In the dynamic landscape of online social networks, recognizing sensitive content is essential for safeguarding user privacy, fostering inclusivity, and enhancing diversity awareness. Building on prior research, this study explores new dimensions and methodologies for detecting sensitive content. We examine the temporal evolution of sensitive content, revealing how patterns shift over time, and address cross-linguistic challenges, emphasizing cultural and contextual nuances in detection. We employ advanced machine learning techniques, including deep learning models and BERT that improve the accuracy and robustness of the detection procedure. In the experimental study, BERT transformer reported the best performance in detecting sensitive content in text. Additionally, we incorporate explainability techniques such as LIME and SHAP to provide deeper insights into the model's decision-making processes, ensuring predictions are interpretable and reliable. Our work enhances the theoretical framework of sensitive content detection in social networks and provide methods that are accurate and scalable and can facilitate the creation of user-centric interaction that prioritize privacy and user experience.
Due to the destruction caused by wildfires underscores the critical need for proactive wildfire management strategies. Particularly, early detection of wildfires can mitigate their impact on the environment and economic loss and save human lives. This research study builds upon our prior work on refining and improving identifying early-stage wildfires. Previous research focused on using a deep-learning model for detecting early wildfires. This study aims to classify the visibility levels of forest sensory images to improve the deep learning model. Due to the constraints of existing public wildfire detection datasets, we have created a specialized forest sensory images dataset tailored specifically for detecting wildfires in their initial stages. We have used an Open-source computer vision library (OpenCV) for image processing integrated with TensorFlow and Keras to create a hybrid deep learning model. The results showed a significant improvement in the accuracy from 84% to 87% in detecting early spots of fire in images captured by surveillance cameras or/satellites. The Convolutional Neural Network model has been evaluated by crucial classification metrics, including accuracy, precision, recall, and f1 values. The study aids in progressing the efficacy of early wildfire detection, furnishing valuable perspectives for urban centers and nations contending with the dangers presented by these catastrophic natural events.
In recent years, there has been concern in Japan that traditional culture cannot be passed on due to a lack of successors, resulting in decline. The tea ceremony, one of Japan's traditional cultures, has seen a decline in the number of participants year by year, and its continuation is at risk. We believe that instructors are indispensable in order to stop the decline in the tea ceremony population. By clarifying the characteristics of tea ceremony movements, we aim to construct a system that can replace instructors, and to help preserve and pass on the tea ceremony. Therefore, in this study, we quantitatively analyze the characteristics of the way the host's body moves during the tea ceremony based on the changes of velocity. Among the operations performed before making tea, the “Preparation,” “Handling the fukusa,” “Re-handling the fukusa,” and “Passing the chasen” are analyzed. The joints of the human body are determined from the 3D coordinates obtained during the operations using Kinect. The velocities of the obtained joint points are calculated, and the analysis is performed focusing on the velocity change. The results of the movement analysis revealed movements specific to experts.
In this paper, taking into account the limitation of not being able to touch or shine a laser on a part of the paintings, including the three-dimensional objects on display, the SfM method was used to acquire three-dimensional shapes from the images, texture as color information and diffuse reflection characteristics were added to the acquired three-dimensional data, and shading was also added in real time to create CG data. Using this CG data and High-Resolution images, museum staff were asked to evaluate the sensory reality of the CG images in comparison with the actual artworks.
At the National Institute for Fusion Science, high-temperature plasma is generated in the Large Helical Device (LHD). Numerous videos capturing plasma discharges during nuclear fusion experiments have been recorded. Analyzing these videos may unveil new discoveries and ensure the safe operation of experiments. Thus, estimating the future states of plasma light emission is essential for optimizing operational parameters in experiments. In current methods for plasma video prediction, a sequence of multiple frames is typically used as input to analyze visual movement trends and predict future plasma discharge conditions or generate future frames in plasma videos. Our study focuses on a more challenging task of video prediction using a single frame of plasma image. We propose a method for predicting sequential future frames containing the image of the visible light from the plasma based on a single given frame by using a conditional GAN model. We experimentally evaluate this approach and find that our proposed method accurately forecasts the future states of plasma images in each frame of the videos.
Data seamlessly becomes a critical component for the state estimation and performance optimization of Cyber-Physical Systems (CPSs). Ensuring timely access to CPS data could be challenging due to operation costs, limited accessibility, security and connectivity, etc. CPS simulators have been thoroughly used in industry to synthesize CPS data. However, a common drawback for CPS simulators is the low integrity and fidelity of the synthesized data to the actual system. In this paper, we propose an extension, calibration, and validation to enhance the data synthesis integrity of an industrial CPS simulator. Behavior patterns are inferred from actual data, using frequency analysis, and applied to the simulation data synthesis, where a calibration using cross-correlation is utilized to fine-tune the data synthesis. According to the analysis results, our enhanced simulation data synthesis demonstrates a correlation of 93.7% to the actual CPS data.
Complex environmental geographical spaces exhibit characteristics such as nonlinearity, multiscale nature, and uncertainty, representing typical spatiotemporal dynamic systems. With the advent of big data, data-driven spatiotemporal prediction models have emerged. However, the high-dimensional complexity of spatiotemporal phenomena makes these models difficult to interpret. This paper proposes a river network feature identification architecture based on coupled dynamic partial differential diffusion equations and graph neural networks. By deeply analyzing the solving process of hydrodynamic-diffusion equations and graph neural networks, spatial and nonlinear information of the environmental field is embedded into the graph network. We design a message-passing network based on an aggregation sampling method for flow field information fusion, allowing multiscale flow field information to be propagated, aggregated, and updated to predict the spatiotemporal evolution of water quality. The proposed model characterizes the constraint relationship between the rate of change of water quality concentration in time and space dimensions, enhancing the model's physical consistency and interpretability.
In data mining using big data, Customer Relationship Management (CRM) is widely practiced to estimate detailed customer needs. CRM primarily utilizes RFM analysis to implement appropriate strategies for each customer category, aiming to increase profits by enhancing customer satisfaction. Additionally, association analysis is employed to extract useful rules and adapt them to these strategies. In this study, we classify purchase history data by product and conduct quarterly association analyses to elucidate customers' purchasing tendencies. Moreover, by extracting association rules for each quarter, this study enables the analysis of seasonal product changes and allows for a more detailed estimation of customer needs.
We propose a method for estimating the position of a light source using Non-Photorealistic Rendering (NPR). Our method uses NPR and adaptive thresholding for preprocessing images, which are input to a Convolutional Neural Network (CNN) model. We train separate models for estimating parameters related to angles in the xy plane and angles in the xz plane, using datasets created in the virtual world on Unity. In addition, we test these models using various datasets created on Unity and datasets created in the real world. As a result, our models demonstrate high accuracy for estimating angles on the xy plane using virtual world datasets, and relatively high accuracy in using the real world dataset. However, for angle estimation on the xz plane, we found that there is room for improvement in both datasets compared to that on the xy plane.
The Budget-based Neighboring Object Group Query (BR-NOGQ) is a novel type of location-based query that considers both the spatial relationships between objects and the user's budget constraints on experiencing these objects. However, processing a large number of BR-NOGQs concurrently can overload a centralized system, leading to poor performance. To address this issue, this paper focuses on developing a distributed processing technique for answering multiple BR-NOGQs using the MapReduce platform. The proposed approach involves designing a grid structure to manage information about different types of objects and developing a MapReduce-based algorithm for efficient distributed processing of multiple BR-NOGQs. Experimental results using a synthetic dataset demonstrate the scalability and efficiency of the proposed algorithm.
Music surrounds us, and there is no denying that music in visual media can shape and evoke emotions. Yet, understanding how musical preference influences emotions through audio and visual stimuli remains an important task. To address this task, we investigated the role of musical preference effect on perceived emotions induced through music and visual stimuli (i.e., animation), using a 7-point scale for valence-arousal ratings and physiological responses in electroencephalogram (EEG) band power. The perceived emotions are categorized into four states: happiness, calmness, fear, and sadness. One emotional state contains 4 sessions: (1) Preferred Music, (2) Unfamiliar Music, (3) Preferred Music+Animation, and (4) Unfamiliar Music+ Animation. Behavior-wise, the rating results showed that the presence of preferred music resulted in higher perceived valence ratings, particularly in happiness. However, no stimulus had a significant effect on perceived arousal ratings and satisfaction ratings. Physiologically, the EEG indexes showed that the presence of preferred music appears to affect an increase in alpha power across various emotions except sadness, whereas unfamiliar music seems to affect beta power, particularly in happiness and calmness. Overall, these findings supported that musical preference is an affective factor reflecting the levels of valence and alpha power in EEG. Nevertheless, our findings did not confirm a significant difference between only preferred music and combining it with animation. Interestingly, we also found that participants were more likely to be attracted to and perceive positive (i.e., happiness and calmness) emotions easily through preferred music and/or visual stimuli than negative (i.e., fear) emotions.
This paper focuses on the reliability of open source software (OSS). In particular, the proposed method is based on the deep learning by using the fault big data obtained from the bug tracking system. The immune system of human body takes on the role of the protection for the pathogenic microorganisms. Recently, the source code and complexity of OSS system become large year by year. Then, we assume that the OSS fault removal system is similar to the pathogenic microorganisms. We propose the OSS fault removal system based on deep learning inspired by immune system in this paper.
Typically, conferences identify and award one or two of the accepted papers as the “best,” highlighting these selections as the premier contributions for that year. Despite their apparent merit, these peer-reviewed selected awards carry inherent limitations and biases. Critics argue that such awards often show favoritism toward renowned researchers and may reflect the preferences of a small group of selectors, usually two or three reviewers, rather than the broader community or conference attendees. These selections are primarily based on the content's comprehension of papers by peers rather than its reception of the author's presentation. This research evaluates the “best paper” awards at an international conference from the attendee's perspective. It details a social experiment conducted at an international conference, where attendees could discuss and vote on the “best paper” after presentations, utilizing an online open forum. The study compared award-winning papers, previously selected by PC members and conference attendees, with other papers from the same conference. The comparison focused on readability, voting stance, and engagement metrics for both full papers and short papers. The findings indicated that the award-winning paper exhibited slightly lower human engagement than that of non-award-winning papers. Furthermore, the regular papers' readability was higher than that of short papers. This research underscores the need for a more democratic, meaningful, inclusive, and insightful evaluation process in future scholarly assessments.
Temporal Fusion Video (TFV) is a technique that amalgamates multiple frames in temporal sequence to generate a cohesive video output, thereby facilitating various applications such as motion analysis, scene summarization, and the creation of unique visual effects. Despite its advantages in capturing fast-moving objects and efficiently summarizing long-duration changes, TFV is susceptible to ghosting artifacts-transparent overlapping images resulting from frame blending. These arti-facts compromise visual quality and hinder accurate analysis. This paper aims to mitigate the challenge posed by ghosting artifacts by proposing an AI model for quantifying them through measurement and regression analysis, with a specific focus on leveraging Automated Machine Learning (AutoML). Our methodology seeks to identify the key factors contributing to ghosting artifacts and model their impact systematically, thus enabling the development of targeted strategies for artifact reduction. The implications of our research are profound, as it offers a quantitative foundation for evaluating TFV methods, thereby enhancing standardization within the field. Moreover, our approach holds promise for improving the practical applications of TFV in domains such as surveillance, entertainment, and sci-entific visualization. By bridging the existing gap in research, our findings introduce a novel and objective method for enhancing the visual quality and applicability of TFV techniques.
In this study, we investigated the factors contributing to the sustained interest of core fans to foster investment in crowdfunding for anime production. A survey of 72 anime fans revealed that among those with low interest in crowdfunding, the quality of storytelling and composition, the quality of animation and video, and official events significantly influenced their continued interest. For those with high interest in crowdfunding, in addition to these factors, the exchange of opinions and information with others also had a strong influence on their sustained interest. Based on these findings, we proposed strategies for creating core fans in the anime fan community, including organizing official events as crowdfunding rewards, emphasizing the quality of the work, promoting the project's development, and providing opportunities for fans to share their opinions and information with others.
Cardiovascular disease is a non-communicable disease that is one of the leading causes of mortality worldwide, and its prevalence is growing year after year. Most cardiovascular diseases require percutaneous coronary intervention (PCI) to remove blood vessel blockages, However, this therapy carries risks of adverse outcomes, such as in-hospital bleeding and mortality. Currently, machine learning is widely used in prediction and decision-making support for clinicians. According to the above reasons, this study demonstrates using and comparing four common machine learning algorithms to predict in-hospital bleeding and in-hospital mortality outcomes. Finally, this study found that the XGBoost model accomplished the most outstanding in-hospital bleeding and mortality results. The performance of in-hospital bleeding shows Accuracy of 0.987 (95% CI: 0.980 - 0.994). In-hospital mortality shows Accuracy of 0.972 (95% CI: 0.962 - 0.982).
The rapid increase in popularity of co-working spaces, commonly used by freelancers and entrepreneurs, gives rise to the need to understand the requirements of their members. Benefits ranging from increased productivity and reduced commuting to cross disciplinary collaboration and networking are stated in the academic literature. A complementary perspective of whether their users agree with these benefits is gained by conducting an aspect-based sentiment analysis, topic modelling with Latent Dirichlet Allocation, and zero-shot text classification. This paper analyses almost 9,000 google reviews of 243 co-working spaces in the UK. The results show that the overall sentiment is positive with less than 7% negative or neutral reviews. The frequently mentioned topics include staff, the social and working atmosphere, community, food & drinks, location, and internet access.
Businesses that had been in decline due to the coronavirus are regaining momentum. Tourism is one such business that is expected to see an increase in domestic as well as international travelers. To attract more of these travelers, we need to rethink regional revitalization centered on tourism. In general, travelers' satisfaction with the destinations they visit varies from region to region, and therefore, measures that emphasize regional characteristics are likely to be effective. To this end, a framework and methodology is needed to examine what each region should emphasize. The purpose of this research is to develop the tourism industry, which is one of the local revitalization measures. This study proposes a method that uses a graphical modeling technique to identify factors that increase tourists' satisfaction with the target area and provides a framework for what municipalities and residents should focus on. The feasibility of this method was also confirmed by applying it to specific regions in Japan.
In recent years, television ratings have been on a declining trend. One reason for this is the regulation of extreme content, which has led to a homogenization of programming. Extreme content is deemed inappropriate for current societal conditions. However, many viewers feel that television has become boring due to the absence of such extreme content. Therefore, we conducted a survey regarding extreme content broadcast on television. Using an online questionnaire, we created a survey with ten questions based on content that viewers judged as extreme from past programs. Additionally, the survey targeted respondents aged 18 to 24, enabling us to examine content considered too extreme to be aired by a demographic known for having a higher tolerance for extreme material. The survey results indicated that actions likely to cause accidents or lifelong disabilities cannot be broadcast. It was also inferred that if the survey included all age groups, the range of acceptable content would likely be even narrower. In this study, we will conduct a survey that includes viewers of all ages. We will use survey tools to gather a wide range of data and perform logistic regression analysis to predict the factors that differentiate the broadcasting of extreme content. Then, based on the collected results and demographic data, we will consider the factors that lead to the rejection of extreme content.
In contemporary design education, the influx of diverse design accomplishments and methodologies, especially those involving interdisciplinary and cross-regional collaboration, poses a significant challenge. Understanding and expressing user and environmental needs within technology, aesthetics, and culture is central to addressing these issues. Since 2016, we have addressed this challenge through a joint design workshop course for 3rd and 4th-year students, aiming to foster cross-cultural understanding and communication skills, thus developing international expertise in design. This course employs a project-based learning (PBL) approach and is a staple in our department's curriculum. It involves mixed groups from our university and international partners working on specific design problems and presenting their solutions. The course, conducted primarily in English, emphasizes non-native linguistic and non-verbal communication through visual expressions like sketches and drawings. One objective is to provide students with the skills necessary for effective cross-cultural communication. Typically, this intensive summer course alternates annually between Japan and partner countries. However, the COVID-19 pandemic 2020 necessitated a shift to an online format to maintain the course's objectives despite travel restrictions. This paper highlights the adaptations made from FY2020 to FY2021 to continue the program online. Our international collaborations led to developing a PBL-based active learning method, crucial for enhancing cross-cultural understanding and communication in design education. Utilizing the technology and methods refined during the pandemic, we successfully conducted a hybrid online/on-site workshop connecting Japan and Taiwan. Online tools facilitated remote group work, discussions, and feedback, accommodating a diverse, multicultural cohort through synchronous and asynchronous communication.
