This study aimed to elucidate gender differences in information processing in the human brain when listening to music. We investigated regions of interest (ROI) related to both gender differences and individual differences by applying representational similarity analysis (RSA) to brain activity measured by functional magnetic resonance imaging (fMRI). In addition, we tested for significant differences between gender and individual differences in order to examine whether differences existed between ROIs related to gender differences and ROIs related to individual differences. As a result, we were able to find ROIs each of which had large gender differences and individual differences, but we were not able to find gender differences as more than individual differences with the present method.

Humans communicate by utilizing both external cues, such as speech and facial expression, which constitutes the information, as well as internal cues like emotions and feelings. Since the other person's internal information cannot be immediately observed, it must be guessed using the information that can be seen from the outside in order to promote smooth conversation. Bayesian inference is one technique for facilitating this estimation; it involves developing several hypotheses beforehand and selecting the best suitable one based on observed data. This method is very effective when the target of estimation is stationary, but when dealing with something that changes constantly, like human emotions, another approach is necessary. Therefore, the focus of this study is on an expanded Bayesian inference that takes into account the symmetry bias, a cognitive bias, and introduces a forgetting rate and learning rate. The efficacy of this estimating approach will be examined by simulating a communication game in which two decision-making agents, each equipped with this estimation method, estimate the other's internal generative models.

To solve large-scale and hard constraint satisfaction problems (CSPs), ant colony optimization (ACO) has been studied as one of the metaheuristics. Although ACO has been effective for solving CSPs, it has been sometimes difficult to solve large-scale combinatorial problems because of falling into the local optima. Lévy ACO, an ACO algorithm using Lévy Flight (LF), has been proposed to escape from local optima. Lévy ACO escapces from local optima by interweaving the local search with a global search using LF. However, the fixed parameter (LF parameter) that determines how often LF is used may reduce the efficiency of the search. In this paper, we propose a method to dynamically adjust the LF parameter of Lévy ACO according to the progress of the search. We demonstrate that our method can solve large-scale and hard graph coloring problems, which is one of CSPs, more efficiently than the previous methods.

Rule-mining algorithms require specific treatment for the rules wherein their items appear only a few times. Each rule-mining algorithm contains a tuning parameter related to the fewness of the related items. A typical method to determine this type of tuning parameter uses validation data. Since validation data are only available with knowledge of the correct rules, it is difficult to determine the parameter. Observing various histograms of the estimated strength, we find that the histograms should be smooth if the parameter is reasonable. This study proposed an unsupervised method to determine the tuning parameter for rule-mining tasks by the histograms of estimated results varying the parameter without knowledge of the correct rules.

Environmental issues such as exhaust emissions from traffic congestion and the spread of electric cars and self-driving cars are important topics in society, and people have been considering various traffic measures against these issues. On the other hand, it is hard to study the effects of traffic policies and laws on traffic flows by applying them to real space. We have developed a multi-agent traffic simulator, MACiMA, and have attempted to reproduce complex traffic phenomena and verify the effects of traffic measures through simulation. MACiMA represents traffic flows by the interaction of agents, but it has a heavy implementation burden to define the behaviour of each agents for the traffic policies in the program code. In this study, we investigated a mechanism for reasoning about the decision-making mechanism of agents based on traffic laws in conjunction with PROLEG, a legal reasoning system using the logic programming language Prolog. Experiments with the created simulator show that the agents dynamically decide their behaviour according to traffic laws and the effects on traffic caused by changes in their proportions.

Autonomous learning agents using online reinforcement learning learn strategies sequentially from state observations obtained from interactions with the environment and internally defined rewards. However, if the state transition changes due to the intervention of other agents, the agent may not be able to learn the strategy it originally wanted to learn or may be induced to learn a specific strategy. In this study, we propose an intervention algorithm and investigate its properties for such an intervention attack on the reinforcement learning process. We formulate the intervention by the intervention agent to the protagonist agent as a 2-player Markov Game, and find that when the protagonist is induced to learn a strategy that maximizes the reward intended by the interventionist, the intervention can fail even in situations where the protagonist always obtains the optimal strategy for his reward. Another problem arises in situations where the protagonist is in the process of learning, for which we devised an improved algorithm.

Hedonic games are mathematical models in which a group of agents is divided into appropriate subgroups, and have been studied as a field of cooperative games. Cooperative games with permission structures, on the other hand, are models in which an agent’s participation in a game is by permission of another agent. In this paper, we introduce a permission structure into SASHG, a type of hedonic game, and consider solutions to hedonic games in which information diffusion, i.e., the incentive to issue as many permissions as possible, holds. Specifically, we first show that Nash stable solutions and information diffusion are incompatible. Given this impossibility, we propose an algorithm with incentives for information diffusion and show the approximate rate of social surplus that can be achieved. As a result, we show the incompatibility theorem of social surplus maximization and Nash stability with incentives for information diffusion, and furthermore, we show that the achievable approximation rate is 0.

In the delivery services in office buildings using automated robots, pedestrians and a wide variety of robots coexist, and the robots travel along the aisles while avoiding these obstacles, so the time required to cross an aisle is not constant, and information on the distribution of travel time is often insufficient or inaccurate. This study proposes a method for estimating the parameters of the probability distribution of travel time using Bayesian inference for multi-agent path finding problems where the travel time varies stochastically, and the probability distribution is not given a priori but must be estimated based on information obtained during operation. Through simulations, we compare the performance of our method with existing methods.

In this study, the authors developed an agent system to support idea creation in SF prototyping. Among the workshop methods for creating the future, SF prototyping is useful for exploring potential possibilities by removing constraints on participants. However, appropriate facilitation is necessary for this purpose. The authors developed a virtual agent system that supports idea creation and analyzed the results. In the process of creating a future, the virtual agent system that supports idea generation that has a face like a person and operates on the screen, and the authors analyzed the results of the idea generation support. The virtual agent was developed to facilitate workshop with understanding the context. From experiments using this, it was found that the number of ideas of participants can be expected to be balanced in the new word creation phase, which is required to generate many ideas to diverge discussions. In addition, in the new character creation phase, which may affect the final product, it was shown that the number of character creations may increase compared to the case of human facilitation.

Home cares and facility cares are important in Japan since the issues arising from aging of the population of Japan will be expected to be socially very serious. During the process of discussions, the prototype of a care matching to optimize QOL will also need to make a care schedule based on the results of a care matching to apply our system to the real-world care facilities. The resilience for a request to change a schedule is going to be very important since a rescheduling often occur on a daily base care activity in the real world. Also, it needs to incorporate requirements from caregivers for scheduling because the process of making a care scheduling in the real world also incorporate requirements from caregivers In this paper, we discuss about our preliminary approach to realize a resilient scheduling system with a value co-creation mechanism.

We consider manipulation on facility location mechanisms which do not satisfy strategyproofness.Specifically, we deal with two mechanisms called the midpoint mechanism and the Nash mechanism. In the midpoint mechanism, the location is determined as half of the sum of the minimum and maximum values among the reported values.In the Nash mechanism, the location of facility is determined as that maximizing the product of utilities of the agents.Agents can improve their utility by manipulation in those mechanisms. In this paper, we investigate how one agent can manipulate the location of a facility in those mechanisms.

This study aims to explore effective methods for promoting university enrollment in regional areas by adjusting enrollment quotas and introducing new scholarships in university entrance exams. Despite attempts to limit university enrollment in urban areas, enrollment remains concentrated in these regions, so it is necessary to implement additional measures.To address this issue, we propose offering new scholarships targeted at urban students to encourage enrollment in regional universities. To evaluate the potential impact of these measures, we employ agent-based simulation to simulate university entrance exams and estimate the effects of quota distribution and scholarship allocation on the number of regional students. Our findings indicate that while the new scholarship alone has limited impact on regional enrollment, when combined with preferential quota allocation for regional universities, it leads to an increase in enrollment in regional core cities.

With the continuing threat of infectious diseases, it is important to visualize the risk of exposure to pathogens and the effects of hygiene behaviors in living situations for people's proper hygiene habits. Focusing on schools where many people gather, we estimate the risk of contact with the virus among students in order to reduce the risk of infection as a public space. A behavioral survey of students in the school is conducted to construct a simulation model. Targeting 4th grade elementary school students with consent, videos are taken with cameras installed in classrooms and corridors, and their behaviors are analyzed mainly before and after class and during breaks. It is found that the virus spreads widely in the classroom by calculating the transmission of the virus through hands from the movement history of the students and the order in which they touched the items. Analysis of communication between students reveals the existence of networks among students. We report a newly constructed simulation behavioral decision model that incorporates these findings.

This study focuses on Dixit, a game of Imperfect-Information Game. In this game, the parent player selects a card from his hand and puts it face down with a hint. The other players put card related to the hint face down from their hands. The other players guess the parent's card from the hint. If all the players can guess correctly on an easy hint, or if none of the players can guess correctly on a difficult hint, the parent player cannot score points. It is important for the parent player to control the relevance of the cards and hints. The analysis of such ambiguous hints is useful for implementing Dixit game AI and understanding human cognitive characteristics. Previous studies have not been able to study the control of relevance. In this study, we annotated and evaluated the relevance scores of cards and hints.

The field of computational social choice brings together principles, techniques, and tools from computer science and social choice theory to create a thriving multidisciplinary field. One of the most well-studied problems in computational social choice focuses on voting rules for selecting the winning candidate in an election. Recent research goes beyond classical voting rules by looking at rules that select multiple winners or drawing on the parallels between machine learning and voting. It is common to encounter voting paradoxes when implementing voting rules in electoral systems. Unfortunately, these paradoxes usually provide little information on the conditions that make them more or less likely to occur. Computer simulations and generative probabilistic models are practical approaches to address this problem. This short paper addresses the problem of evaluating voting rules in competitive computer simulations. Multiagent simulations can provide valuable insights into the performances of competing voting rules defined over parametrically generated problems and populations. The outcomes of this work could improve the designs of electoral systems in the absence of theoretical results to support the optimality of a voting method, and to bridge the gap between axiomatic and experimental analysis of voting systems, leading the way to enhanced explanations and predictions.

According to Toulmin's model of discussion, evidence of a claim that shows the validity of gap between evidence and claim is important. However, not all people make statements while adding appropriate evidence. Therefore, we thought that if a agent could recommend evidence to supplement their arguments during discussions, it would help the discussions to proceed smoothly. The purpose of this research is that the agent provides evidences and helps discussion based on evidence. For this purpose, we propose a web-based evidence recommendation agent. In this paper, we focus on (1) the extraction of claims from opinions and (2) a method for generating search terms and collecting Web articles, and developed a method using GPT-3. To verify the significance of the agent, we conducted a discussion experiment using the agent. In the experiment, the agent failed to recommend appropriate evidences. But, we found that we could use it in ways we hadn't envisioned.

Categorization plays a very important role in human cognition. Categorization efficiently encodes information in myriad feature dimensions. These are automatic, high-speed cognitive processes that enable complex information processing. In previous research on category learning, learning tasks were strictly determined by the experimenter – participants usually do not have been able to seek for a specific set of features during learning. In this study, we used a task in which learners can freely choose what kind of information they want to obtain and examined what kind of information-seeking behaviors learners perform during active category learning.

We propose a task-allocation procedure with preplanning and a resource-re-allocation method by negotiation for path-management of multiple carry robots(agents) in automated warehouses. The task-allocation to agents and the path-management as resource-allocation are key factors to determine the performance of the warehouse. We assume these two factors as separated independent problems, and apply preplanning and negotiation method, respectively. The proposed methods are evaluated by simulation experiments with several problem setting. The results of experiments shows that the proposed method can improve the performance of the warehouse and usage of agents with reasonable computational complexity.

Humans can acquire various manipulation skills in the real world by understanding the structure of the environment and multisensory integration. It is an important step toward the realization of intelligent agents capable of autonomously acquiring diverse skills like humans to learn a manipulation task by model-based reinforcement leaning with a world model from sensor information consisting of multiple modalities. In this paper, we verify by experiments that the learning speed for the Pick and Place task can be improved by attaching a tactile sensor to the end-effector of a robot arm and using it as an input to the world model. We also discuss the need for unified learning environment setup for manipulation tasks.

This study focuses on NewtonianVAE, a world model that can learn a proportionally controllable latent space. To achieve precise control in a physical world, it is necessary to construct a latent space of NewtonianVAE representing a physical world from multi-modal observations. However, learning from multi-modal observations using NewtonianVAE has not been studied. To address this issue, we discuss methods for learning multi-modal observations using NewtonianVAE. In this paper, we propose Multi-modal NewtonianVAE (MNVAE), which uses Mixture-of-Products-of-Experts (MoPoE) to integrate multi-modal observations. MNVAE learns a latent space representing a physical environment and it has the potential for precise control in a physical world.

Multimodal variational autoencoders can acquire a latent representation that integrates information from all modalities by learning an inference model. However, when we want to obtain the shared representation from an arbitrary modality, other modality inputs are missing, which prevents proper inference of the representation. In this study, we reconsider the missing modality problem as part of the amortization gap between amortization inference from any modality and multimodal ELBO, and propose a method to appropriately obtain a shared representation from a single modality input by using iterative amortized inference. However, since multimodal ELBO must be evaluated in the process of iterative amortized inference, missing modality inputs are also required. We, therefore, prepare an inference model that takes only the modality to be inferred as input, distill iterative amortized inference as the teacher and the newly prepared inference model as the student, and verify that an inference model that can acquire a shared representation from a single modality is obtained.

We focus on the task of identifying target objects in domestic environments according to free-form natural language instructions. In this study, we propose a novel transfer learning approach for multimodal language understanding, Prototypical Contrastive Transfer Learning (PCTL) which uses a new contrastive loss, Dual ProtoNCE. Our experiment demonstrated that PCTL outperformed existing methods.

Peripersonal space, where individuals interact with the environment within their reach, has multimodal representations in the brain. It is assumed that the multimodal representation of peripersonal space is acquired through interaction with the environment. In this study, we propose a neural network model that acquires a representation of peripersonal space shared between vision and touch through the experience of vision, touch, and proprioception. Our proposed model reconstructs visual and tactile observations corresponding to proprioceptive inputs after integrating the observations through Transformer based on self-attention mechanism. By learning on camera vision and arm touch of a simulated robot and proprioceptive inputs of camera and arm poses, a spatial representation like a map between the spatial coordinates of peripersonal space and visual and tactile observations was constructed in the model. In particular, the spatial map was shared between vision and touch by sharing part of the visual and tactile decoding module.

This paper describes a domestic service robot (DSR) that fetches everyday objects and carries them to specified destinations according to free-form natural language instructions. We propose Switching Head-Tail Funnel UNITER, which solves the task by predicting the target object and the destination individually using a single model. We conduct physical experiments in which a DSR delivers standardized everyday objects in a standardized domestic environment as requested by instructions with referring expressions. The experimental results show that our method outperforms the baseline method in terms of language comprehension accuracy and the object grasping and placing actions are achieved with success rates of more than 90%.

Research on multi-agent path planning using reinforcement learning methods has recently been developed. However, a common problem in this field is the difficulty of agents learning to cooperate with each other, since each agent is motivated by its own reward. In this study, we examined the impact of considering not only self-reward but also those of others. A world model is introduced to predict the future states of the environment. Considering agents' fairness is expected to be an effective solution to address reward bias among agents and ultimately achieve satisfactory performance in real-world applications, such as operating in crowded environments.

Intelligent robots that can adapt their behavior according to their environment are highly desirable. In particular, camera images are often required to acquire environmental information. Obtaining images from multiple viewpoints can help robots to deal with disturbances and self-occlusions. However, simply combining image features from multiple viewpoints can result in the loss of important information. To address this issue, we propose a deep learning-based approach that uses a gating mechanism to weigh the features of multi-view images based on the current situation. Specifically, we use contrastive learning to align the features and then calculate the weighted average of latent representations using the gating mechanism. We also introduce data augmentation to simulate occlusions and auxiliary costs for action predictive learning. To evaluate our approach, we conducted a real robot experiment, and the results demonstrated the effectiveness of each component of our proposed method.

In this paper, we introduce a reinforcement learning method based on a world model that finds the optimal policy in an environment represented by a graph. There are many environments in virtual and real worlds that are represented by graphs, such as games, transportation networks, knowledge graphs, social networks, and communication networks. Although there are several methods for finding the optimal policy in these environments, existing research has not been able to utilize prior knowledge from similar environments when learning new policies. Therefore, in this study, we propose a method for learning better policies in environments represented by graphs when knowledge of the environment is acquired. We also show that the proposed method outperforms a simple search method without prior knowledge by simulating a maze game represented by a graph.

In this study, we aim to develop a human-like biped robot by integrating the world model with the passive dynamical mechanism that consists of the interaction between the body and the environment. The robot learns jumping movements by watching from third perspective via a camera itself as if it were practicing dance in front of a mirror. Force commands to pneumatic actuator and motor are generated directly from the camera image (end-to-end). In the actual experiment, continuous jumping was obtained using DreamerV2, a world model-based deep reinforcement learning.

In general, robot controls are limited to pre-programmed actions, lacking adaptability to the environment and ease of use for humans. This paper proposes a method that combines a large-scale language model for code generation with a world model for robot control, allowing autonomous robot behavior from natural language sentences. The Language Model Program (LMP) generated by the large-scale language model uses the latent space of the world model learned by NewtonianVAE to enable high-freedom and high-abstraction control. This combination of large-scale language models and world models is important in the context of neuro-symbolic AI, which combines latent and symbolic representations.

With the development of deep learning, significant performance improvements have been achieved in computer vision and natural language processing. In these advancements, scaling laws that demonstrate exponential changes in model performance with respect to model size, dataset size, and computational resources used for training have played a significant role. These scaling laws have been reported to hold for various tasks, including image classification, image generation, and natural language processing. However, it has not yet been verified whether these scaling laws are effective for tasks that involve long-horizon predictions. In this study, we investigate the validity of scaling laws for world models from the perspective of model size. We conduct experiments that scale the model sizes of two world models in a video prediction task conditioned on actions using the CARLA dataset, and verify that the loss function decreases exponentially and the scaling law holds when including large-scale autoencoder.

Visual scene understanding, such as image captioning, can be considered one of the essential topics in the artificial intelligence (AI) field. Image captioning with reading comprehension tasks as an extension of traditional image captioning is more challenging because the generated caption must be related to the text information in the image, and how to read and comprehend text in the context of an image needs to be studied. In this work, we propose multiple image-related attention blocks with multimodal Optical Character Recognition (OCR) information to model the relationship among the global image, multi-level recognized text, and the detected objects in the image. Our model is validated on the standard dataset TextCaps, and the results show that our model outperforms the baseline methods in terms of all evaluation matrices.

Robots are expected to achieve their goals through human-like perception and action. Deep active inference based on the free-energy principle (FEP) is a promising approach. However, most studies have only considered toy problems in simulated environments. To overcome this limitation, we propose another deep active inference framework for real robots. This framework consists of a world model, an action model, and an expected free energy (EFE) model. The world model ensures adequate perception of the environment by minimizing contrastive variational FE, while the action model generates adaptive actions through imitation learning by minimizing contrastive EFE estimated by the EFE model, where each energy is adapted from the original one for contrastive learning. We show that a real robot with the framework can successfully perform a reaching task in both learned and unlearned environments. These results highlight the utility of the FEP with contrastive learning for real-world robot control.

The acquisition of complex human motions in a simulator space is expected to be significant in various scenes such as games and 3DCG animation. One such method is reinforcement learning by using motion-captured human motions as references, but acquiring reference motions is expensive because it requires equipment and actors. However, the cost is high because it requires equipment and actors to acquire reference motions. Against this background, this study investigates the applicability and problems of using the Motion Diffusion Model, a current SOTA model, for generating reference motions.

Soft robots made of flexible materials such as rubber and elastomers are attractive because they can guarantee safety due to their physical softness. However, their flexibility causes difficulty in computing accurate mathematical models, making them difficult to control. In this study, we aimed to obtain a learning-based prediction model of pneumatic artificial muscles, one kind of soft robot, in order to achieve high-precision control of the robots. We created a scaleable data collection device that collects air pressure, muscle length, and load data, and trained a time-series prediction model using 5 hours of collected data. Furthermore, we verified the effectiveness of the method by executing a control task using the learned prediction model.

Robots are expected to perform various tasks in complex environments with a high generalization ability, similar to that of humans. Generally, imitation learning with expert demonstrations has high efficiency but low generalization ability. In contrast, reinforcement learning with explorations has high generalization ability but low efficiency. To combine their strengths, we focus on ``play data'' collected by humans teleoperating a robot with curiosity. Specifically, we propose a framework for real-world robot control and data augmentation based on world model learning from play data. Robot experiments demonstrated that the robot with the framework can perform goal-conditioned object manipulation tasks. Furthermore, we also found that simulation in the world model can create novel combinations that are not included in the original play data. These findings suggest that further learning the augmented data has the potential to enable the robot to acquire higher generalization ability.

In recent aging society, mobile manipulators are expected to replace and support human labor in indoor environments. However, they require human teleoperation in real operation at this point. In this paper, we develop a new teleoperation interface for mobile manipulators replacing previous interfaces. In previous work, teleoperators are forced to operate grasping task in limited or difficult-to-indetify visual feedbacks. In this work, however, a proposed recoginition system assuming world models enables teleoperators to get semantic and spatial information of objects around mobile manipulators. Moreover, we develop another feature which assist grasping objects selected by teleoperators. The experimental results showed that the time of teleoperation was reduced compared to the baseline method, suggesting the possibility of improving the operability of teleoperation.

Collaborative robots are expected to work alongside humans to achieve shared goals. Deep reinforcement learning offers high generalization ability, but its exploration process may create physical risks for human partners. In contrast, deep imitation learning is efficient but may have limited generalization ability since it relies on demonstration data. To address these limitations, this study proposes another deep learning-based framework that utilizes ``play data'' collected through teleoperation of a robot based on an operator's curiosity, during interaction with a human partner. The framework consists of a model that infers a latent representation of achievable goals and a model that generates actions based on the inferred latent goal representation. An additional mechanism optimizes the latent goal representation based on human behavior during interaction through the prediction error minimization mechanism. Experimental results on human-robot collaboration tasks demonstrate the effectiveness of the proposed framework.

Over the past decade, deep learning has made significant strides in improving various domains by training large models with large-scale computational resources. Recent studies showed that large-scale transformer models perform well in diverse generative tasks, including language modeling and image modeling. Efficient training of such large-scale models requires a vast amount of data, and many fields are working on building large-scale datasets. However, despite the development in simulator environments such as CARLA and large-scale datasets such as RoboNet, the scaling to dataset size of the performance of world models, which try to acquire the temporal and spatial structure of environments, has yet to be sufficiently studied. Thus, this work experimentally proves the scaling law of a world model to dataset size. We use VideoGPT and a dataset generated by the CARLA simulator. We also show that the computational budget should mainly be used to scale up dataset size when the number of model parameters is on the order of 10⁷ or larger and the computational budget is limited.

We suggest the use of Optical Character Recognition (OCR) technology to extract data from a designated portion of political funding reports, enabling the creation of a political fund database. Our study underscores the rising popularity of open data in Japan, with the government having developed guidelines to promote its use. However, the lack of structured data and the format of political funding reports present significant hurdles to their analysis. We maintain that a database of political funding reports would enhance transparency and enable greater public access to information. Our research serves as a foundational step towards achieving this objective.

COVID-19 has been raging around the world from the beginning of 2020 to as of January 2023, and it is socially important to build a model to predict and explain its expansion. In this study, we first estimate the number of social contacts within a city for each type of social contact using large-scale GPS data and fit the effective reproduction number to calculate the basic one-to-one effective infection rate. Based on this, mathematical modeling is performed to comprehensively incorporate the effects of delta and omicron variants, vaccination effect and its decay, and herd immunity. The fitting results using real data show that all parameters are consistent with clinical data and other sources and quantitatively clarify the components of the effective reproduction number of COVID-19.

Celebrities and influencers are harassed online on a daily basis. Online harassment mentally disturbs celebrities/influencers and negatively affects society. However, limited studies have been conducted on online harassment victimization of celebrities and influencers, and its effects remain unclear. We surveyed Japanese celebrities and influencers (N=213) about online harassment victimization, emotional damage, and action against offenders and revealed that various forms of online harassment were prevalent. Some victims used the anti-harassment functions provided by weblogs and social media systems (e.g., blocking/muting/reporting offender accounts and closing comment forms), talked about their victimization to close people, and contacted relevant authorities for concrete legal actions (talent agencies, legal consultants, and police). By contrast, some victims felt compelled to accept harassment and did not initiate actions for smaller offenses. We proposed several approaches to support victims, inhibit online harassment, and educate people. Our research would contribute that platforms establish a support system against online harassment.

Many papers have analyzed political discussions between liberals and conservatives on social media. However, only a few portions of users participate in their enthusiastic partisan clusters. The majority of users are in non-partisan clusters with little political interest. It is important to understand their roles and the differences in interest topics from those of the partisan cluster regarding the relative understanding of political discussions. However, few studies have focused on understanding non-partisan clusters, and they are mainly analyses of network structure. This study focuses on each cluster's interest topics, posting tendencies, and tweet reach, especially in political discussions about the discussion on Twitter of the National funeral of former PM Abe. We extracted a non-partisan cluster using clustering and ideological estimation. They had about the same users as the partisan cluster but fewer tweets. We found that while they were interested in events, they were less interested in high-context discussions. In addition, they were more likely to tweet and change their opinions when significant events occurred. They were also centrally located between partisan clusters and exchanged many neutral tweets, suggesting their contribution to stabilizing discussions. This study shows the importance of including non-partisan clusters in political discussions.

The abundance of accessible real-time user information generated from Twitter platform contributes to the utilization in multifarious research domains. Researches detecting Ideologies for Twitter users via Link analysis such as retweet network and follower network shows great results in recent years. However, there are difficulties when taking into account the link relations among countries with different language systems, with the fact that people hardly retweet or follow other people when there is a language barrier. In our research, we fine-tune a multilingual LaBSE model to create U.S. political dimension based on political activists’ user vector embeddings, and project political user vectors from other countries to clarify transnational ideologies. We also build classifier to categorize tweets into various topics to check if the change of topic alters the political ideologies of the same user. We can conclude that user political ideologies show different polarization degrees when it comes to different topics.

Online advertising has grown remarkably, with the display advertising market in Japan accounting for one-third of the total online advertising market. Display advertising consists of images and text, and advertisers maximize sales revenue by contacting consumers through advertisements and encouraging them to make purchases. In today's society, where products are becoming more homogenized and needs are diversifying, appealing to consumer psychology through advertisements is becoming increasingly important. However, it is not sufficiently clear what kind of appeal influences consumer psychology. In this study, we quantified the appeal of texts of advertisements for health products and cosmetics, which were actually delivered in Instagram advertisements, one of display advertisements, by applying Linguistic Inquiry and Word Count (LIWC). The correlation between CTR and the text was analyzed. The results showed that negative appeals that arouse consumer anxiety and a sense of crisis were related to CTR.

The aim of this study is to demonstrate the usefulness and applicability of the method using the word embedding model in vector space, which can realize the method examined in structuralist sociology (Bourdieu, etc.) at higher dimensions. The latter method refers to relational analysis that emphasizes the relationship of relative positions (distance) among actors within a social space. In this study, we collected Twitter data on "parental leave," created high-dimensional vector representation data, mapped it to a three-dimensional coordinate space, and conducted clustering to visualize the various practices of the actors in a certain degree from multiple perspectives, which are difficult to see from the public space.

Meat alternatives, which are the representative products of food-tech, are expected to provide as a solution to the negative impacts of increasing meat consumption. Given their reasonable impact on the environment, human health, and animal welfare, consumers could be encouraged to purchase them by appealing to the moral aspects of alternative products. Here we conducted an online survey of 229 US residents to determine whether morality influences consumers’ willingness to purchase meat alternatives. The results showed that morality was a significant explanatory variable in a multiple regression model for predicting willingness to purchase meat alternatives. However, the effect was limited by product and was only significant for plant-based meats. Our results suggested that the willingness to purchase such products is associated with liberal ideology. These findings suggest that appeal to morality may be an effective way to promote meat alternatives, although the topic may be controversial from an ideological standpoint.

We construct a pre-learning model for individual daily trajectories by inputting travel time and travel location into GPT-2, an autoregressive language model, utilizing the location history of approximately 680,000 smartphones that traversed Urayasu city in August 2022. Additionally, we incorporate environmental factors, such as weather conditions and daily new coronavirus cases, as well as attribute information of the smartphone owners. During the learning process in the model, numerical information is transformed into unique character combinations. By this transformation, we can obtain highly accurate individual daily trajectory models without the need for geographic information.

Modeling route choices of travelers to reach their destinations has various practical applications. However, conventional route choice models use a softmax function to calculate the probability of each route based on its cost, which results in positive flow even for extremely long routes. In this research, we developed a route choice model that can account for zero flow by utilizing the sparse output property of the activation function. The sparse activation function was derived through regularization with generalized entropy, a generalization of Shannon entropy. Our results show that our model can exhibit a range of models, from a softmax-type model to one that captures zero flow. Additionally, we show that parameter estimation can be performed through linear regression. This reduces the computational cost on previous route choice models that could represent zero flow, which required a combination of maximum likelihood method and shortest path search during estimation.

The recent spread of IoT technology has accumulated a so large amount of information, that we cannot process them enough. Data mining methods are required to extract useful information from such information. The objective of this study is to construct an algorithm that can efficiently extract highly frequent or highly localised patterns from a single sequential data. Specifically, we propose a method for extracting these patterns based on the swarm intelligence mechanism. We applied the proposed method to pseudo-created series data, and confirmed that the proposed method extracted patterns depending on their frequency and localisation.

This study presents a method for predicting the GPS signal strength received by devices at each position inside a target floor in a building. The predicted signal strengths can be used as a rough indoor fingerprint localization system without extra infrastructure. Although it has been widely considered as a fact that the current GPS system is still unable to achieve high-precision indoor localization, we attempted to analyze the characteristics of indoor GPS reception by considerable factors. We employ a neural network based system which mainly uses floor plan including wall’s information and window information as well as shapes and heights of neighboring tall buildings. In addition, we integrate the GPS satellite information including its azimuthal angle and elevation angle to estimate line of sight (LOS) from each satellite to the target environment. We evaluate our framework using data obtained in different buildings from different areas in the city and campus.