OPTIMASI MODEL REKOMENDASI PEKERJAAN MELALUI PENINGKATAN ROBUSTNESS VEKTOR EMBEDDING LOWONGAN KERJA DAN RESUME BERBASIS TSDAE

Muhammad Adin Palimbani, Prof. Dr. Drs. Azhari, MT

2025 | Tesis | MAGISTER KECERDASAN ARTIFISIAL

Model rekomendasi pekerjaan berbasis kecerdasan buatan masih menghadapi tantangan besar dalam menangani struktur informasi lowongan kerja yang tidak berlabel, tidak konsisten dan noisy. Judul dan kategori pekerjaan yang diunggah oleh perusahaan sering kali bervariasi, menyebabkan overlapping content dan kesulitan dalam klasifikasi otomatis. Penelitian ini bertujuan mengoptimalkan model rekomendasi pekerjaan adaptif melalui peningkatan robustness vektor embedding lowongan kerja dan resume terhadap data noisy, inconsistent, unstructured dan unlabelled menggunakan Transformer based Sequential Denoising Auto-Encoder (TSDAE). Proses pembentukan embedding akan dilakukan secara terintegrasi melalui 3 tahapan Domain Adaptive Pre-Training yakni menggunakan Pre-Trained Language BERT (baseline), training TSDAE pada domain target dan fine-tuning hasil TSDAE pre-trained. Dataset yang digunakan adalah 4000 informasi lowongan kerja dari Kaggle, O*NET Standard Occupation dan lima Resume sebagai skenario pengujian. Hasil embedding kemudian dikelompokkan dengan K-Means Clustering dengan k=12 dan diukur kemiripannya menggunakan Cosine Similarity. Setiap pekerjaan yang direkomendasikan sistem akan dianotasi oleh lima annotator. Model dievaluasi pada Top-20 Rekomendasi dengan metrik MAP@20, NDCG@20, P@20 dan MRR@20. Model yang diusulkan menunjukkan bahwa TSDAE fine-tuned secara signifikan mengungguli model baseline BERT tanpa TSDAE, dengan mencapai skor rata-rata MAP@20 86%, NDCG@20 93%, P@20 80?n MRR@20 90%. Hasil ini membuktikan pendekatan TSDAE yang dioptimalkan mampu menghasilkan rekomendasi pekerjaan yang adaptif dan relevan terhadap  karakteristik data noisy, inconsistent dan unlabelled.

Artificial intelligence-based job recommendation models still face major challenges in handling unlabeled, inconsistent, and noisy job vacancy information structures. Job titles and categories uploaded by companies often vary, causing overlapping content and difficulties in automatic classification. This research aims to optimize adaptive job recommendation models by improving the robustness of job vacancy and resume embedding vectors against noisy, inconsistent, unstructured, and unlabeled data using Transformer-based Sequential Denoising Auto-Encoder (TSDAE). The embedding formation process will be carried out in an integrated manner through three stages of Domain Adaptive Pre-Training, namely using Pre-Trained Language BERT (baseline), training TSDAE on the target domain, and fine-tuning the results of TSDAE pre-trained. The dataset used is 4,000 job vacancy information from Kaggle, O*NET Standard Occupation, and five resumes as a test scenario. The embedding results are then grouped using K-Means Clustering with k=12 and their similarity is measured using Cosine Similarity. Each job recommended by the system will be annotated by five annotators. The model is evaluated on the Top-20 Recommendations using the metrics MAP@20, NDCG@20, P@20, and MRR@20. The proposed model demonstrates that TSDAE fine-tuned significantly outperforms the baseline BERT model without TSDAE, achieving an average score of MAP@20 86%, NDCG@20 93%, P@20 80%, and MRR@20 90%. These results demonstrate that the optimized TSDAE approach is capable of generating adaptive and relevant job recommendations for noisy, inconsistent, and unlabeled data.

Kata Kunci : Rekomendasi Pekerjaan, Domain Adaptation, TSDAE, K-Means Clustering, Cosine Similarity