Curriculum Classifier

REST API for classifying curricula using machine learning algorithms

The project showcases a rest api that recceives a pdf curriculum and returns the field of ocupation and the level of seniority, along side with the acuracy for each; The machine learning model was trained with 26 areas of ocupation and over 3,000 curricula.

Getting started

After cloning the repository run:

pip install < requirements.txt
fastapi dev server.py

Docs

After running the project locally the documentation is available on:

127.0.0.1:8000/docs

Dataset

This is a custom dataset tailored for this usecase: https://www.kaggle.com/datasets/danicardeal/resume-occupation-and-seniority

Benchmarks and algorithms comparison

PDF Extraction Libraries Benchmark

• Benchmark: Time taken to process all PDFs in the dataset.
• Libraries evaluated: Tika, PyMuPDF, Textract, Pypdfium2
• Library chosen: Pdftotext

Classification of Seniority and Area of Expertise

Seniority Classification

For training the seniority classifier, the text field and the seniority field from the CSV were used.

Area of Expertise Classification

For the area of expertise classifier, the class number and text fields were utilized.

Preprocessing

In the preprocessing phase, the following steps were implemented:

• Spacy

  • Stopwords Removal
  • Lemmatization
  • Tokenization
  • Large model used: en_core_web_lg

• Re

  • Removal of hyperlinks

• CSV

  • Mapping of area of expertise classifications to numerical values for training purposes.

• Models evaluated:

  • Logistic Regression
  • Support Vector Machine
  • Random Forest
  • k-Nearest Neighbors
  • Bernoulli Naive Bayes
  • Naive Bayes
  • CatBoost
  • XGBoost

• Model chosen: XGBoost with parameters

Vectorizer and Model Persistence

The data was vectorized using the CountVectorizer from sklearn. The trained model was exported and loaded using joblib for deployment and inference.

Accuracies

Seniority Accuracy

Area of Expertise Accuracy

Confusion Matrices

Confusion Matrix for Area of Expertise

Confusion Matrix for Seniority

Conclusion

This project demonstrates a robust proof of concept for a REST API capable of classifying curricula into specific fields of occupation and levels of seniority using machine learning algorithms. The model, trained on a custom dataset with over 3,000 resumes spanning 26 areas of occupation, achieves accurate classifications while providing valuable insights into the efficacy of various PDF extraction libraries and machine learning models.

Key Takeaways

1. Effective PDF Processing: After evaluating multiple libraries for PDF extraction, pdftotext was selected for its superior performance in terms of processing time.

2. Comprehensive Preprocessing: Utilizing Spacy for text processing (stopwords removal, lemmatization, and tokenization) and Re for hyperlink removal ensured clean and relevant data for model training.

3. Model Evaluation and Selection: Among the evaluated models, XGBoost emerged as the best performer, providing high accuracy in both seniority and area of expertise classifications.

4. Data Vectorization and Persistence: The use of CountVectorizer for data vectorization and joblib for model persistence streamlined the deployment and inference process, making the system efficient and scalable.

5. Accuracy and Performance: The achieved accuracies and confusion matrices for both seniority and area of expertise classifications highlight the model's effectiveness and reliability.

This project not only showcases the potential for automated resume classification but also serves as an excellent learning experience in handling real-world data, evaluating multiple libraries and models, and implementing a complete machine learning pipeline from data preprocessing to deployment.

Free Software, Hell Yeah!