Byung-Kwon Park

Systematic reviews (SRs) are critical for evidence-based research but are time-consuming and labor-intensive. The rapid expansion of academic publications further challenges the performance and applicability of existing screening and classification methods. While large language models (LLMs) present new opportunities for automation, limited research has examined whether they can achieve classification performance comparable to human reviewers in large-scale, multi-class settings. With the goal of improving classification performance, we proposed an LLM-based framework that leverages full-text key-insight extraction to enhance literature classification. We constructed a manually curated dataset of 900 articles from 17 published SRs to quantitatively evaluate the classification capabilities of LLMs. The results provided empirical evidence of LLMs’ potential in supporting large-scale SRs and introduced a practical pathway for improving efficiency and reliability in evidence synthesis. Empirical results showed that key-insight-based classification (KBC) significantly outperforms abstract-based classification (ABC). We implemented a confidence-weighted voting (CWV) mechanism using multiple LLMs to improve robustness. The CWV method achieved the highest macro F 1-score of 0.796, substantially exceeding KBC (0.732), ABC (0.676), and unsupervised K-means clustering (0.446). By employing zero-shot LLMs, our approach demonstrated the potential for enhanced adaptability across diverse domains and classification tasks without requiring fine-tuning, demonstrating that a carefully designed pipeline can enable LLMs to achieve classification performance comparable to human reviewers.

The accurate identification of collision positions on containers is critical in logistics and trade for enhancing cargo safety and determining accident liability. Traditional visual inspection methods are labor-intensive, time-consuming, and costly. This study leverages data from an Inertial Measurement Unit sensor and evaluates combinations of machine learning models and feature selection methods to identify the optimal approach for collision position detection. Five machine learning models (decision tree, k-nearest neighbors, support vector machine, random forest, and extreme gradient boosting) and five feature selection methods (Pearson’s correlation coefficient, mutual information, sequential forward selection, sequential backward selection, and extra trees) were assessed using three performance metrics: accuracy, execution time, and CPU utilization. Statistical analysis with the Friedman test confirmed significant differences in model and feature selection performance. The combination of k-nearest neighbors and extra trees achieved the highest accuracy of 97.1%, demonstrating that inexpensive IMU acceleration data can provide a cost-effective, efficient, and reliable solution for collision detection. This has strong practical implications for improving accident accountability and reducing inspection costs in the logistics industry.

Recently, the maritime logistics industry has been transitioning to smart logistics by leveraging such technologies as AI and IoT. In particular, maritime big data plays a significant role in providing various services, including ship operation monitoring and greenhouse gas emissions assessment, and is considered essential for delivering maritime logistics services. Marine big data comprise real-world data collected during ship operations, but it is susceptible to loss due to temporal and environmental constraints. To address this issue, an Emulator is proposed to generate supplemental data, including location data, data count, and average distance, using accumulated maritime transport data. This study proposes an Emulator that repetitively generates new data such as location data, data count, and average distance using maritime transport data accumulated up to now. The location data is generated using the cumulative distance and trigonometric ratios based on the location information of standard routes. The data count and average distance are calculated based on user-input parameters such as voyage time and data interval. The generated data is inserted into a database and monitored on a map in real time. Experiments were conducted using maritime transport route data, and the results validated the effectiveness of the Emulator.

Recently, the maritime logistics industry has been transitioning to smart logistics by leveraging such technologies as AI and IoT. In particular, maritime big data plays a significant role in providing various services, including ship operation monitoring and greenhouse gas emissions assessment, and is considered essential for delivering maritime logistics services. Marine big data comprises real-world data collected during ship operations, but it is susceptible to loss due to temporal and environmental constraints. Together with identifying and addressing the root causes of data loss, it is mandatory to supplement the loss by analyzing and utilizing the collected data. This study proposes an emulator that repetitively generates new data such as location data, data count, and average distance using maritime transport data cumulated up to now. The location data is generated using the cumulative distance and trigonometric ratios based on the location information of standard routes. The data count and average distance are calculated based on user-input parameters such as voyage time and data interval. The generated data is inserted into a database and monitored on a map in real time. To evaluate the emulator's performance, experiments were conducted using the maritime transport route data, and the results demonstrated its effectiveness.

The exponential growth of scientific articles has presented challenges in information organization and extraction. Automation is urgently needed to streamline literature reviews and enhance insight extraction. We explore the potential of Large Language Models (LLMs) in key-insights extraction from scientific articles, including OpenAI’s GPT-4.0, MistralAI’s Mixtral 8 × 7B, 01AI’s Yi, and InternLM’s InternLM2. We have developed an article-level key-insight extraction system based on LLMs, calling it ArticleLLM. After evaluating the LLMs against manual benchmarks, we have enhanced their performance through fine-tuning. We propose a multi-actor LLM approach, merging the strengths of multiple fine-tuned LLMs to improve overall key-insight extraction performance. This work demonstrates not only the feasibility of LLMs in key-insight extraction, but also the effectiveness of cooperation of multiple fine-tuned LLMs, leading to efficient academic literature survey and knowledge discovery.

In the logistics and trade that are highly dependent on containers, efficient identification of col-lided positions is of great significance for enhancing cargo safety supervision and accident respon-sibility. Traditional methods that rely on visual inspections require a lot of manpower, are time-consuming and costly. This study proposes a machine learning-based system to identify col-lided positions using the data collected through accelerometers installed on container doors. This study also uses feature selection techniques to reduce the data dimensionality, thereby improving computational efficiency and reducing computational costs. The feature selection methods evalu-ated include: Pearson Correlation Coefficient, Mutual Information, Sequential Forward Selection, Sequential Backward Selection, and Extreme Tree. The machine learning models evaluated in-clude Decision Tree, K-Nearest Neighbor, Support Vector Machine, Random Forest, and Extreme Gradient Boosting. Experimental results show that feature selection effectively reduces data di-mensionality and computational cost while maintaining or improving classification accuracy. The best combination of classification model and feature selection method is the combination of K-Nearest Neighbor and Extreme Tree, which achieved the best balance in terms of accuracy 0.9712, execution time 0.0300 seconds and CPU utilization 0.10%. It has a considerably practical significance in the logistics field in terms of safer and more efficient cargo transportation, reliable accident responsibility identification with reduced computing costs.

Real-time global positioning is important for container-based logistics. However, a challenge in real-time global positioning arises from the frequency of both global positioning system (GPS) calls and GPS-denied environments during transportation. This paper proposes a novel system named ConGPS that integrates both inertial sensor and electronic map data. ConGPS estimates the speed and heading direction of a moving container based on the inertial sensor data, the container trajectory, and the speed limit information provided by an electronic map. The directional information from magnetometers, coupled with map-matching algorithms, is employed to compute container trajectories and current positions. ConGPS significantly reduces the frequency of GPS calls required to maintain an accurate current position. To evaluate the accuracy of the system, 280 min of driving data, covering a distance of 360 km, are collected. The results demonstrate that ConGPS can maintain positioning accuracy within a GPS-call interval of 15 min, even if using low-cost inertial sensors in GPS-denied environments.

The core of dropout prediction lies in the selection of predictive models and feature tables. Machine learning models have been shown to predict student dropouts accurately. Because students may drop out of school in any semester, the student history data recorded in the academic management system would have a different length. The different length of student history data poses a challenge for generating feature tables. Most current studies predict student dropouts in the first academic year and therefore avoid discussing this issue. The central assumption of these studies is that more than 50% of dropouts will leave school in the first academic year. However, in our study, we found the distribution of dropouts is evenly distributed in all academic years based on the dataset from a Korean university. This result suggests that Korean students’ data characteristics included in our dataset may differ from those of other developed countries. More specifically, the result that dropouts are evenly distributed throughout the academic years indicates the importance of a dropout prediction for the students in any academic year. Based on this, we explore the universal feature tables applicable to dropout prediction for university students in any academic year. We design several feature tables and compare the performance of six machine learning models on these feature tables. We find that the mean value-based feature table exhibits better generalization, and the model based on the gradient boosting technique performs better than other models. This result reveals the importance of students’ historical information in predicting dropout.

Cryptocurrencies are highly volatile investment assets and are difficult to predict. In this study, various cryptocurrency data are used as features to predict the log-return price of major cryptocurrencies. The original contribution of this study is the selection of the most influential major features for each cryptocurrency using the volatility features of cryptocurrency, derived from the autoregressive conditional heteroskedasticity (ARCH) and generalized autoregressive conditional heteroskedasticity (GARCH) models, along with the closing price of the cryptocurrency. In addition, we sought to predict the log-return price of cryptocurrencies by implementing various types of time-series model. Based on the selected major features, the log-return price of cryptocurrency was predicted through the autoregressive integrated moving average (ARIMA) time-series prediction model and the artificial neural network-based time-series prediction model. As a result of log-return price prediction, the neural-network-based time-series prediction models showed superior predictive power compared to the traditional time-series prediction model.

Current step-count estimation techniques use either an accelerometer or gyroscope sensors to calculate the number of steps. However, because of smartphones unfixed placement and direction, their accuracy is insufficient. It is necessary to consider the impact of the carrying position on the accuracy of the pedometer algorithm, because of people carry their smartphones in various positions. Therefore, this study proposes a carrying-position independent ensemble step-counting algorithm suitable for unconstrained smartphones in different carrying positions. The proposed ensemble algorithm comprises a classification algorithm that identifies the carrying position of the smartphone, and a regression algorithm that considers the identified carrying position and calculates the number of steps. Furthermore, a data acquisition system that collects (i) label data in the form of the number of steps estimated from the Force Sensitive Resistor (FSR) sensors, and (ii) input data in the form of the three-axis acceleration data obtained from the smartphones is also proposed. The obtained data were used to allow the machine learning algorithms to fit the signal features of the different carrying positions. The reliability of the proposed ensemble algorithms, comprising a random forest classifier and a regression model, was comparatively evaluated with a commercial pedometer application. The results indicated that the proposed ensemble algorithm provides higher accuracy, ranging from 98.1% to 98.8%, at self-paced walking speed than the commercial pedometer application, and the machine learning-based ensemble algorithms can effectively and accurately predict step counts under different smart phone carrying positions.