Penanganan Imbalance Data Klasifikasi Teks Lowongan Pekerjaan: Komparasi Algorithmic vs Data Level

Abstract

Abstract

 Auotomatic classification of job experience levels on job portals presents a crucial challenge in e-recuitment systems for mapping industry requirements. The primary challenges addressed in this study are data imbalance, where the number of ‘Senior’ level vacancies is significantly lower than ‘Fresh Graduate’ and ‘Junior/Mid’ levels, and the semantic ambiguity present in job descriptions. This study aims to compare the effectiveness of handling imbalanced data using two distinct approaches: the Algorithmic Level approach utilizing Class Weighting, and the Data Level approach utilizing the Synthetic Minority Over-sampling Technique (SMOTE), applied to Random Forest and Support Vector Machine (SVM) models. The research was conducted on a dataset comprising 2001 real-world job vacancies in Indonesia, employing TF-IDF for feature extraction. The results indicate that the Algorithmic Level approach using SVM with Class Weight yielded the best performance in detecting the minority class, achieving a Recall of 47% for the Senior level, outperforming SVM with SMOTE, which only achieved 34%. These findings indicate that in high-dimensional text data characterized by significant lexical overlap, synthetic oversampling techniques (SMOTE) tend to introduce noise that obscure the decision boundary, making algorithmic weight modification a more robust solution.

Keywords— Imbalanced Data, Text Classification, SVM, SMOTE, Class Weighting