[Paper Review] Curriculum Vitae Recommendation Based on Text Mining

Honorio Apaza Alanoca, Américo Ariel Rubin de Celis Vidal+1 more|arXiv (Cornell University)|Jul 21, 2020

Data Mining Algorithms and Applications4 references4 citations

TL;DR

This paper proposes a curriculum vitae (CV) recommendation system for the Peruvian job market using text mining and NLP techniques, specifically TF-IDF weighting to rank CVs based on relevance to job offers. The method computes term frequency-inverse document frequency scores across CVs and job descriptions, producing a normalized relevance score (0.000 to 0.549) that serves as a precise, deterministic rating for CV recommendation, with the general TF-IDF average acting as a reliable qualification metric for candidate matching.

ABSTRACT

During the last years, the development in diverse areas related to computer science and internet, allowed to generate new alternatives for decision making in the selection of personnel for state and private companies. In order to optimize this selection process, the recommendation systems are the most suitable for working with explicit information related to the likes and dislikes of employers or end users, since this information allows to generate lists of recommendations based on collaboration or similarity of content. Therefore, this research takes as a basis these characteristics contained in the database of curricula and job offers, which correspond to the Peruvian ambit, which highlights the experience, knowledge and skills of each candidate, which are described in textual terms or words. This research focuses on the problem: how we can take advantage from the growth of unstructured information about job offers and curriculum vitae on different websites for CV recommendation. So, we use the techniques from Text Mining and Natural Language Processing. Then, as a relevant technique for the present study, we emphasize the technique frequency of the Term - Inverse Frequency of the documents (TF-IDF), which allows identifying the most relevant CVs in relation to a job offer of website through the average values (TF-IDF). So, the weighted value can be used as a qualification value of the relevant curriculum vitae for the recommendation.

Motivation & Objective

• To address the challenge of matching job seekers with relevant job offers in the unstructured, text-based data prevalent in Peruvian job portals.
• To develop a recommendation system that leverages explicit textual content from CVs and job postings to improve hiring efficiency.
• To evaluate the effectiveness of TF-IDF in quantifying the relevance of CVs to specific job offers using real-world Peruvian employment data.
• To provide a data model for automated CV recommendation based on term relevance and inverse document frequency.
• To support university programs and job market analysis by identifying key skills and profiles in demand.

Proposed method

• Web scraping using Scrapy and BeautifulSoup to extract 10,000 job offers and 10,000 CVs in Spanish from Peruvian job platforms like Computrabajo and Bumeran.
• Text preprocessing via NLTK, including ASCII removal, lowercase conversion, punctuation and accent stripping, and stopword filtering.
• Computation of Term Frequency (TF) and Inverse Document Frequency (IDF) for each term across the corpus, using the formula: $ \text{TF-IDF} = \text{TF} \times \text{IDF} $, where $ \text{IDF} = \log\left(\frac{N}{\text{df}(t)}\right) $.
• Aggregation of TF-IDF scores per CV per job offer, computing document-specific averages and a general average across all documents.
• Normalization of relevance scores using min-max scaling: $ X_{cs} = \frac{X - X_{\text{min}}}{X_{\text{max}} - X_{\text{min}}} $, with values ranging from 0.000 to 0.549.
• Construction of a 5000×9998 matrix mapping CV relevance to job offers, enabling systematic analysis and recommendation.

Experimental results

Research questions

• RQ1 How can unstructured textual data from Peruvian job portals be effectively transformed into structured relevance scores for CV recommendation?
• RQ2 To what extent does the TF-IDF weighting method accurately reflect the relevance of a CV to a specific job offer?
• RQ3 Can the general average of TF-IDF scores across documents serve as a reliable, deterministic rating for CV recommendation?
• RQ4 How does the normalization of TF-IDF values improve interpretability and usability in a recommendation system?
• RQ5 What insights can be drawn about skill and experience demand in the Peruvian job market from the term relevance analysis?

Key findings

• The TF-IDF method successfully quantified the relevance of CVs to job offers with a deterministic, non-probabilistic calculation based on term frequency and inverse document frequency.
• The highest normalized relevance score achieved was 0.549, while the lowest was 0.000, indicating a clear range of candidate suitability.
• CVs 2 and 8 were identified as the most relevant for Job 7, based on the highest average TF-IDF scores across relevant terms.
• The general average TF-IDF score (0.092) across all documents served as a robust, consistent metric for ranking CV relevance.
• The data model produced a 5000×9998 matrix of CV-to-job relevance scores, enabling scalable and systematic candidate recommendation.
• The study demonstrated that normalized TF-IDF scores can be directly used as a qualification value in a recommendation system without requiring probabilistic modeling or user feedback.