Claim Missing Document
Check
Articles

Found 14 Documents
Search

Genetic algorithm-optimized BERTopic with SHAP explainability for institutional research trend analysis Muhammad Dedi Irawan; Yustria Handika Siregar; Hewa Majeed Zangana; Ali Ikhwan
IAES International Journal of Artificial Intelligence (IJ-AI) Vol 15, No 4: August 2026
Publisher : Institute of Advanced Engineering and Science

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.11591/ijai.v15.i4.pp3818-3826

Abstract

Institutional research grant titles constitute short-text grey literature characterized by heterogeneous semantic structures, making topic identification and research trend analysis challenging. This study proposes an integrated bidirectional encoder representation from transformers-based topic modeling (BERTopic) framework combining genetic algorithm (GA)-based hyperparameter optimization and Shapley additive explanations (SHAP)-based interpretability to improve semantic topic quality and model transparency. GA was applied to optimize dimensionality-reduction and density-based clustering parameters, while SHAP was used to estimate the contribution of bigram features to the surrogate classifier’s predictions of BERTopic-generated topic labels. Experimental results demonstrated that the proposed framework improved topic coherence from 0.367 to 0.543 while reducing the outlier ratio from 21.12% to 13.55%. In addition, the number of topics decreased from 46 to 10, resulting in a more compact and less fragmented topic structure. The resulting topic structure revealed dominant themes related to higher education, religious moderation, Islamic counseling, halal tourism, and sharia banking. Overall, the proposed framework contributes to the development of more coherent, interpretable, and semantically robust topic modeling for institutional short-text grey literature analysis.
Performance Evaluation of Sensor Data Filtering Methods for Signal Processing in TVET Learning Applications Farid Baskoro; Hisham A. Shehadeh; Hewa Majeed Zangana; Tri Wrahatnolo; Puput Wanarti Rusimamto; Fendi Achmad; Aristyawan Putra Nurdiansyah
Buletin Ilmiah Sarjana Teknik Elektro Vol. 8 No. 4 (2026): August
Publisher : Universitas Ahmad Dahlan

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.12928/biste.v8i4.16883

Abstract

Technical and Vocational Education and Training (TVET) learning requires sensor measurement data that are stable, accurate, and easy to interpret. Raw LiDAR sensor data often contain fluctuations that may interfere with the readability results. This study employed an experimental-comparative design by comparing Moving Average, Median Filter, Savitzky-Golay, Butterworth, and Simple Kalman Filter. The data acquisition system used a VL53L0X LiDAR sensor and ESP32 microcontroller. Data processing was conducted in MATLAB on 10,500 samples at a sampling frequency of 50 Hz. The evaluation was carried out based on error metrics, signal stability, noise reduction, and filter responsiveness. The raw data had a standard deviation of 111.26 and still showed fluctuations that required reduction. A Greenhouse–Geisser-corrected repeated-measures ANOVA showed a significant effect of filtering method on segment-level residual RMSE, F(1.10,44.92)=26.23, p<0.001, partial η2=0.390. Bonferroni-adjusted comparisons showed that Savitzky–Golay produced significantly lower residual RMSE than the other methods, indicating stronger preservation of the raw-signal pattern. The results showed that Savitzky–Golay achieved the best overall trade-off, with the lowest residual deviation, the highest estimated SNR of 32.154 dB, and good pattern preservation without excessive smoothing. Butterworth and Simple Kalman provided stronger fluctuation reduction, although Kalman introduced greater deviation and a 39-sample delay. Moving Average offered simple smoothing, whereas the Median Filter was more suitable for impulsive noise and outliers. This study contributes a comparative evaluation of filtering methods from both signal-processing and TVET pedagogical perspectives, supporting filter selection based on smoothness, readability, noise reduction, and responsiveness in signal processing.
Power System Stabilizer Optimization Based on Modified Black‑Winged Kite Algorithm Widi Aribowo; Laith Abualigah; Diego Oliva; Nur Vidia Laksmi B; Fithrotul Irda Amaliah; As’ad Shidqy Aziz; Hewa Majeed Zangana
Buletin Ilmiah Sarjana Teknik Elektro Vol. 7 No. 4 (2025): December
Publisher : Universitas Ahmad Dahlan

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.12928/biste.v7i4.14669

Abstract

This article presents a Modified Method for tuning the parameters of a power system stabilizer (PSS). This article suggests a different approach that modifies the Black Kite Algorithm (BKA). The Black Kite (BKA) method is inspired by the migratory and predatory habits of the black kite. BKA combines the Leader and Cauchy mutation strategies to improve the algorithm's capacity for global search and convergence rate. This article includes comparative simulations of the PSS objective function and transient response to verify the effectiveness of the suggested strategy. The study validates the proposed method through comparison with both conventional techniques and the original BKA. Simulation results demonstrate that, when benchmarked against competing algorithms, the proposed method consistently yields optimal performance and exhibits faster convergence in certain scenarios. Notably, it reduces undershoot and overshoot by an average of 65% and 90.22%, respectively, compared to the PSS-Lead Lag method. Furthermore, the proposed approach not only minimizes overshoot and undershoot but also achieves a significantly faster settling time.
Hybrid Vision Transformer for Brain and Lung Tumor Detection: A Multi-Modal Approach Using MRI (BraTS) and CT (LUNA16) Datasets Hewa Majeed Zangana; Mohammed Aquil Mirza; Sharyar Wani; Xinwei Cao; Marwan Omar
Buletin Ilmiah Sarjana Teknik Elektro Vol. 7 No. 4 (2025): December
Publisher : Universitas Ahmad Dahlan

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.12928/biste.v7i4.14766

Abstract

The integration of artificial intelligence (AI) into medical imaging has transformed clinical diagnostics, yet existing CNN-based systems still struggle with capturing global spatial context and generalizing across modalities. This study addresses this gap by proposing a hybrid Vision Transformer (ViT) architecture for tumor detection in MRI and CT scans, evaluated on two benchmark datasets: BraTS (brain MRI) and LUNA16 (lung CT). The research contribution is a unified, end-to-end transformer model that processes heterogeneous modalities without traditional feature-level fusion. The proposed method incorporates convolutional layers for local feature extraction alongside transformer blocks for long-range dependency modeling. Extensive experiments demonstrate that our model achieves a 2.5% higher Dice score and 3.1% higher F1-score compared to state-of-the-art CNN-based baselines, with accuracy reaching 95.4% on BraTS and 93.6% on LUNA16. Attention-based heatmaps further enhance explainability by highlighting clinically relevant tumor regions. These results show that hybrid transformers offer a robust and interpretable framework for multi-modal tumor detection, paving the way for more reliable and transparent AI-assisted radiological diagnostics.