Machine Learning-Based Recommendations and Classification System for Unstructured Resume Documents

Resume screening represents a crucial step within a company's hiring process, permitting evaluation of potential employees prior to their recruitment. Furthermore, it facilitates the preparation of candidates for comprehensive examination [1]. The ultimate objective of resume screening is to identify suitable individuals for job vacancies, while concurrently preventing the company from overlooking exceptional candidates [1]. However, manual screening of resumes is a time-consuming and complex process, particularly given that each job vacancy typically attracts hundreds of resumes [2].

Although resumes generally adhere to a standard format, the specific skills and qualifications demanded by each role may necessitate varying levels of specificity in their creation [3]. The majority of contemporary job applications require electronic resumes. However, the development of resume screening approaches based on modern methodologies, knowledge discovery, and social networks remains in its infancy, with substantial research required prior to the implementation of commercial systems [4].

One notable limitation of resume screening is its reliance on candidate-provided content entered into job portals or electronic forms during the job application process. While job seekers may possess unique attributes that distinguish them, discrepancies may arise if the data entered does not match the content of the candidate's resume. At its most effective, resume screening could yield a list of individuals with similar resumes, and vice versa, but this could introduce significant bias and inaccuracies particularly in the case of numerous unsupervised searches [5]. Therefore, it is imperative that uploaded resume documents are screened for an effective job-matching profile. Consequently, the extraction of useful data from these documents is essential. In this context, Natural Language Processing (NLP) can be employed to extract job-matching attributes from the resume document [6]. The extracted attribute values can then be compared to the skill set requirements for recruitment, facilitating the identification of relevant jobs with commensurate salaries.

What is Natural Language Processing?

Natural Language Processing (NLP) is a process of breaking down the given input into component parts and interpreting their meaning. The process includes converting the input into a form that a computer can understand [7]. NLP can be employed to extract high-quality information from documents and queries by developing the computer’s ability to understand and communicate with humans in the human language.

The rising use of social media applications such as Facebook, LinkedIn, Twitter etc. for job-seekers’ screening activities is likely to propel the growth of the NLP market. NLP being a demanded technique is used by researchers in extracting the most relevant information, identifying negative character traits, and more. In the recruitment industry resumes need to be screened, and the process requires to be automated, and NLP is one among the appropriate method to interpret the resume [8]. Even well-written resumes will never be 100% successful without proper interpretation or extraction. The research is motivated by several factors, including the exponential growth in the job market and the subsequent increase in the number of job applications, which has made the recruitment process more complex and time-consuming. Additionally, the unstructured data in the form of pdf resumes has made it even more challenging to extract structured information to automate the screening process [9].

The proposed system is motivated by the need to ease the recruitment process by providing an automated solution to screen candidates, categorize them based on their skills, and recommend job updates or rewrites. This will help recruiters to efficiently and accurately identify the most suitable candidates for a job.

Furthermore, the proposed system is also motivated by the need to reduce the mismatch between candidate skills and the skills demanded by industries, which can lead to ineffectiveness in all phases of the recruitment process. The system aims to ensure that the skills of the candidates are aligned with the skills required by the industry, resulting in more effective job candidate selections.

Overall, the research is motivated by the need to improve the recruitment process by providing an automated and accurate system to screen candidates, categorize them, and provide job recommendations. The proposed system is expected to improve the efficiency and effectiveness of the recruitment process, resulting in more successful job placements.

This paper utilizes NLP for effective data extraction from resume documents. NLP helps in analyzing every aspect of a resume and helps not only to screen the appropriate resume but also helps in recommending candidates to update or rewrite the resume if it is not on par with a standard resume. The satisfactory resumes are then classified according to the job description.

The aim of the proposed system is to extract the data from the pdf files and to classify the resume as satisfactory and unsatisfactory resumes, then recommend rewriting to further strengthen the unsatisfactory resume and classify the satisfactory resumes based on the skill set.

3.1 Overview of the proposed system

The overview of the proposed system is depicted in Figure 1. The system can be fragmented into three major modules.

1. To Extract or scrape the essential data from the resume, that is in pdf format:

For this task ‘PyPDF’ package is used. PyPDF is a Python library that allows for the manipulation of PDF files. PyPDF provides functionality for extracting information from PDF documents by analyzing their internal structure.

1.png

Figure 1. Resume recommendation and classification system overview

2.png

Figure 2. NLP model

To extract information from a PDF document using PyPDF, one would typically follow the following steps:

1. Open the PDF file using PyPDF.
2. Access the desired page or pages within the document.
3. Extract text or other content from the page(s).
4. Close the PDF file.

To accomplish these steps, PyPDF provides a variety of operational codes that can be used to interact with PDF documents. Some of the key operational codes for extracting information from PDF documents using PyPDF include:

• PdfFileReader: A class that represents a PDF file and provides functionality for reading and parsing the file.
• getPage: A method of the PdfFileReader class that allows for the selection of a specific page within the document.
• extractText: A method of the PdfFileReader class that extracts text from a selected page.
• numPages: A property of the PdfFileReader class that provides the number of pages in the PDF document.

By using these operational codes, PyPDF enables users to extract information from PDF documents and perform a variety of other operations on the files.

The first step is to install the 'PyPDF' package, once the package is installed, the next step is to import the required libraries and open the PDF file using the 'PyPDF' package.

Next, we need to extract the text from the PDF file. We can do this by iterating through each page of the PDF file and extracting the text using the 'extractText()' function.

After the text is extracted, it may contain unwanted characters or formatting, which needs to be cleaned. Pre-processing can be done to clean the extracted text. This may include removing special characters, removing stop words, and tokenizing the text.

In conclusion, the 'PyPDF' package is a useful tool to extract essential data from resumes in PDF format. Once the text is extracted, it can be further processed to clean and refine the data, making it useful for various applications, including candidate screening and job matching.

2. To classify the resume into satisfactory and unsatisfactory resumes:

The data extracted from the pdf is pre-processed so that unnecessary data that does not impact the result is removed. The important function of this pre-processing is to extract only useful data. NLP (Natural Language Processing) is utilized in pre-processing. The process of pre-processing using NLP is depicted in Figure 2 and the algorithm is discussed as follows:

Algorithm for Pre-processing the data extracted using PyPDF

_______________________________________________

Input: Text extracted from a PDF document using the PyPDF package

Output: Pre-processed text that is cleaned and refined for further use

________________________________________________

1. Remove special characters from the text using regular expressions
   • Initialize a regular expression pattern to match all non-alphanumeric characters
   • Use the re.sub() function to replace all matches with a space character
2. Convert the text to lowercase
   • Use the lower() method to convert all characters to lowercase
3. Remove stop words from the text using the nltk package
   • Initialize a set of stop words using the stopwords.words('english') method
   • Tokenize the text using the word_tokenize() function from the nltk package
   • Use a list comprehension to remove all stop words from the tokenized text
4. Tokenize the text using the nltk package
   • Use the word_tokenize() function to split the text into individual tokens
5. Perform stemming using the Porter Stemmer algorithm from the nltk package
   • Initialize a PorterStemmer() object
   • Use a list comprehension to apply the stemming algorithm to each token in the tokenized text
6. Return the pre-processed text as a list of stemmed tokens

_________________________________________________

Further, the classification is performed using machine learning models. Four benchmark models namely logistic regression [18], Gaussian Naïve Bayes [19], SVM, and Decision tree Classifier [20] are trained using a set of satisfactory and unsatisfactory resumes. Once the model is trained, it is used for the classification of the resumes as satisfactory and unsatisfactory. Satisfactory resumes are considered for job recommendations and unsatisfactory resumes are recommended for upgrade or rewrite. The satisfaction of a training structure in machine learning approaches can be judged based on several criteria, including:

1. Accuracy: This measure how well the model predicts the correct class labels for the test data. Higher accuracy indicates better performance.
2. Precision: This is the ratio of true positive predictions to the total number of positive predictions. It measures the ability of the model to identify true positives and avoid false positives.
3. Recall: This is the ratio of true positive predictions to the total number of actual positive instances in the test data. It measures the ability of the model to identify all positive instances, including those that might be missed.
4. F1 Score: This is the harmonic mean of precision and recall. It provides a balanced measure of both metrics and is a good indicator of overall performance.

The results of the models are recorded for accuracy, Precision, Recall, and F1Score and later compared in the results section.

The output from the NLP model is fed to the machine learning models. To better understand the standard of the resume, a word cloud is generated. Word cloud is a way of visualizing content or data. The most frequently used words in the resume will be generated as a word cloud for a better understanding of what skill set the candidate has. Figure 3 shows the word cloud of an unsatisfactory resume where the recruiter could not come to a conclusion on the expertise of the candidate, whereas the word cloud depicted in Figure 4 clearly showcases the expertise of the applicant.

3.png

Figure 3. Word cloud generated for an unsatisfactory resume

4.png

Figure 4. Word cloud generated for a satisfactory resume

3. Classify satisfactory resumes for job recommendations based on skill set:

The skill set of the satisfactory resumes extracted using NLP is stored and the distinct skills are identified. For instance, the skillset required for Business analytics is extracted from the set of business analytics resumes as [‘Tableau’, ’Pandas’, ‘Artificial Intelligence’, ’Pyspark’, ‘IBM Watson’, ’Sklearn’, ‘R programming’, ‘Keras’, 'Map Reduce’, ’Modelling’, ’Regression’, ‘Dundas BI’, ’Patterns’, ’Data Mining’, ‘Plotly’, ’Text Mining’, ’Oops’, ’Deep Learning’, ’Web Analytics’, ’Time Series’, ’Regression’, ’Tensorflow’, ’Azure’, ’Linear Regression’, ’Logistic Regression’, ’Decision Tree’, ’Random Forest’, ’Data Structure’, ‘Keras’].  The skillset likely grows with the increase in the number of resumes used for training. As more resumes for a specific job role are used for training the model becomes more effective. The resumes are categorized based on the extracted data for each job description and the distribution of the categories is visualized in the results section. The job recommendations are done based on the category of the resume. The job description or requirement is used as the input and the appropriate candidate’s resumes are fetched as recommendations based on the extracted info. The performance of the models is measured using accuracy, where accuracy is the number of correct predictions among the total predictions made, it is calculated using Eq. (1) as below.

Accuracy $=\frac{\text { Number of predictions made correctly }}{\text { total number of predictions made }}$           (1)

The percentage of accurate predictions made by the model is measured by accuracy. Accuracy alone might not be sufficient in cases with unbalanced data, where the proportion of instances in each class differs dramatically.

What percentage of the positive forecasts is actually positive, according to a criterion called precision? It shows how well the model is determining a class's true positive values.

What percentage of all the real positive records are accurately identified, according to recall? It shows how successfully a model can locate each instance that belongs to a class. Recall is calculated for each class in multiclass classification tasks and then averaged using either a macro-average or a micro-average.

Recall is a metric used to measure the ability of a model to identify all relevant instances of a particular class. Macro-average and micro-average are two ways to calculate the average recall score across all the classes in a multiclass classification task.

In macro-average recall, the recall score for each class is first calculated separately, and then the average of those scores is taken. This approach treats each class equally, regardless of its size or frequency in the dataset.

On the other hand, in micro-average recall, the recall score for each class is calculated and then averaged, taking into account the size or frequency of each class. This approach gives more weight to classes that have more instances in the dataset.

The main difference between macro and micro-averaging is that macro-averaging gives equal weight to each class, while micro-averaging gives more weight to larger classes. Therefore, if there is a class imbalance, micro-average recall would be more appropriate than macro-average recall. In the proposed system, recall is calculated using macro average.

The weighted harmonic mean of precision and recall is known as the F-Score, and a score of 1 denotes the highest possible precision and recall values. It is a helpful indicator for assessing the model's overall performance.