Local railways are facing a decline in users, and additionally, they are challenged by the difficulty of accurately grasping actual usage conditions and latent demand—information that is essential for formulating policies aligned with local realities. This study proposes a method to address this issue by leveraging social media (SNS) data and large language models (LLM) to conduct large-scale analyses of users’ raw voices related to local railways. Specifically, we develop a technique to automatically extract usage purposes from SNS posts and apply it to multiple railway lines with differing demand characteristics for comparative analysis. Through this analysis, we aim to clarify differences and commonalities in demand patterns across lines, and particularly to visualize latent demand and usage characteristics in low-demand local railways. This method is expected to contribute to the formulation of effective policies that are grounded in the actual conditions of local railway systems.

In greenhouse cultivation, detailed monitoring of plant growth and properly controlling the cultivation environment are necessary to achieve stable production. In this study, we present an approach to measuring plant growth using first-person view (FPV) videos captured by growers during cultivation management. FPV videos often contain blur caused by motion and many frames that include non-plant areas such as aisles and equipment, making the extraction of images suitable for plant diagnosis a key challenge. To address this problem, we propose a method that combines FPV videos with the grower’s motion data, acquired by an IMU sensor, to automatically extract images suitable for plant diagnosis from FPV videos. We developed a prototype system and conducted experimental measurements targeting tomato cultivation in greenhouses. Using the collected data, we extracted target images based on growers’ activities and applied them to plant diagnosis. The results indicate that the proposed method has the potential to provide a low-cost, easy-to-use approach for assessing plant growth in greenhouses.

Tissue segmentation in pathological images requires high expertise and considerable effort. An interactive semantic segmentation tool was developed that sequentially learns annotation inputs from a user and refines predictions to improve accuracy. The tool utilizes a Random Forest classifier to segment Masson’s Trichrome-stained images into three classes: “Nucleus,” “Collagen,” and “Background.” Experiments to evaluate the effectiveness of the tool were conducted by single trained subject. Evaluation, based on Intersection over Union (IoU) and the number of annotation operations demonstrated that the tool maintained accuracy comparable to manual annotation while improving work efficiency by approximately 1/10. Although transfer learning was attempted to adapt the model to new images, a clear improvement in efficiency was not obtained. A limitation in nucleus recognition accuracy was suggested. Further improvements in both accuracy and efficiency are expected through the addition of shape-based features.

In data science, system modeling is required to handle huge amounts and types of data. In this paper, we apply a new feature space generation method to multi-class logistic regression. We consider the improvement in accuracy and changes in characteristics of the feature space. The improvement is due to the synergistic effect of high dimensionality caused by randomization and nonlinearization. We verify that this method has desirable properties due to the matrix rank and eigenvalue distribution by using specific numerical examples.

In recent years, the challenges faced by local communities have become increasingly complex, making it essential to accurately capture the opinions and needs of residents in order to address them. Against this background, cross tabulation, which aggregates responses by combining two survey questions, has been widely used as a method to reveal the relationships between items in questionnaire data. However, as the number of questions increases, the interpretation of cross-tabulation results becomes more difficult, raising the risk of overlooking important relationships. In this study, we propose a method that represents multiple cross-tabulation results as a tensor and classifies them by extracting major response patterns through nonnegative CP decomposition. Furthermore, we report the results obtained by applying the proposed method to resident survey data collected in Seika Town, Kyoto Prefecture.

A growing number of local governments are utilizing video-sharing platforms as a means of regional promotion, producing videos designed to showcase the unique attractions of their respective areas. On these platforms, viewers can freely post comments on the videos. While many comments express positive opinions and favorable impressions, some include defamatory or malicious content. This study analyzes comments posted on regional promotion videos produced by local governments using text mining techniques. By applying document clustering to the comments associated with highly rated videos, the study aims to identify the key themes and content characteristics that contribute to high evaluations.

We discuss parameter estimation of decaying oscillations using autoencoders. Decaying oscillations commonly occur in many physical systems and analyzing them can reveal the characteristics of the underlying physics. Rapidly decaying signals are challenging to analyze using conventional methods, and observational data often contain noise. Under such conditions, machine learning offers the potential for efficient and high-precision parameter estimation. In this study, we utilize the latent space of autoencoders to estimate the frequency, phase, decay time, amplitude of decaying oscillations in noise, and evaluate the estimation accuracy.

This study aims to identify knowledge gaps in parenting information by analyzing questions, answers, and search histories collected from a mother-focused social networking service (SNS). Specifically, we define knowledge gaps as topics that are frequently asked or searched yet remain insufficiently answered. First, question texts in the Q&A data are vectorized and clustered based on semantic similarity. Then, the answer rate and users’ search frequency within each cluster are quantified and visualized to detect such gaps. The analysis revealed that knowledge gaps were particularly prominent in areas such as infant sleep problems, weaning practices, fever management, and marital stress. This approach provides a structural understanding of mothers’ real information needs and has the potential to reveal blind spots in public information provision by administrative and medical institutions.

This paper proposes a support system that enhances users’ exercise motivation through non-verbal behavioral expressions by virtual audience robots in accordance with the difficulty and performance of the user’s exercise. Previous studies have shown that the presence of an audience can positively affect motivation through the audience effect and increased self-efficacy. In this paper, we developed a virtual reality (VR) environment where users perform boxercise while being observed by audience robots capable of expressing cheers, movements, and facial expressions. These behaviors are dynamically altered based on the user’s performance and exercise difficulty. Evaluation experiments confirmed that the audience robots’ behavior enhances the enjoyment of movement.

Onomatopoeia, a general term for mimetic words and sound-effect words, is said to be one effective means of conveying the speed or intensity of an action to others. Therefore, this research aims to investigate the effects of coaching using onomatopoeia on physical movements in sports coaching. Thin-film pressure sensors (Force Sensing Resistors) were embedded in handballs. Undergraduate and graduate students were given instructions using various onomatopoeic terms, and pressure data were measured when they gripped the ball. Additionally, their subjective interpretations of each onomatopoeic term were surveyed. The results showed that instructions using onomatopoeia influence movement through multiple factors, including both individual interpretation and physical expression. Furthermore, it was confirmed that repeated instruction led to changes in physical movement, bringing it closer to the intended interpretation.

In recent years, the effectiveness of virtual crossmatch (VXM) for evaluating antibody responses to donor human leukocyte antigens (HLA) in kidney transplantation has gained attention. In VXM, the amount of donor-specific antibody is inferred from the mean fluorescence intensity (MFI) obtained from interactions between HLA antigens immobilized on beads—mimicking donor lymphocytes—and recipient antibodies. Although the MFI reflects the antibody level against a given HLA, it in fact represents the aggregate of antibodies targeting finer structural units within the antigen, called eplets; antibodies bind at the level of eplets rather than whole HLA molecules. Thus, an HLA’s MFI can be understood as the sum of antibody bindings to multiple eplets present on that HLA, yet current analytic methods cannot deconvolve antibody contributions at the eplet level. To address this limitation, we propose applying an evolutionary algorithm to estimate eplet-level antibody contributions while simultaneously satisfying constraints based on least-squares error and maximum error, and to maintain solution diversity within the search space.

In this study, we aim to develop an interactive game-support agent that enhances user persistence and motivation by fostering a sense of rivalry tailored to the user’s play level, based on the three factors of rivalry proposed by Kilduff et al. The agent is designed to act not merely as an opponent but as a peer, thereby suppressing potential negative aspects of rivalry, such as hostility and irrational behavior, while promoting user skill improvement and continuous learning support. We employ Dobutsu Shogi (Animal Shogi) as the task, as it is fully solved and allows objective evaluation of gameplay and match design according to the user’s skill level. Through game evaluations and impression evaluations, we verify the effectiveness of the proposed agent.

Aircraft are regarded as a promising means for wide-area and maneuverable exploration of Mars. However, controlling high-speed vehicles under the uncertainties specific to the Martian environment remains challenging. This study conducts robust trajectory optimization to investigate the impact of wind disturbances considered as an uncertainty on the flight path of a Mars exploration aircraft. An evolutionary algorithm was employed for the optimization process, while Polynomial Chaos Expansion (PCE) was adopted to quantify uncertainties. For the evaluation, an aerodynamics–flight coupled simulation was performed by solving equations of motion while sequentially retrieving aerodynamic data from a pre-constructed database. The results of the robust optimization revealed that strong tailwinds promote nose-up attitudes, leading to an extension of flight time, while strong headwinds hinder nose-up maneuvers, resulting in shorter flight time and reduced range. These insights provide critical guidance for the control profiles and path-planning strategies of both Earth-based demonstration missions and future Mars mission designs.

This study proposes a piecewise modeling method for nonlinear control systems by optimal control performance. It is straightforward to construct piecewise multi-linear models as a convex combination of vertex values. However, it is very difficult to find the dividing positions of the piecewise regions because it is a nonlinear programming problem. In this paper, a piecewise modeling algorithm based on particle swarm optimization is proposed to find the multiple dividing positions using optimal control performance. An example is considered to demonstrate the effectiveness of the modeling method using numerical simulations.

Damage caused by wild animals and birds affects not only agricultural crops but also the natural environment itself. The Ministry of Agriculture, Forestry and Fisheries (MAFF) of Japan is promoting ”Smart Wildlife Damage Countermeasures” that utilize ICT, creating a demand for more effective methods to mitigate such damage. In this study, we developed an animal species identification model for infrared images using YOLO (You Only Look Once), a convolutional neural network-based object detection algorithm, to validate the effectiveness of training with infrared imagery. Concurrently, we trained the model using two datasets of different sizes to investigate the amount of image data required for effective learning and to consider practical deployment methods.

In this study, we investigated the applicability of CLIP, a vision-language model, to automate the affective evaluation of product images. For three product categories (chairs, cups, and pens), we calculated impression scores using CLIP using an ensemble of object-conditional prompts and analyzed the correlation with human subjective evaluations (Likert scales). Experimental results confirmed a moderate positive correlation with impression words related to visual atmosphere, such as “cute” and “casual,” demonstrating CLIP’s effectiveness. On the other hand, correlations with physical and cultural attributes, such as “heavy” and “formal,” were weak or even negative. Furthermore, correlation analysis of antonyms revealed that CLIP was unable to preserve semantic oppositional structures (e.g., heavy vs. light). These results suggest that CLIP is capable of expressing impressions of visual surface features, but has limitations in its ability to infer underlying physical properties from visual information.

With the rapid development of artificial intelligence, pose estimation has attracted attention not only for humans but also for animals. However, datasets for animals remain limited, and the high cost of annotation poses a major obstacle to practical deployment. This study investigates a method to construct animal pose estimation models with limited data by leveraging the abundance of pre-trained models for human pose estimation and applying task editing through Task Arithmetics. Specifically, difference vectors were derived from human and animal joint estimation tasks and applied as subtraction operations to examine the feasibility of knowledge transfer in low-sample domains. The experimental results suggest that smaller learning rates improve the effectiveness of Task Arithmetics, while differences in annotation policies have little impact on estimation accuracy.

With the growing importance of data utilization in contemporary society, the demand for data-driven decision-making has been steadily increasing. In such contexts, decision-making often involves evaluating multiple alternatives, necessitating comparative analysis. Effective comparison requires collecting information on each alternative and assessing them from a shared perspective. However, even when the items being compared are the same, the appropriate perspectives for comparison may vary depending on the analytical objective. This study proposes a system that supports the comparison of two items by dynamically extracting relevant perspectives tailored to the user’s intent. The system utilizes ChatGPT to gather information about each alternative and identifies shared perspectives to present their differences and similarities. Experimental results confirm that the similarities and differences provided by the system are effective in supporting meaningful and context-aware comparisons.

In this paper, we propose a method for utilizing images at the summary level in multimodal neural machine translation. Conventional models typically extract and use only the image information related to the next token being predicted, but we show that this approach can lead to over-translation. To address this issue, we propose MVNMT, a new model that employs image information to model the features of the entire sentence (summary) and integrates those features into the decoder. MVNMT uses a variational autoencoder to extract a shared latent representation from both text and images. Our experimental results demonstrate that MVNMT outperforms conventional text-only translation models in terms of translation metrics and effectively mitigates the over-translation problem compared to MNMT models that utilize image information at the token level.

In reinforcement learning (Q-learning), multiple agents can achieve efficient learning by updating the Q-table cooperatively while simultaneously trying in parallel. In this study, we propose a switching reinforcement learning model by simultaneously analyzing agent clustering and Q-learning for each cluster, assuming that multiple agents are solving a problem in several different environments. But it is also assumed to be unknown in which environment each agent is solving the problem. We calculate fuzzy memberships following the Fuzzy c-Means (FCM) method using the acquired gain based on the policy of each cluster as the clustering criterion, and learn the Q-table for each environment in parallel by updating the Q-values with membership weighting. In addition, by introducing deterministic annealing of the fuzziness degree of the partition, we achieve both robust model estimation and maximization of the acquired gain.