Person-job fit is the extent to which a candidate is well-suited for a position. Determining this person-job fit is currently a laborious task. Qin et al. introduce a novel approach that uses neural networks to automatically guess the person-job fit based on information extracted from job postings, resumes, and previously filled vacancies.
The challenge of person-job fit has been studied by many from different perspectives. However, the solutions that have been proposed to date typically rely on manual inspections, which are labour-intensive, inaccurate, and prone to subjectivity.
The authors therefore propose a model that they call an ability-aware person-job fit neural network (APJFNN). This model should be able to determine whether a candidate’s resume matches the requirements in a job posting.
The APJFNN consists of three components:
Hierarchical ability-aware representation
Person-job fit prediction
The text from job postings and resumes can’t be directly processed by neural networks, so it’s first converted into some numbers that together capture both the meaning of each word and the position of the word in the text.
The result of this conversion is passed to the next component.
The second component consists of four parts that determine the importance which words and phrases should receive more attention:
Some words and phrases in a job posting are more important than others, e.g. mentions of skills. The component therefore estimates the importance of each word in the job posting.
Since required skills are always listed in descending order of importance, their importance is also estimated based on where in the job posting it’s mentioned. The however.
When a recruiter looks at a resume they are mostly interested in the specific skills that were listed in the job posting. The component therefore estimates the importance of each listed skill in the resume based on whether it appears in the job posting.
The skills listed in the resume’s “work experience” section are always listed in reverse chronological order, which again means that more importance should be ascribed to skills that appear earlier in the resume.
Now that the words in the job posting and resume have been weighted in various ways, they are fed into a neural network that learns to predict how well the resume matches the job posting.
The APJFNN was trained and evaluated using a dataset provided by . The dataset contains about 18,000 job postings and 899,000 resumes from applicants for those job postings, of which about 12,800 were accepted. An equal amount of true positives and (randomly selected) true negatives were selected to train the model.
The performance of the APJFNN model is compared with five classical supervised learning models: logistical regression, decision tree, AdaBoost, random forests, and gradient-boosting decision tree.
The results suggest that the APJFNN model performs a lot better than existing algorithms.
The model is also fairly robust, showing only small differences in performance when the size of the training dataset is lowered from 80% to 40% of the dataset.
Another nice bonus is that the model can use the importance estimations to explain what it bases its predictions on.
The authors introduce a neural network model to determine how well a resume matches the requirements in a job posting
Not all words in a resume or job posting are equally important
The newly proposed model performs better than existing supervised machine learning algorithms