This project goes beyond basic classification by integrating advanced NLP techniques, rigorous model evaluation, and an interactive user interface. Below are the main components and innovations included in the system, as detailed in the accompanying academic report.
Resumes are automatically classified into one of 24 professional job categories (e.g., IT, Healthcare, HR). After extensive experimentation with various machine learning models and text embeddings, the best-performing configuration was a Support Vector Machine (SVM) trained on Sentence-BERT (SBERT) embeddings, achieving 76% accuracy on an independent test set.
Given a job description, the system identifies the top 5 most relevant resumes based on semantic similarity using SBERT and cosine similarity. This module was validated against LLM-based human-like assessments (via Gemini), confirming the system's practical alignment with "human evaluation".
Four feature extraction methods were tested: TF-IDF, Word2Vec, Doc2Vec, and Sentence-BERT. These were combined with Logistic Regression, Random Forest, and Support Vector Machine classifiers. Extensive grid search, 5-fold stratified cross-validation, and SMOTE were applied to address class imbalance. Wilcoxon signed-rank tests were used to assess statistical significance among model performances.
To interpret model predictions, SHAP (SHapley Additive Explanations) was used. These techniques help reveal the most influential words contributing to classification decisions, even when using dense embeddings like SBERT.
A user-friendly GUI built with Flask (Python) and HTML/CSS/JavaScript offers two main functionalities:
- Resume Classification Page: Users can paste or upload a resume (PDF) and receive the top 3 predicted job categories.
- Job Description Matching Page: Recruiters can input a job description and retrieve the top 5 matching resumes from the database.
To run this project, you need to install all the required Python packages. All dependencies are listed in the requirements.txt file.
Note: This project was developed and tested using Python 3.11. It is strongly recommended to use this version to ensure compatibility and reproducibility.
The requirements.txt file was generated using:
pip freeze > requirements.txt-
Clone the repository:
git clone https://github.com/andreabochicchio02/ResumeClassification.git cd ResumeMining -
(Optional) Create and activate a virtual environment: You can use either a standard Python
venvor a Conda environment to isolate dependencies and avoid conflicts with other projects. -
Install all dependencies:
pip install -r requirements.txt
Now you're ready to run the project!
This project is licensed under the MIT License.