and c can represent the number of events.

At step S120, the correlation can be obtained by calculating the following mathematical expression 3 in the learning unit (200).

[Mathematical Formula 3]

Here, C(S, P) is the correlation, E(S) is the analysis relevance value, S is the user specification, and P is the corporate preference. Here, P refers to a specific specification among the user specifications, and it can be used to determine which specification has the smallest computational load (i.e., the largest amount of information obtained) when classified.

A job search matching method according to one embodiment of the present invention may include a step (S130) of determining the ranking of companies satisfying the company conditions according to the correlation.

In step S130, the learning unit (200) can determine the ranking of the user specifications according to the correlation. The learning unit (200) can determine the ranking in order of decreasing correlation. The learning unit (200) can place the user specifications with low correlation in the recommendation priority.

The above user specifications may include at least one of school, major, gender, age, region, and MBTI.

At step S130, the learning unit (200) can determine a ranking including corporate preferences based on a combination of multiple user specifications.

A job search matching method according to one embodiment of the present invention may include a step (S140) of presenting a list of companies to the user according to the company ranking according to the correlation.

At step S140, the visualization unit (300) can present a list of companies to the user according to the company ranking according to the correlation.

In step S140, the visualization unit (300) can provide a result according to the ranking of the correlation of corporate preferences according to a plurality of the user specifications. The visualization unit (300) can provide a result according to the ranking of the correlation of corporate preferences according to a plurality of combinations of the user specifications.

In step S140, the visualization unit (300) can provide data as a visualization so that the user can easily check which parts are visually similar when the user receives a recommendation for a similar company and checks the information of the recommended company announcement. The visualization unit (300) can use a Wordcloud method that variously visualizes key words that frequently appear in the content of each company announcement. The visualization unit (300) can also provide the user with a function to search for similar companies in an intuitive and interesting way by showing similar companies on an additional web page on a two-dimensional coordinate plane using TSNE (a technique for reducing the dimension of high-dimensional vector data).

In step S140, the visualization unit (300) can output the system developed up to the 'data embedding' step by loading it on a web server using Flask. Data transmission between the Backend and the Frontend is performed by transmitting data to the frontend as a parameter when the backend calls the render_template function that shows an html file to the client. The data received from the Backend can be displayed in components placed on the html using JavaScript and jQuery in the Frontend. Data transmission from the Frontend to the web server can be implemented by using the get method of the form tag, or by including data in the URL called from the frontend and transmitting it, and then processing and using it in the backend.

The UI of the web can be composed intuitively and concisely using various Bootstrap elements (ex. progressbar, radio button, table, card).

If the user needs to wait for the system to perform some operation, such as when the system is learning something new or generating a visualization, you can use a progressbar to let the user know what operation is in progress.

The composition of HTML elements can be designed in terms of layout and size using CSS grammar in the style tag of <head>.

In the case of html files that display visualization information, since newly changed information must be provided, not data cached in the web browser, caching can be prevented using Cache-Control, and when calling a visualization image, a random number is added to the URL and transmitted to the server, so that even if an image file with the same name is overwritten with other visualization information, the changed image can be provided to the user.

An example web page might look like this:

Start page: hello.html,

Job posting link input page: apply.html,

Loading page: index.html,

Analysis results page: result.html,

Company job posting details page: linktoimg.html,

Page that reports that there is no 'welfare' information: nobokri.html,

A page that reports that there are no similar businesses: nolist.html,

Page showing visualization information: display_plot.html

Users enter the Jobplanet job posting link in apply.html, and then enter the criteria (welfare or qualification requirements) for which they would like to receive recommendations for similar companies.

At step S140, the visualization unit (300) extracts similar companies from the learned model based on the links and criteria entered by the user on the web server and provides them to the user in a table format through result.html.

At step S140, the visualization unit (300) provides the user with a WordCloud result that analyzes keywords of the job posting of the corresponding company and the raw job posting content that has not been preprocessed by clicking [View Details] of each company in result.html.

At step S140, if the visualization unit (300) orders a recommendation for welfare benefits when there is no information about ‘welfare benefits’ in the link entered by the user, the system requests the user to re-enter the link by showing nobokri.html.

At step S140, if there is no company similar to the company of the link input by the user (if there is no company with similar job and criteria), the system displays nolist.html to inform the user that there is no similar company, and recommends other companies that are not similar in 'input criteria' but have similar 'jobs'.

At step S140, the visualization part (300) allows the user to move from result.html to display_plot.html, which shows visualization information.

At step S140, the visualization unit (300) draws a scatter plot on HTML using Plotly.js with TSNE coordinate data received from the server in display_plot.

At step S140, the visualization unit (300) displays two plots at this time: a scatter plot visualizing companies with similar user input companies and ‘user-selected criteria’ and a scatter plot visualizing companies with similar ‘jobs’.

By clicking on the marker of each plot, the user is directed to linktoimg.html and provided with the job posting information of the company corresponding to the marker.

At step S140, the visualization unit (300) provides not only the WordCloud results and raw job posting content in linktoimg.html, but also the ‘similarity’ according to each criterion so that the user can directly check the reliability of the data recommended by the system.

In addition, linktoimg.html provides information on which marker the user clicks on the 'Visualization Results' page. If the user clicks on the red marker, it informs them that it is the job posting they entered, and if the user clicks on the blue marker, it provides information on 'what criteria' the job posting was analyzed to be similar to the user-entered job posting.

The embedding vector of each job posting is a 200-dimensional high-dimensional data.

At step S140, the visualization unit (300) reduces the data to two dimensions using the TSNE module to provide visualized information to the user, and then provides it by visualizing it on a two-dimensional plane.

At step S140, the visualization unit (300) does not visualize information on all companies, but only visualizes and provides information on companies (the most similar companies) that are a certain distance away from the user input company.

In step S140, the visualization unit (300) performs visualization so that, when the distance between two arbitrary markers is extremely small, the position difference between the two markers is so small that they appear as one marker, by slightly changing the position of one of the two markers so that they do not overlap.

In step S140, the visualization unit (300) creates a WordCloud module using tokenized welfare/qualification data as keywords and provides it.

At this time, if the user selects ‘welfare’ as the criterion for judging similarity, the WordCloud result created with ‘welfare’ tokenized data is shown, and if the user selects ‘qualification’, the WordCloud result created with ‘qualification’ tokenized data is shown.

The present invention has been described with reference to preferred embodiments thereof. Those skilled in the art will appreciate that the present invention may be implemented in modified forms without departing from the essential characteristics of the present invention. Therefore, the disclosed embodiments should be considered from an illustrative rather than a restrictive perspective. The scope of the present invention is indicated by the claims, not the foregoing description, and all differences within the scope equivalent thereto should be interpreted as being included in the present invention.

Claims (10)

A step for receiving at least one desired corporate condition and user specification from the user;
A step of learning and calculating a correlation between companies satisfying the above business conditions and company preferences according to at least one of the above user specifications;
A step of determining the ranking of companies satisfying the above corporate conditions according to the above correlation; and
Including a step of presenting a list of companies to the user according to the ranking of the companies according to the correlation,
The step of entering at least one of the above business conditions and user specifications is:
We embed job postings from multiple companies that include the company whose link the user entered into a 200-dimensional embedding vector.
The steps presented to the above user are:
After reducing the above embedding vector to two dimensions, only the company of the link entered by the user and the companies located a certain distance away from the company of the link entered by the user are visualized as points on a two-dimensional coordinate plane - the company of the link entered by the user is expressed as a red dot on the coordinate plane, and the companies located a certain distance away are expressed as blue dots -
A job search matching method that provides the user with, when the user clicks on any point on the two-dimensional coordinate plane, the similarity between the job posting content of the company corresponding to the point and the company of the link entered by the user - the similarity being the result of a word cloud created using tokenized data of one of the welfare benefits or qualification requirements of the job posting. In the first paragraph,
The above correlation operation step is,
A step of matching at least one user specification with a corporate preference;
A step of learning by defining the correlation between the change in the above user specifications and the change in the above corporate preference as a probability; and
A job search matching method characterized by including a step of calculating a correlation between the corporate conditions and corporate preferences according to at least one user specification according to the above probability. In the second paragraph,
The above correlation operation step is,
A job search and recruitment matching method, characterized in that the above correlation is calculated by the following mathematical formula 1.
[Mathematical Formula 1]

(Here, C(S, P) is the correlation, E(S) is the analysis relevance value, S is the user specification, and P is the corporate preference.) In the first paragraph,
The above user specifications are,
A job search matching method characterized by including at least one of school, major, gender, age, region, and MBTI. In paragraph 4,
The steps to determine the ranking of the above companies are:
Determine a ranking that includes corporate preferences based on a combination of multiple above user specifications,
The above user presentation steps are:
A job search matching method characterized by providing results according to the above ranking. Data processing unit that receives at least one desired corporate condition and user specification from the user;
A learning unit that learns and calculates correlations between companies satisfying the above business conditions and company preferences according to at least one of the above user specifications;
A ranking unit that determines the ranking of companies that satisfy the above corporate conditions according to the above correlation; and
Including a visualization section that presents a list of companies to the user according to the ranking of companies according to the correlation;
The above data processing unit,
We embed job postings from multiple companies that include the company whose link the user entered into a 200-dimensional embedding vector.
The above visualization section,
After reducing the above embedding vector to two dimensions, only the company of the link entered by the user and the companies located a certain distance away from the company of the link entered by the user are visualized as points on a two-dimensional coordinate plane - the company of the link entered by the user is expressed as a red dot on the coordinate plane, and the companies located a certain distance away are expressed as blue dots -
A job search matching system characterized in that, when the user clicks on any point on the two-dimensional coordinate plane, the similarity between the job posting content of the company corresponding to the point and the company of the link entered by the user is provided to the user - the similarity being a word cloud result created using tokenized data of one of the welfare benefits or qualification requirements of the job posting. In Article 6,
The above learning department,
A job search matching system characterized by matching at least one user specification and a corporate preference, defining a correlation between a change in the user specification and a change in the corporate preference as a probability and learning it, and calculating a correlation between the corporate condition and the corporate preference according to at least one user specification according to the probability. In Article 7,
The above learning department,
A job search matching system characterized in that the above correlation is calculated by the following mathematical formula 1.
[Mathematical formula 1]

(Here, C(S, P) is the correlation, E(S) is the analysis relevance value, S is the user specification, and P is the corporate preference.) In Article 8,
The above user specifications are,
A job search matching system characterized by including at least one of school, major, gender, age, region, and MBTI. In Article 9,
The above learning department,
Determine a ranking that includes corporate preferences based on a combination of multiple above user specifications,
The above visualization section,
A job search matching system characterized by providing results according to the above ranking.