Job Recommendation System

LinkedIn Job Listings Analysis Using Machine Learning

---

1. Introduction

The job market is rapidly transforming due to technological advancements, remote work trends, and evolving industry demands. LinkedIn, as a leading professional networking platform, offers an extensive repository of job postings, providing opportunities for data-driven insights.

This project analyzes LinkedIn job listings to identify high-demand roles, in-demand skills, and salary benchmarks. It also implements a recommendation system to match job seekers with roles aligned to their profiles. The use of ML, NLP, and explainable AI ensures scalable, accurate, and interpretable insights.

---

2. Relevant Work

• Traditional ML models like Logistic Regression and Decision Trees are widely used for job classification.
• Neural network embeddings improve personalized job recommendations.
• Explainable AI (XAI) techniques like SHAP and LIME enhance model interpretability.
• This project combines traditional ML, neural embeddings, and XAI to address limitations of prior studies, ensuring robust evaluation across multiple metrics.

---

3. Methodology

3.1 Data Preprocessing

• Cleaned missing values and outliers
• Applied feature engineering:
  • TF-IDF for text data
  • One-hot encoding for categorical variables
  • Normalization for numerical fields

3.2 Model Development

• Classification: Logistic Regression, Random Forest, KNN for job types
• Regression: Neural Networks, Gradient Boosting for salary prediction
• Recommendation: Neural embeddings for personalized job suggestions

3.3 Evaluation

• Metrics: Accuracy, Precision, Recall, F1-score (classification), MSE (regression), Precision/Recall (recommendation)
• SHAP visualizations used for explainability

---

4. Models

Model	Use Case	Key Notes
Logistic Regression	Job type classification	Interpretable baseline
Naive Bayes	Text classification	Simple but affected by class imbalance
K-Nearest Neighbors (KNN)	Personalized recommendations	High accuracy for job recommendations
Random Forest	Classification	Robust feature importance analysis
Gradient Boosting	Salary prediction	Captures complex interactions
Neural Networks	Recommendations	Outperforms traditional models using embeddings

Hyperparameter optimization was applied for all models.

---

5. Data

• Size: 124,000 LinkedIn job postings
• Fields:
  • Job Titles & Descriptions
  • Work Types (remote, hybrid, onsite)
  • Salaries (min, max, median)
  • Company Info (industry, HQ, employee count)
  • Skills (technical & soft skills)

5.1 Data Cleaning

• Categorical missing values → "Unknown"
• Salary missing values → median imputation

5.2 Data Enrichment

• Merged external datasets for skills and industries to enhance predictive power

5.3 Data Visualization

• Job titles, locations, work types, and experience levels analyzed
• Correlation heatmaps used for numerical feature relationships

---

6. Feature Engineering

• Textual data: TF-IDF vectorization for job descriptions and skills
• Categorical encoding: One-hot encoding for work type and experience level
• Numerical normalization: Min-Max scaling for salary fields

---

7. Training and Testing

• Data split: 80% training / 20% testing with stratified sampling
• Cross-validation: 10-fold
• Hyperparameter tuning: Optuna
• Oversampling: Handled class imbalance
• Ensemble methods: Combined multiple models for stability

---

8. Results

8.1 Logistic Regression

• Accuracy: 65.26% | AUC: 71.85% | F1-score: 55.32%
• Moderate performance; captures text-based patterns

8.2 Naive Bayes

• Accuracy: 68.99% | AUC: 68.00% | F1-score: 12.20%
• Poor recall due to class imbalance

8.3 KNN

• 10-fold CV Accuracy: 97.09% | Test Accuracy: 94.97% | F1-score: 92.12%
• High performance for personalized recommendations

8.4 Random Forest

• Test Accuracy: 92.09% | AUC: 97.41% | F1-score: 87.85%
• Robust and interpretable

8.5 Gradient Boosting

• Test Accuracy: 91.67% | AUC: 96.76% | F1-score: 87.24%
• Strong predictive power

8.6 Job Recommendations

• Neural embedding-based system provides personalized job suggestions
• SHAP plots highlight skills and location influencing recommendations

8.7 Explainable AI

• SHAP visualizations provide transparency and trust
• Helps stakeholders validate and interpret recommendations

---

9. Discussion

• KNN and Random Forest excel at classification; Neural Networks for recommendations
• Challenges: class imbalance, high computational costs
• Feature engineering and XAI improved accuracy and trust

---

10. Conclusion

• ML and NLP successfully analyzed LinkedIn job listings
• High-performing models for classification, regression, and recommendation
• SHAP visualizations enhanced interpretability
• Future work: longitudinal data, real-time recommendation optimization

---

11. Folder Structure

Job-Recommendation-System/

─ cleaned_text.csv

─ cleaned_text_with_job_info.csv

─ postings.csv

─ companies/ # Company-related datasets

─ jobs/ # Job-related datasets

─ mappings/ # Mappings for skills, industries.

─ ML4GEN.ipynb # Main project notebook

─ NLP.ipynb # NLP analysis notebook