Jurnal Teknik Informatika dan Sistem Informasi

Application of Conventional Modeling and Deep Learning to Stock Data with Outliers

2025-06-13T12:21:10+00:00

Apple Inc. stock (AAPL), one of the leading technology companies, is one of the concerns of investors as it continues to see an increase in the number of users every year. Therefore, forecasting Apple's stock price is important to help investors mitigate risks and optimize investment decisions. This forecasting can be done using two main approaches, namely conventional approaches such as Autoregressive Integrated Moving Average (ARIMA) and deep learning-based approaches such as Long Short-term Memory Network (LSTM). This study aims to find the best model using both methods, as well as compare the accuracy of the models based on datasets with outliers and datasets with handled outliers. The dataset analyzed in this study comes from weekly AAPL stock closing price data for 500 periods, from January 26, 2015 to August 19, 2024 obtained from Yahoo Finance. This study obtained the ARIMA(1,1,1) model as the best model for both datasets, with the outlier-handled dataset producing better test MAPE, while the dataset with outliers had better training MAPE. The LSTM method produced smaller MAPE values than ARIMA, demonstrating its superiority in capturing the fluctuating patterns of the AAPL stock data. Outlier handling was shown to improve model accuracy, as seen in the outlier-handled dataset. This research provides insight into the effectiveness of statistical and deep learning methods in modeling stock prices, and emphasizes the importance of outlier handling in financial data analysis.

Length of Stay Prediction Analysis for Pulmonary Infection Patients using Classification Algorithms

2025-09-10T07:34:14+00:00

This study investigates the relationship between age, gender, and other factors in attributes with Length of Stay (LOS) in patients with pulmonary disease. The main objective of the study was to help predict the LOS of new patients presenting with the same diagnosis and to help reduce the cost of care related to the duration of hospital stay. The theory used in this study is that the factors of age, gender, diagnosis, leukocyte values and chest X-ray results can affect the duration of their stay in the hospital. Data for this study was obtained from the medical records of one of the hospitals in West Java during the study period for approximately 3 months. The methods and techniques used are Artificial Neural Network-MLP (ANN), naïve bayes, J48 and Random Tree to analyze and model the relationship between input variables (age, gender, secondary diagnoses and others) and output variables (LOS). The results of this study are expected to provide a better understanding of the factors that influence LOS in patients with pulmonary diseases, as well as contribute to the development of prediction methods that can help better patient management and clinical decision-making in hospitals.

Binary Logistic Regression and Support Vector Machine for Classifying the Human Development Index

2025-08-28T05:01:56+00:00

Binary Logistic Regression and Support Vector Machine (SVM) are two widely used classification methods in data analysis, especially for problems with categorical target variables. In this study, these two methods are compared to classify the Human Development Index (HDI) status of Indonesia in 2024. The initial data consists of five predictor variables, but after conducting a correlation analysis to avoid multicollinearity, only three variables were used in the modeling. The Synthetic Minority Over-Sampling Technique (SMOTE) was applied to address class imbalance. Binary Logistic Regression was chosen due to its good interpretability, while SVM was used as a comparison due to its robustness against outliers. Evaluation results show that Binary Logistic Regression achieved an accuracy of 87.85%, slightly higher than SVM, which reached 86.92%. Therefore, Binary Logistic Regression is considered more optimal in classifying HDI status on the data that has been balanced and simplified. This study contributes to the application of statistical methods and machine learning in supporting human development analysis based on data.

A Trade-Off Analysis of Efficiency and Stability in Predictive Microclimate Control

2025-09-28T18:46:53+00:00

Precise microclimate control is a crucial aspect in various IoT applications, yet commonly used reactive, threshold-based control strategies often prove to be inefficient. This study presents a simulation-based comparative analysis to quantitatively evaluate the performance between a reactive control strategy and a more intelligent predictive one. Using time-series humidity data from a Tropidolaemus sp. terrarium, a SARIMA forecasting model was developed and validated to drive the predictive controller. The performance of both strategies was then benchmarked in a simulation environment based on two key metrics: actuation efficiency and environmental stability. The results demonstrate that the predictive controller is significantly more efficient, reducing actuator activations by up to 47% compared to the reactive controller. However, this study reveals a fundamental trade-off: this efficiency is accompanied by a decrease in stability due to an overshoot phenomenon caused by a rigid control action mechanism. This study concludes that the superiority of proactive prediction must be synergized with adaptive action mechanisms to achieve holistic system optimality, while also presenting a simulation methodology as an efficient framework for evaluating intelligent control systems.

Improving Classification and Regression Tree Performance Using Bagging in Heart Disease Diagnosis

2025-10-27T15:49:36+00:00

Heart disease is one of the leading causes of death worldwide, necessitating fast and accurate diagnostic methods for effective prevention. One approach that can be used is data mining, particularly classification methods to analyze health data. The Classification and Regression Tree (CART) algorithm is known for its interpretability but has a drawback in terms of model stability against data variation. To address this issue, the Bootstrap Aggregating (Bagging) technique is applied to improve the model’s stability and accuracy. This study aims to implement and evaluate the effectiveness of the Bagging technique in enhancing the performance of the CART algorithm for heart disease diagnosis. The data used in this study consists of three datasets available on the Kaggle platform: Heart Disease, Heart Disease Cleveland, and Heart Disease Prediction. The model is built under two conditions: using default parameters and using parameters optimized through the Grid Search method. The research process includes data preprocessing (data type adjustment, handling missing values, and outlier detection), training of two types of classification models (single CART and CART with Bagging), and evaluation based on accuracy metrics. The results show that the application of the Bagging technique consistently improves the accuracy of the CART algorithm. Under default parameters, accuracy increased from 72.89% to 78% (Heart Disease), 81.89% to 85.78% (Heart Disease Cleveland), and 77.44% to 82.44% (Heart Disease Prediction). With tuned parameters, accuracy increased from 75% to 84% (Heart Disease), 77% to 83% (Heart Disease Cleveland), and remained at 83% (Heart Disease Prediction). Therefore, the Bagging technique is proven effective in enhancing the accuracy and stability of the CART model for heart disease diagnosis.

Analysis on the Implementation of an Indonesian Automatic Short Answer Grading System

2025-11-04T04:55:52+00:00

In the education field, many types of questions have been developed to measure the students understanding for the material that has been given, such as multiple choice, short answer, essay, and others. Assessment for essay-type questions often takes up a lot of assessors’ time. A solution to overcome this problem is the development of an automatic essay assessment system. This system is developed in various literature on Automatic Essay Scoring and Automatic Short Answer Grading. This study uses a Multiclass Support Vector Machine (SVM) model in building an automatic assessment system. There are several findings from the results of this research. For the dataset used in this research, a combination of unigram and bigram cosine similarity, type-token ratio, and word count ratio features implemented together with RBF kernel with γ = 100 produces the highest precision value at the validation stage. At the evaluation stage, with a precision metric value of 0.49 and RMSE of 2.77, this model is considered less accurate. This is because the KNN and Logistic Regression models have higher evaluation metric values. The Logistic Regression model is more recommended for this automatic short answer grading system, because this model can provide more balanced and accurate predictions based on the lowest RMSE value and the precision, recall, and f1-score values that tend to be stable.

Switching Hybrid Model for Handling User and Item Cold-Start

2026-01-15T07:26:04+00:00

Recommender systems face significant challenges under cold-start conditions, where information about users or items is still limited. This study proposes a hybrid switching approach that adaptively combines Content-Based Filtering (CBF), User-Based Collaborative Filtering (CF), and Item-Based CF based on the number of user and item interactions. The evaluation was conducted through cold-start scenario testing for a single user, accuracy measurement using RMSE and MAE with 5-Fold Cross-Validation, and adaptivity testing under varying levels of cold-start conditions (5%, 20%, and 50%). Experimental results show that the hybrid model effectively handles all cold-start scenarios by falling back to CBF or CF User-Based when data is insufficient, and opting for CF Item-Based when sufficient information is available. The model achieved the best performance with an average RMSE of 0.8165 and MAE of 0.6592, along with low standard deviations, indicating stable performance across folds. Furthermore, the hybrid system demonstrated dynamic adaptability to data completeness levels, with a gradual shift in fallback algorithm usage as cold-start severity increased. Therefore, the hybrid switching approach not only excels in accuracy but also offers flexibility and robustness, making it an effective solution for improving the quality of recommender systems in scenarios with incomplete data.

Optimizing Nutritional Needs for Wasted and Severely Wasted Toddlers Using Differential Evolution

2025-12-20T08:56:32+00:00

Abstract — Malnutrition in toddlers, particularly wasting and severe wasting, remains a significant challenge in Indonesia, particularly in the East Nusa Tenggara (NTT) province. This study aims to develop a daily food menu optimization system for wasted and severely wasted toddlers aged 12-59 months using the Differential Evolution (DE) algorithm. The system is designed to balance macronutrient (energy, protein, fat, carbohydrates, fiber) and micronutrient (calcium, iron, zinc, copper, phosphorus, vitamin C) needs. The utilized database consists of food items commonly found and easily accessible in NTT, categorized into staple foods, side dishes, vegetables, and fruits. The DE algorithm was implemented to generate optimal, varied, and affordable menu combinations. The results show that the DE algorithm successfully created balanced menu recommendations. The optimal configuration was achieved with a population size of 20 and 1,500 iterations, consistently producing valid menu solutions with efficient computation time. This system proves to be an effective tool for addressing toddler nutritional fulfillment by considering local food variety and affordability.

The Efficiency Paradox: User Perceptions and Systemic Barriers in the Implementation of Electronic Medical Records

2026-01-08T06:57:36+00:00

This study aims to analyze the implementation of electronic medical record systems in improving the efficiency of health services. The approach used is a qualitative method with data collection techniques through in-depth interviews and observations of several health service units. The results of the study indicate that this system has a positive impact on accelerating the process of recording and accessing patient medical data, as well as reducing the use of physical documents. However, several challenges were also identified, such as limited technical training for healthcare staff and unstable internet connectivity. The conclusion of this study is that the electronic medical record system has the potential to improve the efficiency of healthcare services, but it requires support from comprehensive training and adequate technological infrastructure.

Hybrid Fuzzy Logic and Profile Matching to Improve in Hypertension Drug Classification

2026-02-06T08:10:23+00:00

Classification of hypertension drugs has been carried out using various methods, but the combination of Fuzzy Logic and Profile Matching (F-PM) for hypertension drug classification has not been widely reported. This study develops a new proposal with a different approach, namely combining Fuzzy Logic with the Profile Matching method. This method was evaluated using fifty clinical datasets taken from www.kaggle.com. Experimental results show that the application of Fuzzy Logic to the Profile Matching method can increase accuracy by 20.18% or 98.39%. This study also compares it with other classification methods. The results of the performance comparison show that the proposed approach is superior. This approach can be a reference for many future studies.

Daily Nutritional Planning System Based on Portion Optimization and K-Means Clustering

2026-02-22T11:42:28+00:00

Abstract — Personalized daily nutritional planning is a complex challenge due to the difficulty of translating individual nutritional needs into accurate food portions, exacerbated by the high prevalence of dual nutritional burdens in Indonesia. This study aims to design and implement an intelligent daily nutrition Decision Support System (DSS) capable of generating measured menu recommendations. The research method employs a hybrid approach, integrating an expert system knowledge base (Mifflin-St Jeor, FAO, IOM rules) with an inference engine based on dynamic portion optimization using linear programming (PuLP). Furthermore, unsupervised machine learning (K-Means) is applied to cluster food items to generate educational nutritional labels. The system was implemented as a web application using Python Flask and tested through case studies and functional verification. The main finding shows that the optimization engine successfully generated daily meal plans with specific grammages that closely approximated the target calories and macronutrients (case study caloric deviation <3%). The K-Means integration also proved effective in providing functional labels (e.g., "Pure Protein", "Energy Dense") for food items. This study concludes that a hybrid architecture based on dynamic portion optimization can provide a diet planning tool that is more quantitatively accurate and informative than traditional qualitative approaches.

Digital Filter Impact on Convolutional Neural Network Performance for Environmental Sound Classification

2026-02-22T11:45:52+00:00

Often, telephony-style bandwidth restriction techniques are applied raw to environmental sound classification systems without sufficient validation. To test their effectiveness, this study evaluates the impact of various digital filters (Low-Pass, High-Pass, Band-Pass, Band-Stop) on CNN performance on the ESC-50 dataset. After establishing the Log-Mel Spectrogram as the best input feature (surpassing MFCC), experiments proved that standard Band-Pass filters (300-3400 Hz) and Low-Pass filters actually reduced accuracy. This confirms that environmental sounds require a broad frequency spectrum (broadband), especially at high frequencies. Positive findings were obtained from the use of a low-order High-Pass Filter (HPF) (FIR-32) with a cut-off of 1000 Hz, which successfully increased accuracy to 66.20% above the baseline. Spectral analysis shows that this configuration successfully removes low noise without triggering transient smearing (time distortion). Therefore, this study recommends low-order HPF as the new standard, while suggesting the use of adaptive filters (learnable filters) in the future.