New adverse events not detected in pre-marketing clinical trials sometimes emerge after medicines have entered the market. Therefore, systems that can extract the terminology of diseases and symptoms from unstructured data such as medical records are needed to enable automatic monitoring of adverse events. In this study, we analyzed nursing records of patients treated with vancomycin hydrochloride using a disease extraction system and a terminology list of allergic skin symptoms to identify cases with these symptoms and their occurrence frequency in a period with many occurrences and other period. The accuracy of the proposed method in detecting cases was assessed using the F-measure, which was found to be 0.71-0.74, and changes in the occurrence frequency between two periods were almost equal for the proposed method and visual confirmation by the researcher. These results suggest that the proposed method can be used for monitoring allergic skin symptoms due to vancomycin hydrochloride.

Medication management in residential care facilities is fraught with challenges, particularly concerning the occurrence of incidents when non-medical staff assist residents with their medications. To mitigate incidents, understanding the root causes is essential, typically through incident report analysis. Our study developed a multi-label classifier to identify factors contributing to medication-related incidents in residential care facilities from 7,121 incident report descriptions. Nine factors were identified: procedure adherence, medication, resident, resident family, non-medical staff, medical staff, team, environment, and organizational management. Multiple labels were assigned to each description. Due to the scarce labels for resident family and non-medical staff, these were excluded for model development. We fine-tuned three pre-trained models (two BERT and one ELECTRA), all achieving promising results. The F1 scores exceeded 0.6 across most categories and exact match accuracy also exceeded 0.6, demonstrating the effectiveness of our model in identifying factors from medication-related incident reports in residential care facilities.

It has been a while since the importance of listening to patient real voice was highlighted for better quality of adverse event (AE) evaluation. Our laboratory has reported novel deep leaning (DL) models that detect AE signals for hand-foot syndrome (HFS) or AEs limiting patients’ daily lives (AE-L) from cancer patient authored narratives. This study was designed to evaluate the DL models by applying them to S records in pharmaceutical care records following SOAP format, identifying characteristics and utility of the DL models. From 30,784 S records for 2,479 patients with at least one prescription history of anticancer drugs, our DL models extracted true AE signals with more than 80% accuracy for both HFS and AE-L, being also able to screen important AE signals requiring medical support from healthcare professionals. Our DL models could screen clinically important AE signals that would require intervention to treat the symptoms.

In Soccer penalty shootouts, it has been shown that the success rate of kicks varies depending on environmental factors such as the order of kicks and the prestige of the competition. In this study, we focused on the influence of such environmental factors and examined the possibility that leading or not leading situation influenced the success rate. We also attempted to elucidate the background against which this influence occurred. The results did not support the hypothesis that the success rate varied depending on the leading or not leading situation, but revealed that kickers were more likely to aim for the lower goalmouth in situations where they were assumed to feel strong pressure due to environmental factors. It was also confirmed that the success rate was higher when aiming at the upper goalmouth. These results suggest that environmental factors may influence the success rate through which zone (upper or lower) to target.

The present bias is a bias that leads to an overestimation of immediate costs and benefits, and can hinder the success of human behavior to achieve long-term goals. Methodologies to support goal attainment by mathematically modeling the present bias and providing appropriate interventions based on the present bias have become an important research subject in the field of behavioral economics. Existing studies have proposed mathematical models that are easy to handle analytically, but these models can only handle discrete time intervals and are limited in the discount functions that they can handle. In this study, we propose an extension of this model to continuous time using a variational method. The proposed model not only describes human behavior in continuous time, but also has the advantage of being able to handle a wide class of discount functions. Furthermore, we consider the problem of intervention optimization in this model and theoretically derive the optimal intervention strategy.

This study proposes a method for predicting the extent of human damage caused by bears from the situation when one encounter the bear using a large language model. As the situation of encouter, we utilize the date and time, location, sex and age of the victim, number of people at the time of the encounter with the bear. In addition, we textualize past findings on bears' attacks. We consider multiple labels as human damages such as death, serious injury, minor injury and sights. We conducted finetuning of Japanese BERT using the labeled encounter situations as training data. We evaluated proposed method on the human damage data by bears in Hokkaido and some areas in Honshu from 2021 to 2023. We confirmed that the proposed method improved the accuracy compared to machine learning methods.

In recent years, the development of the data distribution market has led to active data transactions between different companies and organizations. When trading data, ensuring the quality of data meets the request of purchasers is essential. However, the need for standard indices for assessing the data quality has resulted in variability in quality evaluations of individuals' background knowledge, skills, or experiences. In this paper, we proposed the "quality metadata," which is the automated visualization for supporting the evaluation of the data quality by inputting the actual datasets. The experimental findings indicated that evaluators' tendencies in assessing data quality varied based on their data analysis experience and their use of quality metadata by different evaluation indices.

We propose a framework for generating datasets that can appropriately evaluate the performance of the proposed algorithms in competitions. In most competitions which students and engineers studying machine learning participate in, the dataset is selected ad hoc. Therefore, there has been an issue, such as the use of dataset that would yield high performance no matter what algorithm is used. To resolve these problems, we conbine Item Response Theory and Conditional VAE. Item Response Theory is a theory for creating test questions and evaluating examinees' abilities. Conditional VAE is a method for generating images according to parameters. Experimental results show that our method generates a dataset which can evaluate the performance of algorithms appropriately more than MNIST.

Knowledge Tracing (KT) has been studied actively to facilitate effective student learning with optimal support based on student learning data. Important tasks of KT are tracing the evolving abilities of students and predicting their performance accurately. Recently, Deep item response theory (Deep-IRT) methods combining deep learning and item response theory have been proposed to provide educational parameter interpretability and to achieve accurate performance prediction. However, earlier Deep-IRTs estimate a student’s ability value using only a most recent latent ability parameter. Because current ability estimates cannot reflect past ability history data adequately, the parameter interpretability and the performance prediction accuracy might be impaired or biased. To overcome this difficulty, we propose a new DeepIRT with a temporal convolutional network that convolves past multidimensional ability states. The proposed method stores the student’s latent multi-dimensional abilities at each time point and produces a result that comprehensively reflects the long-term ability history data during performance prediction.

With the spread of computer-based testings and learning systems, it becomes possible to collect examinees' response data and addressing times that cannot be obtained from paper tests. The previous research pointed out that predicting the examinees' addressing times is important for adaptive learning. This study proposed Deep-IRT to predict the examinees' addressing times based on the Deep-IRT method which provides high accuracy of examinees' response prediction and the parameter interpretability. The proposed method predicts examinees' addressing times by two independent networks: an examinees' speed network and an item network. Empirical experiments demonstrate the effectiveness of the proposed method.

Causal inference using large language models (LLMs) has become an important research topic in recent years. In addition, research and development on prompt engineering has been actively conducted to improve the accuracy of LLMs responses. In particular, metacognitive prompting that apply human introspective thinking are known to significantly improve response accuracy in various tasks. In this study, we evaluate the effectiveness of metacognitive prompting on necessary/sufficient cause decision problems. The results show that metacognitive prompting was not necessarily effective. On the other hand, it is found that we can lead to the correct answers to the judgment problems which cannot be solved at all by using the metacognitive prompting, by providing multiple examples of similar problems with correct answers.

In recent years, Large Language Models (LLMs) have been developed and have shown high performance on various tasks. This high level of performance is achieved by learning from large corpora of documents written by humans. However, since humans are subject to various cognitive biases, leading to irrational judgments, LLMs can also be influenced by these cognitive biases resulting in irrational decision-making. For example, changing the order of options in multiple-choice questions affects the performance of LLMs due to order bias. Our research aims to mitigate such cognitive biases and prompt LLMs to make rational decisions. In our proposed methods, we apply cognitive biases mitigation methods used in crowdsourcing to prompts that are input to LLMs. To test the effectiveness of our methods, we conduct experiments on GPT-3.5 and GPT-4 to evaluate the influence of six cognitive biases on the outputs before and after applying our methods. The results showed that our proposed method can mitigate the influence of cognitive biases.

As Artificial Intelligence (AI) achieves high predictive accuracy, its utilization in supporting human predictive tasks has advanced significantly. As AI becomes more advanced, humans are challenged to comprehend and retrace how the algorithm came to a result. Explainable AI (XAI) has been developed to bridge this gap by providing rational explanations to help comprehension. Despite this, it remains unclear what constitutes an effective explanation to foster trust in AI. This study focuses on two factors affecting AI trust, the importance of decision outcomes and decision-making content, to explore how the basis for human judgment without AI. Our findings suggest that the necessity of explanation for AI trust varies with the context of AI use, indicating that the explanation needed to gain human trust differs according to the scenario with AI.

To apply Large Language Models (LLMs) in the real world, it is crucial that the text they generate is of value to humans and of a quality that is acceptable to humans. This study aims to find evaluation functions that correlate with human evaluations of fashion coordination descriptions generated by LLMs. Identifying such evaluation functions could allow for the improvement of the accuracy of fashion coordination description generation models in a direction aligned with human values, and potentially automate the entire process from description generation to evaluation. In this research, fashion coordination descriptions generated by LLMs were evaluated by skilled fashion stylists, and a dataset was created based on their evaluation. Using this dataset, we sought to find evaluation metrics that correlate with human evaluations. The candidates for these functions were functions used in the abstractive summarization task.

Large language models (LLMs) often inherit biases from vast amounts of training corpora. Traditional debiasing methods, while effective to some extent, do not completely eliminate memorized biases and toxicity in LLMs. In this paper, we introduce a novel approach to debiasing in LLMs based on unlearning techniques by performing gradient ascent on hate speech against minority groups, i.e. minimizing the likelihood of biased or toxic content. Specifically, we propose a mask language modeling unlearning technique, which unlearns the harmful part of the text. This method enables LLMs to selectively forget and disassociate from biased and harmful content. Experimental results demonstrate the effectiveness of our approach in diminishing bias while maintaining the language modeling abilities. Surprisingly, the results also unveil an unexpected potential for cross-domain transfer unlearning: debiasing in one bias form (e.g. gender) may contribute to mitigating others (e.g. race and religion).

Crowdsourcing is a popular way to obtain labeled data at relatively low cost. Existing methods for inferring true labels from crowdsourced noisy annotations assume that there is one true label for each item, but sometimes it is more appropriate to assume that each item has multiple true labels depending on the attributes of the workers. In this study, we propose a model that estimates the abilities and labels based on worker attributes, which considers a diversity of labels. The proposed model is an extension of the Dawid-Skene model and assumes that there is a true label for each combination of attributes. Experiments on synthesis data, where each item had multiple true labels depending on the attributes of the workers, showed that the existing model underestimated the ability of a minority group, but the proposed method accurately estimated it.

In recent years, large multimodal models(LMMs) have demonstrated innovative performance in various tasks such as text analysis, transcription, and optical character recognition. On the other hand, the data usage policies of LMMs depend on developers and thus there is a potential risk that confidential information could be stored or used as training data. Various LMM applications have been explored in academia and industry, and there is a growing demand for technologies that utilize LMMs while preserving data privacy. One such application of LMMs is the automation of data annotation tasks. Traditional data annotation is performed manually, requiring much time and cost. Also, the quality of the annotation is highly dependent on the individual abilities of each annotator. Therefore, LMMs have been expected to be faster and more accurate annotation resources. In our study, we propose a framework that balances annotation accuracy and data privacy preservation in data annotation tasks. Our experiments employed LMMs for image annotation and showed a reduction in privacy leakage risks while maintaining annotation accuracy.

In this study, we conduct a symbol emergent communication experiment in which human-machine pairs play a joint attention naming game (JA-NG). In this experiment, we will investigate whether distributed Bayesian inference can be performed between human participants and computers. As a result, we confirmed that ARI increases when humans and computers play JA-NG. We also confirmed that humans who performed JA-NG with a model following the acceptance probability based on the Metropolis-Hastings method had a higher ARI and converged to a more correct posterior distribution of signs than humans who performed JA-NG with other comparative models. These results suggest that humans can also play the Metropolis-Hastings Naming Game (MHNG), which allows for a certain degree of integration of perceptual information between humans and machines. This supports the idea that MHNG is a valid theory to explain human symbol emergence, and at the same time, it indicates the possibility that humans and machines can learn together in a co-creative manner. Since this experiment was conducted and analyzed with a small number of participants 9, further recruitment of participants and statistical testing is necessary to gain confidence in the findings.

The social impact of the spread of COVID-19 infection continues to be significant. Signs of the spread of infection serves as a basis for various decision-making. However, when COVID-19 was reclassified as a Class-5 infection, the daily full count became a weekly sentinel reporting. Daily data is required to calculate the effective reproduction, which is one of the epidemiological indicators, making it difficult to calculate. In this study, we constructed a prediction model with the aim of understanding the spread of COVID-19 infection at an early stage based on the number of sentinel reports. From the number of weekly sentinel reports, the model could estimate the effective reproduction by predicting the number of infected people daily. A simulation conducted using data during the 8^th/9^th waves showed a correlation between the calculated effective reproduction and the subsequent spread of the epidemic, suggesting its usefulness in the early stages of the epidemic.

When people form impressions of others in face-to-face communication, gesture style (i.e. the way of gesturing) impacts their impressions, such as well-mannered, honest, and enthusiastic. In this study, we trained a deep-learning model for gesture style transfer using YouTube videos, and developed a system that generates CG avatar animations from motion data to which the style transfer was applied. We also proposed a method to select an appropriate style by fine-tuning a Large Language Model (LLM). Finally, a user study was conducted to verify whether the impression of the motion generated by applying the selected style was close to what the user intended.

Robots are used in a variety of environments, including factories, homes, hospitals, and schools. In some of these environments, robots are assembled by people before being used. At this time, people find more value in the robot than if they simply purchased it (IKEA effect). The more widely robots are used, the more important it becomes to improve the relationship between people. We focused on the IKEA effect before and after assembly work as a way to improve the relationship between humans and robots. In this study, we experimentally examine the effects of the working relationship before and after the assembly task and between participants on the robot, which is the object of empathy. The results showed that participants showed high empathy after assembling the robot.

This study explores the dynamics of internal state changes in agents through reinforcement learning to enhance socially adaptive behaviors in public interactions. Recognizing the necessity for communication robots to adjust their actions based on the behaviors and situations of surrounding individuals, we design agents capable of modifying their internal states in response to the internal states of others. By simulating interactions, we investigate the impact of these experiences on the agents' internal state changes. Our approach leverages Q-learning to model the adaptive changes in agents' internal states, focusing on interactions that necessitate consideration of others. In conclusion, our findings suggest that for scenarios requiring estimating others' latent internal states, merely inferring these states from momentary behaviors is inadequate. There is a need for models that incorporate temporal dimensions to make more accurate predictions.

In this paper, we propose a transit allocation method for Smart Access Vehicle Sercvice. This method is devised to enable efficient allocation of vehicles to demand that extends over multiple business areas. There are two types of proposed methods: the Last method and the Next method. Simulation experiments were conducted to compare the two proposed methods with the current allocation method. The results of the experiments show that the proposed methods have advantages in several criterion of usabilities. From experiment results, we could find that the Next method was superior than the Last method.

Evacuation planning during disasters is considered to be a trade-off between complexity and efficiency. In this paper, we formulate the efficiency and complexity as a multi-objective optimization problem and show that various analyses can be performed by clarifying the Pareto solution.

In dynamically changing and harsh disaster environments, such as the spread of fire, evacuation guidance systems must be adaptive and distributed to ensure continued functionality even in the event of partial system destruction or malfunction. This study introduces an efficient evacuation guidance method tailored for scenarios involving the spread of fire, utilizing distributed graph-based optimization algorithms. The proposed method aims to minimize total evacuation time while simultaneously directing the crowd away from the fire, all without relying on a central control mechanism. To evaluate the effectiveness of the method, an integrated simulation model encompassing fire spread, evacuation agents, and evacuation guidance signs has been developed. Simulation experiments show that the proposed method provides efficient evacuation guidance in a dynamic environment.

The Social Force Model, one of the most typical pedestrian simulation models, demonstrates excellent results in simple walking spaces. However, it can’t reproduce the actual behavior of pedestrians in walking spaces with complex geometry. In this study, the authors applied the Social Force Model to pedestrian simulation in walking spaces with complex geometry by dividing them into subareas. The simulation results showed that the proposed method reproduced the behaviors that the conventional SFM could not.

The increasing severity of environmental problems and natural disasters and the dramatic development of computational technology have led to a growing interest in pedestrian traffic simulation. In this study, we focus on pedestrian waiting behavior in pedestrian traffic simulation. Understanding waiting behavior is important because it is universal in most traffic situations and can cause bottlenecks in traffic flow. However, pedestrian waiting behavior is not explicitly addressed in the Social Force Model (SFM), one of the most popular pedestrian traffic simulation models. In this study, we decided to extend the SFM to represent pedestrians with orderly queue-type waiting behavior and to evaluate the impact of waiting pedestrians on traffic flow. This study made it possible to quantitatively clarify the delay of pedestrian traffic due to passage restrictions caused by waiting queue.

One-class classification based image anomaly detection aims to identify anomalies in images by using only normal images. Recent methods that use a memory bank to store deep feature representations of normal images have achieved high performance and are robust to data variation. However, to make the inference procedure efficient, feature reduction is usually conducted over the memory bank. Thus, statistical information may also be reduced. Therefore, we propose an anomaly detection method that combines feature statistics and clustering. Specifically, the anomalies are identified by a weighted anomaly score based on K-center clustering with feature reduction and Mahalanobis distance without feature reduction. As a result, the proposed computation of anomaly score incorporates both the idea of efficient nearest neighbor retrieval and the idea of out-of-distribution detection in the memory bank. In the experiments, we evaluate the anomaly detection performance at pixel level on the MVTec AD benchmark and show the superiority.

In the manufacturing process of industrial products such as electronic substrates, visual inspections are conducted to remove product defects. In recent visual inspection systems, anomaly detection using machine learning is known as a highly accurate method . While these methods are evaluated under ideal experimental settings with clean training data, the inclusion of noisy data can degrade accuracy under practical settings. In this paper, we propose the sample selection method for noisy training data in anomaly detection. We improve convenience of the conventional method, SoftPatch. SoftPatch does not identify the noisy data at the data-level because it performs selection at the image patch-level. It also cannot be applied to other anomaly detection methods. On the other hand, our proposed method identify the noisy data and applicable to other methods. Experimental results using industrial product data demonstrate that our proposed method maintains the accuracy as the conventional method while improving convenience.

Anomaly detection based on one-class classification, a crucial aspect in industrial manufacturing, aims to identify anomalous samples that deviate from patterns established exclusively from normal ones. Conventional methods using a memory bank generally address anomaly detection by storing normal patterns in advance. However, increasing the number of normal samples for training leads to high computational complexity. Furthermore, relying only on a random subsample can result in a weak performance when the subsample is not representative of the structure of the data. In this study, we propose a feature selection based on subspace structure for anomaly detection, which selects only the features representing normal patterns. Specifically, we capture normal patterns with a few samples by constructing the subspace structure based on Nearest Subspace Neighbor (NSN). Experimental evaluations demonstrate that our method maintains high anomaly detection performance despite employing a limited number of selected features.

Image classification using machine learning is well-researched, yet categorizing “kawaii” images from human sensibility remains challenging due to its subjective aspects. Previous studies on “kawaii” images have achieved 70.2% accuracy by extracting color and shape features. We proposes a classification method based on the constituent objects in “kawaii” images, which has not been used by previous research. In our experiments, we created feature vectors that not only quantitatively but also semantically represent the objects in the image, and input them to a machine learning learning-based classifier for classification. As a result, classification was possible with an accuracy of up to 71.9%. Experiments were also conducted with different conditions for searching from within the image, and the features and results were discussed.

Recent video diffusion models have utilized attention layers to extract temporal features. However, attention layers are limited by their memory consumption, which increases quadratically with sequence length. This limitation presents challenges when attempting to generate longer video sequences. To overcome this challenge, we propose leveraging state-space models (SSMs). SSMs have recently gained attention as viable alternatives due to their linear memory consumption relative to sequence length. In the experiments, we first evaluate our SSM-based model with UCF101. In this scenario, our approach outperforms attention-based models in terms of Fr'echet Video Distance (FVD). In addition, to investigate the potential of SSMs for longer video generation, we perform an experiment using the MineRL Navigate. In this setting, our SSM-based model can save memory consumption for longer sequences, while maintaining competitive FVD scores.

In machine learning, exploiting symmetry in data is crucial to improving both learning efficiency and accuracy. Symmetry detection algorithms in time series data have received much attention in unsupervised learning to uncover the core physical principles of the data. Most existing work focuses on basic two-dimensional symmetry and is inadequate to handle more complex forms, such as the three-dimensional rotations that are common in real-world scenarios. To overcome this limitation, we introduce a new model that can learn such complex symmetries in uniformly varying time series data. Unlike conventional approaches that exploit symmetries in data space, our model adopts a latent variable framework and assumes existing symmetries in this latent space. By applying the identifiability theory of nonlinear ICA, we theoretically and experimentally prove that the symmetries detected by our method are consistent with the true symmetries from time series data whose symmetries are broken in data space.

This paper evaluates and compares several multivariate time series clustering methods by incorporating technical indicators into time series data for clustering. We have proposed a method for clustering time series data by calculating technical indicators used in the financial sector for time series data and then compressing them to two dimensions using UMAP for clustering on a two-dimensional plane. In this paper, we create new artificial datasets to evaluate and compare our proposed UMAP-based method (UMAP SC) with multivariate time series clustering methods: GAK k-means, k-Shape, and Soft-DTW k-means.

Random Convolutional Kernel Transform (referred to as ROCKET) is an algorithm for learning and inference by a linear model that convolves random kernels on time-series data and uses the maximum value and the proportion of positive values of the inner products obtained for moving windows as the features. In this presentation, we propose a method to calculate for each inference which time periods the ROCKET model focuses on using the regression coefficients of the linear model corresponding to each feature. Numerical experiments using artificial data were conducted to verify the validity of the method. The results show that for some anomaly patterns, it is possible to identify the time period that contributed to the classification.

Purpose:Apply the Stochastic Block Model to reveal the latent structure of networks. By stacking layers, blocks are treated as upper nodes, allowing for the analysis of more complex network structures. Validate using both artificial and real financial data to detect structural changes during normal and abnormal periods and analyze the causes of anomalies. Method Overview:Estimate adjacency matrices from variables and train using Markov Chain Monte Carlo methods. Use the Akaike Information Criterion to estimate the optimal structure. For financial data, apply the Dynamic Time Warping and generate adjacency matrices based on co-occurrence relationships to estimate the structure with this model. Do anomaly detection by comparing structures during abnormal and normal periods. Results:Confirmed the effectiveness of the model with artificial data. Detected structural changes in financial time series data and achieved anomaly detection. The detected abnormal periods coincided with financial crises, and considerations were made about their causes.

In this paper, we focus on the residual returns that are not explained by the common factors in financial asset returns. We propose a novel method to extract well-behaved residual returns based on principal component analysis (PCA). Traditional PCA requires determining the number of common factors, presenting a trade-off: increasing the number reduces common factors but also increases the potential for noise. Our proposed method randomly divides returns into two groups, extracts factors (PC) from one, and estimates eigenvalues from the other. Then, by creating a projection matrix that aims to transform eigenvalues to the same level, the proposed method can extract residual returns with better and more stable properties than PCA. Finally, we demonstrate that our method is capable of extracting residual returns with desirable properties through analysis based on both synthetic and real market data.

This paper describes our study of AI player creation for the picture-guessing card game Dixit. In the game, the active player makes up a phrase (or sentence) that describes the image on one of her hand. The relation between the image and the phrase had better be moderate as she scores nothing if either all other players or no player guess correctly. As an extension of the previous research, we propose a method to give multiple meanings to images by selecting words whose vectors are similar to the result of adding vectors representing each group of words in the image. It was demonstrated through post-experimental questionnaires that more human-like associations were made compared to the previous research. We also inferred that humans can successfully associate words with low relevance to the image through two-step association. Currently, we are adding usable parts of speech and comparing the language models to implement two-step association.

This study deals with "Street Fighter League", which is official team league matches using the fighting game "Street Fighter series" released by CAPCOM. Machida and Adachi (FIT2023) evaluated players by the number of wins and losses per battle. Machida’s result is different from the official ranking. Under the rules of this tournament, it is conceivable that a player could strategically and intentionally lose one battle to win a match. We introduce the ratio of points gained and lost, such as K/D ratio in games such as FPS. In our schemes, among the 32 players, three of them changed their ranking by more than 10 compared to the official ranking, and one compared to the Machida’s scheme. We can find four strong players. Furthermore, the proposed method can be applied in the future when the number of participating players and e-sports tournaments increase, and different tournaments can be compared.

In recent years, there have been many studies attempting to analyze the narrative structure of entertainment media such as manga and novels, and to automatically generate narratives. In addition, there are various subgenres of horror works, and classification of subgenres has been conducted mainly from a literary perspective. However, few studies have analyzed and classified subgenres from a scientific perspective. In this study, a total of 111 films in three horror subgenres were selected and their endings were analyzed using factor analysis, chi-square test, and residual analysis to extract ending patterns and clarify differences among horror subgenres. As a result, three ending patterns were extracted from the factor analysis, chi-square test, and residual analysis, and the differences in endings among the three horror subgenres were clarified. In addition, we generated patterns for the ending parts of the stories.

While numerous guides to ideation exist, universally applicable, academically systematized methods remain elusive—especially in advertising, where individualistic approaches prevail. This research addresses the need for personalized ideation support, transcending advertising, with the potential to enhance work environments and elevate idea quality. In our experiment, 15 persons in idea-centric professions used a custom app to document 245 instances of ideation. Analysis focused on the interplay between the relaxed phase and the focused phase, revealing unique patterns in ideation and associated heart rate fluctuations. Ongoing data collection and analysis will inform the development of AI to bolster creativity.

Deep learning has been applied to optical music sheet recognition (OMR). However, OMR processing from various sheet-music images still lacks precision to be widely applicable. We propose a measure-based multimodal deep-learning-driven assembly (MMdA) method enabling end-to-end OMR processing from various images including inclined photo images. Using this method, measures are extracted using a deep-learning model, aligned, and resized to be used for inference of given musical-symbol components by using multiple deep-learning models in sequence or in parallel. The use of each standardized measure enables efficient training of the deep-learning models and accurate adjustment of five staff lines in each measure, which enables locally inclined sheet-music images to be precisely positioned. Thus, a score can be reproduced from the inclined image with the proposed MMdA method while current OMR applications cannot. Multiple musical-symbol-component deep-learning feature-category models with a small number of feature types can represent a diverse set of notes and other musical symbols including chords. The proposed MMdA method provides a solution to end-to-end OMR processing and enhances the utility of OMR of mobile phone- based sheet-music photo images.

The healthcare industry faces challenges such as handling a larger amount of sensitive personal information compared to other sectors and the significant impact that uncertain information can have on individuals. Therefore, we have developed voluntary guidelines in the healthcare industry with the aim of providing healthcare service providers using AI-generated data a set of checkpoints to self-assess and ensure that their services do not cause undue harm to users. These guidelines serve as a reference for businesses that aim to provide such services, helping them avoid providing users with unfair disadvantages.

In nursing homes, nursing care records are an indispensable source of information for understanding the condition of patients and improving operations. However, caregivers generally work under an overcrowded schedule, and because of the existence of skill differences, actual caregiving behaviors and observations are not fully reflected in the records. This study examines the discrepancies between actual caregiver behavior and documented records in nursing homes. A digital record of caregiver behavior will be digitally recorded for one large nursing home and reconciled with the home's caregiver records. Generalizing the record content for computerized matching will simplify the matching of actual behaviors to documented records. We proposed a new time study to obtain actual behavior and developed an application to record behavioral data based on it. To verify the effectiveness of the system, we conducted a behavior data acquisition experiment in a university and confirmed that the system could acquire sufficient behavior data to be matched with nursing care records. In addition, since the granularity of records differs between behavioral data and nursing care records, we proposed a method of aggregation.

In this study, we developed a smart mirror, a device for estimating and measuring mental health based on individual response characteristics through simple interactive interaction. More than 320 million people suffer from mental illness worldwide but delayed detection of mental illness cause severe psychosis and, worst of all, suicides.Research has shown that early detection and early intervention have a dominant effect on remission in many cases of mental illness, and daily monitoring is essential for early detection. Therefore, we proposed a mental health estimation method that quantifies individual response characteristics based on device-based measurement and uses labeling by physicians as supervised data. In this study, we used a smart mirror as a measurement device, as we hypothesized that measuring individual response characteristics from simple interactions would reduce the burden of measurement on the user and enable stable data measurement. The results show that mental health can be estimated by using a sequential estimation method that classifies negative and neutral/positive responses based on the individual response characteristics obtained through a short interaction. This study shows that smart mirrors provide an objective, efficient, and non-invasive approach to mental health estimation and new possibilities in mental health care.

Critically ill patients with several life-support tubes and equipment are admitted in intensive care units. Therefore, physical restraints are commonly used to prevent contingencies such as self-extubation. In this study, we attempted to detect postures leading to self-extubation from the security footage of the ward. For patient posture estimation, we used MediaPipe, which can detect the three-dimensional coordinates of each body part. Time series data representing changes in movement of each upper body part were obtained from 13 self-extubation videos for which consent was granted. The time series data were placed on the video for visual confirmation of the posture. In three cases where the deviation between the actual posture and the coordinate position estimated by MediaPipe was considered small, we confirmed whether changes in posture leading to self-extubation could be detected using the singular spectrum transformation method. Consequently, large change values were observed in all three cases at approximately the time they started the action of grabbing the tube. All three cases where large change values were obtained at the time of self-extubation had room lighting turned on, and they were all bright videos. Under certain conditions, the potential for detecting removal actions from image-based posture estimation was suggested.

Pancreatic cancer is one of the most difficult cancers to treat, and it's important to identify the relations in cancer advance and genes. One of the way is to analyze about the factors and the gene mutations that cause cancer advance from pancreatic cancer DNA information. Even if there are only a few metastases, this study found that pancreatic cancer advance is influenced by the site of metastasis and the number of mutations, and identify their genes. It applys PCA(principal component analysis) and VAE(Variational Autoencoder).

The digitalization of information such as specifications and inspection information, which were previously managed on paper, is now in progress for the purpose of improving the work efficiency and labor saving of civil engineers. On the other hand, many documents in the civil engineering field are in pdf format and come in a variety of formats. In some cases, scanned data of old documents are used as references, which cannot be handled by text extraction tools or optical character recognition (OCR) technology. In recent years, multimodal models have been used for OCR and document understanding, and it is expected that multimodal models will be used in the civil engineering field as well. In this study, we measure how well multimodal models can recognize and understand documents in the field of civil engineering that contain many technical terms and are written in Japanese. We also conduct a qualitative analysis and discuss the possibility of using multimodal models in the field of civil engineering.

Chaos Engineering (CE) is an engineering technique that improves the resiliency of communication networks by intentionally injecting faults in the networks and taking precautions in advance. Although some CE tools partially automate the CE workflow, the parts that require try-and-error along with domain knowledge and text understanding remain manual. Here, we propose ChaosEater, a system for automating the entire CE workflow with Large Language Models (LLMs). In this paper, we share the architecture and experimental results demonstrating our system's effectiveness in automating the CE.

Large Language Models (LLMs) are expected to solve the challenges current recommendation systems face. We investigate biases that arise when using LLMs as recommendation systems and verify the validity of existing calibration methods. We first demonstrate the rating bias caused by using LLMs in a few-shot manner on actual data. We then apply several calibration methods from other classification tasks, such as sentiment analysis and natural language inference, to mitigate this bias. Our findings reveal these methods to be insufficient for recommendation tasks. We further question the assumptions made by the existing methods and discuss strategies for improvement.