Prediksi Registrasi Mahasiswa Baru pada Perguruan Tinggi Keagamaan Menggunakan Pendekatan Machine Learning dan Deep Learning dengan Interpretasi SHAP

Kurniawan, Alit Fajar (2026) Prediksi Registrasi Mahasiswa Baru pada Perguruan Tinggi Keagamaan Menggunakan Pendekatan Machine Learning dan Deep Learning dengan Interpretasi SHAP. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025241046-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Prediksi registrasi mahasiswa baru menjadi kebutuhan penting bagi perguruan tinggi keagamaan dalam mendukung perencanaan daya tampung, strategi penerimaan, dan pengelolaan sumber daya. Permasalahan yang sering muncul adalah data penerimaan mahasiswa baru masih banyak digunakan sebatas laporan deskriptif, sehingga belum mampu memberikan gambaran prediktif terhadap jumlah registrasi pada tahun berikutnya. Penelitian ini mengembangkan framework prediksi registrasi mahasiswa baru berbasis Multivariate Time Series forecasting dengan pendekatan feature engineering, Machine Learning, Deep Learning, dan Explainable AI. Data yang digunakan merupakan data penerimaan mahasiswa baru pada 40 program studi di perguruan tinggi keagamaan selama periode 2017–2025 dengan total 360 observasi. Tahapan penelitian meliputi persiapan dataset, Exploratory Data Analysis, feature engineering, sequence building, temporal split, Scaling train-only, implementasi model, evaluasi, forecasting tahun 2026, dan interpretasi SHAP. Feature engineering dilakukan melalui pembentukan fitur share, pressure, lag feature, dan target forecasting t+1 untuk merepresentasikan daya tampung, peminat, jalur penerimaan, serta pola historis registrasi. Sepuluh model digunakan sebagai pembanding, yaitu Linear Regression, Random Forest, XGBoost, LSTM, Transformer, LSTM-Attention, BiLSTM, BiLSTM-Attention, Hybrid LSTM-Transformer-MLP, dan CNN-LSTM. Hasil rata-rata dari lima kali iterasi menunjukkan bahwa CNN-LSTM memberikan performa terbaik dibandingkan sembilan model lainnya dengan MAE sebesar 6,0394, RMSE sebesar 8,6185, MAPE sebesar 27,7548%, dan R-Square sebesar 0,9167. Model ini kemudian digunakan untuk forecasting registrasi mahasiswa baru tahun 2026. Interpretasi SHAP menunjukkan bahwa reg_span_lag1, reg_um_lag1, total_reg_lag1, dan jml_dt merupakan fitur yang paling berkontribusi terhadap prediksi total registrasi. Hasil penelitian menunjukkan bahwa pendekatan yang diusulkan menghasilkan model prediksi yang akurat dan dapat diinterpretasikan untuk mendukung perencanaan penerimaan mahasiswa baru berbasis data.
=====================================================================================================================================
Forecasting new student registration has become increasingly important for religious higher education institutions in supporting admission capacity planning, recruitment strategies, and resource management. However, new student admission data are still primarily used as descriptive historical reports and have not been fully utilized to predict future registration. This study proposes a multivariate time series forecasting approach for predicting new student registration by integrating feature engineering, machine learning, deep learning, and Explainable Artificial Intelligence (XAI).
The dataset consists of new student admission records from 40 study programs at a religious higher education institution during the 2017–2025 period, comprising 360 observations. Feature engineering was performed by constructing share features, pressure features, lag features, and forecasting targets (t+1) to better represent admission capacity, applicant numbers, admission pathways, and historical registration patterns. Ten predictive models were evaluated, including Linear Regression, Random Forest, XGBoost, LSTM, Transformer, LSTM-Attention, BiLSTM, BiLSTM-Attention, Hybrid LSTM-Transformer-MLP, and CNN-LSTM.
The average results of five independent runs showed that CNN-LSTM outperformed the other nine models, achieving an MAE of 6.0394, RMSE of 8.6185, MAPE of 27.7548%, and an R-Square of 0.9167. The selected model was then used to forecast new student registration for 2026. SHAP interpretation identified reg_span_lag1, reg_um_lag1, total_reg_lag1, and jml_dt as the most influential features in predicting total registration. These findings demonstrate that the proposed approach provides accurate and interpretable registration forecasts, making it a valuable tool for supporting data-driven planning in new student admissions.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	CNN-LSTM, Feature engineering, Multivariate Time Series forecasting, Prediksi Registrasi Mahasiswa Baru, SHAP, CNN-LSTM, Feature engineering, Multivariate Time Series forecasting, New Student Registration Prediction, SHAP
Subjects:	Q Science > QA Mathematics > QA76.758 Software engineering
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Alit Fajar Kurniawan
Date Deposited:	01 Aug 2026 02:45
Last Modified:	01 Aug 2026 02:45
URI:	http://repository.its.ac.id/id/eprint/141547

Actions (login required)

View Item