SESyahril Efendi

Syahril Efendi

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Syahril Efendi has a total of 51 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Development of the 3DSCO model using long short-term memory and reinforcement learning for supply chain management in dropshipping-based online businesses

    by Zulfahmi Indra, Syahril Efendi, Tulus Tulus, Poltak Sihombing

    The object of this study is the supply chain and inventory management process in a dropshipping-based online business, represented by the Amazon Sale Report.csv dataset. The problem solved is the low ability of traditional supply chain models to provide adaptive and predictive systems to manage product demand, inventory, and distribution in a dynamic digital environment. This study aims to develop a 3DSCO (Data-Driven Dropship Supply Chain Optimization) model, a data-driven supply chain optimization model that integrates Long Short-Term Memory for demand forecasting and Reinforcement Learning for adaptive inventory control. The study uses 6,000 transaction data taken from Amazon Sale Report.csv through a systematic random sampling technique. The results show that the developed model is able to predict demand patterns accurately, achieve stable convergence during the training process, and produce more efficient inventory control policies compared to conventional approaches. The results show that the proposed model achieved an RMSE of 4.28, MAPE of 6.87%, and R2 of 0.94 for demand forecasting. The model also reduced the stockout rate to 7.8%, decreased average holding cost to 95.3 compared with 120.5 under EOQ, and achieved a fill rate between 94.2% and 99.3%. These results demonstrate the effectiveness of the proposed framework in improving logistics performance and reducing inventory-related costs. The study findings can be applied to e-commerce companies, dropshipping businesses, and digital businesses that possess historical sales data and require data-driven decision-making systems to improve supply chain efficiency, reduce operational costs, and enhance responsiveness to changes in market demand

  • Open Access

    Development of a TRIE-BERT pipeline for automatic spacing and low resource language classification in Batak Toba and Angkola scriptio continua texts

    by Muhammad Anggia Muchtar, Opim Salim Sitompul, Maya Silvi Lydia, Syahril Efendi

    Batak Toba and Batak Angkola texts written in scriptio continua form without spaces are the object of this study. The work solves a low-resource variety classification problem where the two varieties are similar and missing word boundaries introduce segmentation noise. A hybrid TRIE-BERT pipeline was developed in which trie automation performs deterministic spacing, the restored spacing is fixed, and the spaced text becomes a stable input interface for a Bidirectional Encoder Representations from Transformers (BERT) classifier. Experiments used a Batak lexicon of 19,070 word entries and 8,000 sentences, 4,000 per variety, evaluated under four data schemes from 1,000 to 8,000 sentences and five epoch settings from 5 to 50 with an 80:20 split. After lexicon recalibration of about 70 sentences, spacing reached 98 percent accuracy. The best setting at 8,000 sentences and 50 epochs achieved 0.85 test accuracy with 0.343 training loss, 0.85 ROC AUC, and 0.85 F1-score, exceeding a long short-term memory recurrent neural network baseline (LSTM-RNN) at 0.80 accuracy, 0.397 loss, 0.803 ROC AUC, and 0.80 F1-score. Class-wise evaluation yielded precision 0.81 and recall 0.92 for Toba and precision 0.90 and recall 0.79 for Angkola, explaining averaged precision 0.86 and recall 0.85. The improvement is associated with the combined use of deterministic trie-based boundary recovery and contextual BERT classification, where spacing is fixed before classification to reduce token ambiguity and stabilize the input structure. The results support Batak text processing pipelines that require automatic spacing and variety detection under limited labels, provided lexicon coverage is maintained and spelling variation is controlled

  • Open Access

    Identifying the best models for BISINDO alphabet gesture classification to support the communication needs of the deaf community

    by Ilka Zufria, Tengku Henny Febriana Harumy, Syahril Efendi

    The object of this research is the BISINDO alphabet gestures which are static hand movements used by the Deaf community in Indonesia to communicate, with each letter having a unique hand pattern influenced by culture and regional variations. The problem solved is the accuracy and performance of BISINDO hand gesture classification which is less than optimal due to complex gesture variations and computational limitations on lightweight devices. This study examines the performance of BISINDO alphabet gesture classification using deep learning models with 4 architectures: VGG-19, ResNet-50, MobileNetV2, and Inception-V3. This object is a collection of images used as a dataset consisting of 10,400 images (7,280 training, 2,080 validation, and 1,040 testing). The results show that the MobileNetV2 and VGG-19 architectures achieve the highest accuracy of 100%, followed by Inception-V3 (99%) and ResNet-50 (98%). The results of this study indicate that the superior performance of MobileNetV2 is due to its efficient depthwise separable convolutional architecture, while the superiority of VGG-19 lies in its very deep architecture and the use of small convolutional filters to capture very detailed hierarchical features of gestures. Meanwhile, Inception-V3 excels thanks to the inception module that captures gesture features at various scales. The results of this performance comparison, MobileNetV2 is the most superior because of its computational efficiency that supports high performance, as well as adaptation to complex variations in BISINDO alphabet gestures. The results of this study can also be applied in the fields of education and communication for the deaf community, especially in embedded applications, thus enabling real-time sign language accessibility

  • Open Access

    SAGOMECON: An Adaptive ε-Constraint-Based Optimization Method for Multi-Criteria Decision-Making in Collaborative Industrial Networks

    by M Mesran, Poltak Sihombing, Syahril Efendi, Muhammad Zarlis

    SAGOMECON: An Adaptive ε-Constraint-Based Optimization Method for Multi-Criteria Decision-Making in Collaborative Industrial Networks


  • Open Access

    A hybrid multi-scale convolution neural network with attention and texture features for improved image classification

    by Irpan Adiputra Pardosi, Tengku Henny Febriana Harumy, Syahril Efendi

    The object of this study is the classification of low-resolution and multi-class images, represented by the CIFAR-10 benchmark dataset. It is challenging to accurately classify low-resolution and multi-class images because traditional CNNs usually have trouble identifying both global and complex texture patterns. To address this issue, this study employs the CIFAR-10 dataset as a representative benchmark for real-world scenarios where image quality is limited, such as in low-cost medical imaging, remote sensing, and security surveillance systems. The limited discriminability of traditional CNNs in these situations is the primary issue addressed. The proposed method employs three parallel convolutional streams with distinct kernel sizes (3 × 3, 5 × 5, and 7 × 7) to capture hierarchical spatial patterns, followed by the integration of two attention mechanisms – squeeze-and-excitation and convolutional block attention module – that adaptively emphasize the most relevant spatial and channel-wise information. In addition, structural texture descriptors such as Gray-level co-occurrence matrix, local binary pattern, and Gabor filters are computed independently and later fused with the deep representations to enrich the feature space. Experiments were carried out on the CIFAR-10 dataset under varying levels of class complexity: 10, 5, and 3 categories. The results reveal that the hybrid approach significantly improves precision, recall, and F1-score across all scenarios, with the highest accuracy of 90.87% obtained when only three classes are involved. These improvements are explained by the complementary nature of deep and handcrafted features, which together enable the model to learn both global semantics and fine-grained local textures can achieve higher classification accuracy, improved reliability, and reduced misclassification errors, ultimately enhancing the effectiveness of applications ranging from medical decision support to intelligent surveillance.


  • Open Access

    Enhancing retrieval performance in social media with corpus-based query expansion using bidirectional encoder representations from transformers

    by Roberto Kaban, Poltak Sihombing, Syahril Efendi, Maya Silvi Lydia

    This study focuses on a collection of tweets related to government services (e-government), which are preprocessed and transformed into a domain-specific corpus for query expansion. Conventional IR models struggle with unstructured and noisy content containing informal language and abbreviations, which reduces retrieval accuracy. To overcome these issues, this study proposes a hybrid query expansion (QE) model named ROCBERT-QE, which combines corpus-based retrieval (CBR) with bidirectional encoder representations from transformers (BERT). The model applies dual expansion, using corpus-based co-occurrence frequencies to capture lexical relationships and BERT embeddings to preserve semantic context. A domain-specific corpus consisting of 5,017 preprocessed tweets related to Indonesia’s National Health Insurance (BPJS) was constructed, encompassing 6,215 unique terms that represent linguistic variation and informality in public discourse. Experimental results demonstrate that ROCBERT-QE outperforms baseline retrieval methods such as TF-IDF, BM25, and standard BERT. For single-word queries, Recall reached 0.8574 and Precision 0.8807, while for sentence-level queries, Recall was 0.8932 and Precision 0.9175. The synergy of frequency-based and contextual expansion enables effective handling of lexical noise and semantic ambiguity. The results confirm the scientific potential of combining corpus-based and transformer-based approaches in IR tasks involving unstructured content. Practically, ROCBERT-QE can be applied for real-time analysis of citizen discourse in e-government contexts, such as service evaluation, policy feedback, and early detection of public issues. The framework is scalable and adaptable to other domains with informal or multilingual data characteristics

  • Open Access

    Identifying the impact of Metaperceptron in optimizing neural networks: a comparative study of gradient descent and metaheuristic approaches

    by Darwin Darwin, Tengku Henny Febriana Harumy, Syahril Efendi, Carles Juliandy, Binarwan Halim

    This study investigates the application of the Metaperceptron framework as an adaptive optimization mechanism in training neural networks for polycystic ovary syndrome (PCOS) diagnosis. The research addresses the persistent challenges in conventional optimization methods such as slow convergence, local minima entrapment, and hyperparameter sensitivity that hinder the efficiency and generalization capability of artificial neural networks. By integrating Metaperceptron with both gradient descent (GD) and genetic algorithm (GA), this work demonstrates significant improvements in convergence speed and diagnostic accuracy. Specifically, Metaperceptron-enhanced GD reduced convergence time by nearly 40% while maintaining high accuracy (0.8950 for single-layer neural network and 0.9100 for multi-layer neural network). These results were achieved through dynamic learning rate adjustment and meta-level control over search strategies, enabling better exploration-exploitation balance during training. The findings are explained by the framework’s ability to adaptively respond to gradient landscapes and dataset characteristics, offering a more stable and efficient optimization process. Practical implementation of the proposed method is feasible under conditions where data quality and representativeness are ensured, particularly in medical diagnostics and other domains involving imbalanced or noisy datasets

  • Open Access

    Evaluating deep learning architectures for CO2 emissions forecasting: TCN, LSTM, and hybrid approaches with hyperparameter optimization

    by Roni Yunis, Tengku Henny Febriana Harumy, Syahril Efendi

    The object of the study is CO2 emission prediction using deep learning models. The problem lies in developing accurate models capable of handling temporal dependencies and periodic patterns in CO2 data. To address this, three deep learning models – temporal convolutional network (TCN), long short-term memory (LSTM), and a hybrid TCN-LSTM are evaluated. These models are optimized using random search and Bayesian optimization. Results indicate that the Hybrid TCN-LSTM model, optimized via random search, performs best, achieving MAE: 1.0269, R2: 0.9305, and MAPE: 4.47%. TCN excels at capturing periodic patterns through dilated convolutions, while LSTM handles long-term dependencies. Their integration combines these strengths, improving accuracy. Optimal hyperparameters (learning rate: 0.000539, dropout rate: 0.5) enhance robustness. Random search outperforms Bayesian optimization in navigating complex search spaces and avoiding local optima. Key findings include the hybrid model's ability to address short-term periodicity and long-term trends, and Random Search’s reliability over Bayesian methods in this context. These insights advance time series forecasting methodologies and support robust predictive frameworks. Practically, they aid environmental policy, energy planning, and carbon trading by enabling data-driven decisions for emission reduction. However, implementation requires high-quality historical data and sufficient computational resources


  • Open Access

    Evaluating the impact of weight initialization on recurrent and transformer-based models in financial asset price prediction

    by Andri Andri, Tengku Henny Febriana Harumy, Syahril Efendi

    The object of this research is a deep learning model based on recurrent neural network (RNN), long short-term memory (LSTM), and transformer, applied to predict financial asset prices using historical time series data. The main problem addressed is the absence of a systematic study evaluating the combined effect of weight initialization methods and activation functions in time-series prediction models, particularly regarding convergence speed, prediction accuracy, and the model’s ability to capture price variability. The results show that RNN and LSTM have better training stability, are able to converge in one epoch, and provide high prediction performance (RMSE < 3.7, MAPE < 0.015, R2 close to 0.9999). In contrast, Transformer showed lower prediction performance (RMSE around 37, MAPE around 0.58, R2 between 0.9884–0.9885) and tended to overfitting on various strategy combinations. In RNN and LSTM models, the All-Zeros (AZ) and ReLU combination specifically degrades stability and leads to overfitting. The superiority of RNN and LSTM is attributed to their sequential architectures, which are more effective at learning short-term temporal patterns and more robust to suboptimal weight initialization. Therefore, selecting an appropriate combination of weight initialization and activation function plays a key role in enhancing model performance. These findings contribute empirical evidence to the importance of configuration choices in deep learning for time-series forecasting. The results can be applied in the development of deep learning-based financial asset prediction and recommendation systems, particularly for assets with long historical records and volatile market conditions


  • Open Access

    Development of optimization algorithms to vehicle routing problem

    by Muhammad Amin, Syahril Efendi, Mahyuddin K. M. Nasution, Marischa Elveny

    The Capacitated Vehicle Routing Problem with Time-Dependent Demands (CVRPTD) is a significant optimization challenge in the logistics and transportation domain, characterized by dynamic customer demands, strict time windows, and heterogeneous vehicle fleets. This study focuses on urban parcel delivery operations as the primary object of research. The problem addressed involves the inefficiency of conventional vehicle routing strategies in adapting to time-varying customer demands and operational constraints, which often lead to increased costs and service delays. This study aims to minimize total operational costs while ensuring compliance with capacity constraints, service continuity, and demand fluctuations. A comprehensive mathematical model is developed based on a fully connected, directed acyclic graph G=(V, A), incorporating decision variables that represent vehicle routing sequences, timing, and vehicle type assignments. This study addresses the Capacitated Vehicle Routing Problem with Time-Dependent Demands (CVRPTD) in urban parcel delivery, where traditional routing methods struggle with dynamic demands and operational constraints. A mathematical model using a directed acyclic graph is developed, optimized via a gradient-based method with Hessian approximation, LU decomposition, and quasi-Newton techniques. Experiments on datasets with up to 200 customers and 20 vehicles with reductions ranging from 1.79% to 12.75%. The most significant improvement was observed in Sidorame Timur, where the optimization distance decreased by 12.75%, indicating high accuracy in route optimization. For the SCP, the proposed algorithm achieved a 6.46% improvement in solution quality over traditional greedy algorithms