In recent years, marketing using attribute information associated with member accounts of online services has been widely used. However, the majority of users are non-members who use services without registering for an account, and it is difficult to implement measures using attribute information for these non-member users. In order to deal with this situation, semi-supervised learning is an effective way to increase the number of users with attribute information by predicting it from the history data of member users who have attribute information, using the history data of non-member users as well. One of such semi-supervised learning methods is the Ladder Network, which is a neural network based model with adding and removing noise. This model provides highly accurate prediction for image data, and is also considered to be useful for predicting user attributes from historical data, where the feature vector is high-dimensional. However, this method cannot be applied to the case where the label takes ordered value, such as the user's age category. In this study, we propose an extended model based on the Ladder Network that incorporates a mechanism that can appropriately predict the user's attribute information. We also conduct an evaluation experiment using actual browsing history data to show the effectiveness of the proposed method.

In recommendation systems, users and objects are vectorized based on data describing the relationship between the object and the user, such as the users' five-point rating of the object and similarity between objects, and the user's preference of an unknown object is predicted according to the distance between the user and the object. There are many types of similarities and previous research improves performance by converting them into features and mixed by means of DNNs. However, it is difficult to analyze which similarities are important for the individual user because of the complex mechanism of DNNs. In this study, we propose a model that predicts the mixing ratio of similarities for each user and calculates the vector of objects so that the mixing ratio of the vector directly corresponds to that of similarity. We also show that interpretable mixing improves precision experimentally.

Collaborative Metric Learning (CML) is a recommendation model based on implicit data, i.e. behavior history such as clicks and browsings. CML learns an metric space to embed not only the relationship between users and items, but also the similarity between items and that between users. Moreover, CML recommends the items which are close to each user in the trained embedding space. However, CML tends to learn by focusing on items that are popular among many users, and the accuracy of embedded representations of other minor items is often neglected. On the other hand, it is necessary to learn embedded representations of many minor items that match the user's preferences with high accuracy in order to provide unexpected recommendations that users may not have recognized. In this study, we propose a method to learn the embedded representations that capture user's preferences by weighting according to the number of observations of implicit data, and to make unexpected recommendations that include minor items. Finally, we apply the proposed method to actual movie evaluation data set, and show the usefulness of the proposed method in making unexpected recommendations based on the users' preferences.

The competition for customers among video distribution services is intensifying. In general, the user's purchasing actions for video contents (items), unlike daily necessities, have a strong influence on the their real-time interests during viewing items (consumption). In other words, the user's interest after the consumption of an item is determined by the influence of the previous item for the user's interest persistence (interest persistence probability under the item). Therefore, it is important to select and evaluate items based on the interest persistence probability under the item in order to have users use the service for a long time is important. Hidden Semi-Markov Models (HSMM) was proposed as a model for predicting the next item to be consumed by a user while taking into account the user's interest persistence. If the interest persistence probability under an item can be calculated and analyzed using HSMM, the new insights leading to the marketing strategies can be expected. In this study, we propose an analysis process using item clustering based on the distribution of the interest persistence probabilities under the items, utilizing the characteristics of HSMM. In addition, we show the effectiveness of our proposed method by applying the actual data set.

The purpose of this study is to create a highly accurate emotion classification model for voice by analyzing the entire voice waveform and extracting the features of voice data for each emotion. In this paper, we created an emotion classification model using neural network for the purpose of more accurate emotion classification of voice data. We then evaluated the model and compared its accuracy with that of a model created using decision trees.

Predicting future stock price fluctuations is very important for investors to avoid losses in their investments. The VI index is one of the indicators of the overall risk of the stock market; the higher the VI index, the more likely it is that prices will fluctuate, and if we can predict an increase in the VI index, we may be able to predict future price fluctuations in the stock market. In addition, social events and the opinions of others influence people's investment behavior, and these are expected to be related to the VI index. Therefore, this study proposes a model that predicts the increase in the VI index using the content of social media posts by stock message boards and the content of newspaper media articles.

We propose a method to optimize the investment ratio and learn behaviors considering the risk in Compound Deep Reinforcement Learning. Compound reinforcement learning is reinforcement learning that aims to learn behaviors that maximize the compound returns with the betting fraction parameter. We can maximize the compound return by optimizing the betting fraction. Previous work on compound deep reinforcement learning uses the given betting fractions in the range of zero to one, and it does not consider the investment risk. We propose a method to optimize the betting fraction by adding a network to compound deep Q-network. We also propose a method to learn behavior that reduces the variance of returns.

Many supervised deep neural networks have been studied to predict stock prices. Among these studies, recent methods used the attention mechanism to extract relevant time-series data and improve prediction accuracy. Although the advantage of the attention mechanism has been demonstrated based on various experimental results, relationship between these studies and concept drift studies have not been discussed. This paper discusses this relationship. Coexistence and transition of multiple concepts and exhaustive analysis of concept drift are discussed in this paper.

Predicting market impact is important for all traders to formulate, evaluate and improve their trading strategies. In the past, there have been studies that approximated market impact from historical trading data. There have also been studies that calculate market impact indirectly by modeling limit order books(LOB). However, these models are extremely generalized and have difficulties in terms of practicality. On the other hand, in recent years, deep learning has been used to analyze LOB, which is an important factor in considering the market impact. However, the main purpose of these studies is mainly stock price prediction or optimization of the overall execution cost of trades. There is little research on the evaluation of market impact alone. In this paper, we propose a new model for evaluating market impact using deep learning. First, we predict stock prices based on the LOB information, and then evaluated the market impact using the prediction. Through experiments, we have found that the market impact calculated by the new model was consistent with that of existing studies, and LOB analysis by deep learning captured significant information in market impact calculation.

The objective of this study is to develop a highly accurate stock price prediction using news information. The authors have previously reported the construction of a sentiment classifier for news using machine learning to improve the accuracy of news sentiment analysis.In this paper, we propose a stock price prediction model that takes into account the impact of news by classifying news into positive, negative and neutral using the classifier constructed using BERT and incorporating the positive and negative classification results into the stock price data.The effectiveness of the proposed method is compared and verified by simulations, aiming at highly accurate stock price prediction.

Deep sequential generative models have been used in various tasks such as time-series prediction, unseen sequence generation, and time-series anomaly detection. In this report, we focus on models so-called sequential variational auto-encoders and propose an efficient learning framework by sequential Bayes filtering. Although similar prior works provide tighter ELBOs which are lower bounds of the log marginal likelihood, several problems such as the low spread of particles in latent space remain. The proposed framework overcomes these problems by emphasizing practical use and outperforms the prior works for several datasets in predictive ability.

In recent years, second-order optimization with a fast convergence rate has been used in deep learning owing to fast approximation methods for natural gradient methods. Second-order optimization requires the inverse computation of the information matrix, which generally degenerates in the deep learning problem. Therefore, as a heuristic, a damping method adds a unit matrix multiplied by a constant. This study proposed a method for scheduling damping motivated by the Levenberg-Marquardt method for determining damping and investigated its effectiveness.

One of the recent issues in AI is the black box inside the inference results of machine learning. As an approach to solving this problem, the fusion of fuzzy inference and reinforcement learning, which is based on rules that follow human subjectivity, is an effective method. Igarashi et al. proposed a policy gradient method that uses fuzzy control rules as policies. In their framework, we approximated the membership function with a sigmoid function and learn the parameters in the sigmoid function and rule weights in the speed control problems of a car. As a result of the learning experiment, it was confirmed that appropriate parameter values were obtained. However, even in this case, the approximate form of the membership function was designed by a human. Therefore, we attempted to approximate the membership function with a neural network to see if we can learn the shape of the membership function from scratch. As a result of the learning experiment, we obtained a function shape that closely resembled the shape of the human-designed membership function from the initial values of random parameters. This suggests that the proposed learning method can acquire human fuzzy concepts from scratch.

In this work, we are interested in studying the potential of learning algorithm design that is driven by novel, diverse notions of "risk" that go well beyond the traditional choice of expected loss. In particular, we look the impact that introducing a generalized, scalable, bidirectional dispersion term has on how risk is measured, and the repercussions it has in the dynamic setting of stochastic optimization.

In recent years, machine learning has made it possible to utilize various types of data. On the other hand, as the opportunities to use data increase, data breach from machine learning models has been pointed out. For example, consider a model that provides input candidates while writing an e-mail. If this model outputs a credit card number that exists in the training data, this is a data breach. In this research, as a countermeasure against data breach, we tackled the task of text generation by RNN using DP-Adam, an optimization algorithm that satisfies differential privacy. We tested whether it is possible to prevent the exposure of dummy data, which is assumed to be personal information. As a result, we confirmed that the DP-Adam model and the L1-regularized model avoided the dummy data exposure. However, the text generated by the L1-regularized model contained words that did not exist.

Item stock allocation optimization is a problem for e-commerce companies that operate multiple warehouses to improve the efficiency of item allocation among the warehouses. This problem is very important for reducing shipping and inventory costs. Since the objective function in this problem has some parts that are difficult to represent by formula and to grasp the features, evolutionary computation, one of black box optimization methods, is suitable as the solution. We apply constrained NSGA-II, a multi-objective evolutionary computation method, for minimizing both shipping and inventory costs simultaneously, while considering various constraints such as the capacity of warehouses. In usual evolutionary computation, since the crossover is operated randomly, the charactoristic structure in a good solution may be destructed. Thus we propose a method by which crossover is operated by each item or warehouse, for the purpose to sustain the structure of solutions. By comparing the evolutionary transition of the proposed method with uniform crossover, we confirm the effectiveness.

In hybrid cloud data utilization, data movement across the countries will occur, and there is a risk of unintentionally violating the laws and regulations of export control. The purpose of this study is to establish a technology to reduce the risk of violation of laws and regulations in data movement, by associating technical documents and legal statements of export control. In order to respond to the differences in the vocabulary of Japanese business documents and legal statements, the proposed method calculates the similarity of the sentences through their English translations. Our prototype associated 73% of technical documents with US EAR items. In addition, by cooperating with the access management of the data, the prospect of reducing the risk of illegal export of technical information is obtained.

Academic-corporate collaboration is evolving. In the era of digital convergence, especially in recent years when AI has widely spread and become prevalent in various fields, academic-corporate collaboration has pursued efficiency and effect, and we can observe a certain number of cases where researchers themselves bridge academia and industry with two hats of academia and corporation. In this research, we focus on researchers who conduct academic-corporate collaboration alone (we call "ambidextrous researchers") and study the impact of academic-corporate co-authored papers. Through the study, we used Scopus as a data set of papers and extracted highly cited papers for analysis.As a result, it is confirmed that the number of citations of academic-corporate co-authored papers, especially those including "ambidextrous researchers" as authors, is significantly higher in the top highly cited papers. In this paper, we propose the method for analysis and present the impact of academic-corporate co-authored papers quantitatively.

AI and robotics technologies should prevent people from being socially isolated, especially in societies having a large single population. Here, we discuss a use of these technologies as a mediator between humans. The goal is to maintain and facilitate human communication with the help of such an AI agent. As an initial attempt, we target elderly users who live separately from their family members. Then, we introduce a social mediator robot that handles the exchange of messages between the user and his/her family members. Most importantly, the robot is designed to act to facilitate communication between both parties. In this paper, we will explain our research projects studying how the robot should act in this facilitation process. Both verbal and nonverbal intervention approaches are discussed.

In recent years, research in China has experienced an explosion in the number of researchers and papers and an increase in the percentage of highly cited papers. As a background, "returnee researchers" who have returned to their home countries after researching abroad due to China's sea turtle policy continue to achieve high research performance after returning to China. However, the number of returnee researchers is too small to explain the burst of research in China. This study measured the secondary impact of returning scientists in China on domestic Chinese researchers who co-authored with returnees. We found that co-authorship with returnee researchers improves domestic researchers' quantity and quality. Moreover, returnees make domestic researchers more multidisciplinary with co-authorship. The analysis of the secondary effects of returnee researchers in this study provides valuable insights into the overall picture of research within China.

Currently, "quality of education" is a key issue in the field of education. To improve this, it is effective to reflect on lessons. However, in actual educational settings, it is difficult for teachers to record every child in class. To address this issue, we believe that the amount of events occurring to each child can be estimated from the amount of behavior, and we are working to extract the amount of behavior using AI technology from the data measured by our class situation sensing system. Based on this, we have examined the possibility of estimating class participation attitudes using behavior quantities by comparing the extracted behavior quantities of individual children with manually assessed class participation attitudes. In this presentation, we report on the current state of the art regarding the estimation of class participation and discuss how it can be used in teacher reflection situations.

Today, many devices such as smartphones and cars are equipped with cameras thanks to the development of various types of hardware and software. In particular, in-vehicle cameras are expected to be applied not only to vehicle operations such as automated driving technology, which has been rapidly developing in recent years, but also to recognition of the outside world, such as detection of changes in objects on the road. Generally, deep learning methods are used for such processing involving recognition from video images, but this requires a large amount of teacher data, which is currently selected manually. On the other hand, manually assigning teacher labels to video data takes several times longer than the actual video time. Therefore, the cost of labeling is not only a problem, but also the mislabeling caused by human identification errors. Therefore, we propose an efficient annotation workflow to reduce the man-hours required to create teacher data for in-vehicle camera videos. Based on a model that captures the characteristics of in-vehicle camera videos, we show that this workflow can be used to perform teacher labeling (annotation) with a minimum number of man-hours. By capturing the characteristics of the in-vehicle camera video and reflecting them in the identification model in the workflow, the proposed workflow resulted in a reduction of about 77.5% of man-hours.

Ethnicity is an essential factor in predicting the flow of international trade. However, it is challenging to collect ethnic data, so there was a limit to analyzing international trade with ethnicity. In this study, we classify the ethnicity of corporate executives by surnames. And we analyze transactions between ethnic groups in international trade. In addition, we measure the influence of ethnicity using the gravity model. We find that the ethnic factor significantly affects international trade, and the impact varies from country to country.

This study aims to improve the accuracy of the proposed method to visualize cracks in concrete using pix2pix from radar images by adding quasi-3D information. In a previous study, ground-penetrating radar tests were conducted on concrete specimens in which artificial defects were embedded in varying positions, angles, and sizes, and cross-sectional images including geometric information of the defects and corresponding radar image pairs were obtained. The dataset was applied to pix2pix to construct a model that outputs the cross-sectional image from the radar image. In this study, we further attempt to apply a method to output the cross-section image by applying two radar images, in which the radar scanning positions are shifted by about the maximum size of coarse aggregate in concrete. The proposed method can visualize defects with high accuracy, even in cases where the reflection intensity of electromagnetic waves becomes small, and the accuracy of conventional methods decreases.

In recent years, with the growing need for a safe, secure, and comfortable environment, abnormal detection plays an important role to prevent terrorist attacks, incidents, and accidents, for example, detecting falling objects on the road or suspicious objects in facilities such as train stations. In previous work, the background subtraction method has been used to detect such objects. However, it has the problem of false detection of swaying grass and trees, changes in sunlight, etc. In this study, a new abnormal detection method is proposed that combines VAE (Variational Auto-Encoder) and NNS(Nearest Neighbor Search) using frame subtraction images to detect falling and suspicious objects from surveillance cameras. Experiment results show that for data 1 (Ayabe), the G-mean value was 0.975 by our proposed method, compared with 0.876 by the previously reported VAE and 0.663 by the background subtraction method using OpenCV. Furthermore, an incremental learning framework is constructed by feeding back the user’s classification result to reduce false detection. Experiment result on data 2 (Yasu) shows that the G-mean Value was improved by 0.072 with our method.

Designers for condominiums refer existing floor plans to create new floor plans that current trends and potential customer needs are taken into account. In this study, we applied semantic segmentation to the floor plans to extract the attributes of rooms and their boundaries, in order to create a database of the floor plans useful to the design work. We investigated the effects of pre-processing such as binarization and enlargement, and the differences between DeepLabv3 and HRNetOCR on the accuracy to enhance the segmentation. As a result, we confirmed that the practical accuracy can be obtained by considering both overall and local features in the inference of floor plans.

A technique for detecting cracks in river concrete revetments using image segmentation technology is being studied. However, the area of the crack to be classified is very small compared to the background. Models trained on such imbalanced data are known to be unstable in prediction accuracy due to optimization that is more influenced by classes with more data than classes with fewer data. Therefore, in order to improve the crack detection accuracy, this paper tries to adopt Focal Tversky Loss, which is a loss function robust to imbalanced data. The model using Focal Tversky Loss showed higher crack detection accuracy than the commonly used Binary Cross Entropy and Dice Loss. In addition, by introducing an attention mechanism into the segmentation model, a visualized image representing a judgment reason as to which part of the image was focused for segmentation was generated.

It is difficult for users to find paintings that match their tastes among the huge number of paintings available in museums and on the Internet. A painting recommendation system is a useful solution to this problem. However, it is difficult to construct a recommendation system based on user history, which is used for movies and music, because the unit price and the market size of paintings themselves are different. In order to construct a recommendation system that uses only the image features of the painting itself, this study explores the extraction of image features by machine learning. As features, we used features obtained from general-purpose image classification models and object detection models. The relationship between these features and the user's subjective preference for paintings was investigated by subject experiments.

Anomaly detection by generative models is achieved by comparing the reconstruction and the original image. However, existing generative models often lead to a blurred reconstruction and the loss of original image features (e.g., the orientation). They are practically problematic in industrial anomaly detection, such as the detail flows being overlooked and the need to align the orientation of target objects. Therefore, the generative models have only achieved inferior anomaly detection performance compared to batch-based models and the models for latent features. This paper proposes the reconstruction without diffusion by a diffusion model. This method reconstructs an image well while preserving original features and outperforms existing methods in the industrial dataset MVTeC AD.

Monitoring of respiratory status, which is an important vital sign, is essential for perioperative management. It has become increasingly important in recent years due to the spread of COVID-19 infection. To automate the respiratory monitoring, we have measured and analyzed the movement of thorax as a respiratory signal using a displacement sensor. In this study, we used Convolutional LSTM to discriminate normal and abnormal respiration based on temporal changes in frequency components obtained by complex wavelet transform. To improve the accuracy, we focused on signal preprocessing and network structure. We verified the usefulness of a network structure that combines quantization to reduce the number of patterns to be learned, blocking to determine the appropriate length of data for each learning iteration, and templating patterns by convolutional layers. In our experiment, the proposed method achieved high precision and recall of (99.8%, 99.6%) for normal respiration and (97.7%, 99.1%) for abnormal respiration.

This study targets the prediction of anomalies in navigation and radio equipment onboard merchant ships. These devices are used for voice and other communications with business management centers on land via satellite. However, as these devices deteriorate over time or become overloaded, they can cause abnormalities that prevent voice communication. The objective of this research was to be able to automatically identify the signs of such abnormalities based on the characteristics of the data, and to prompt the replacement of the equipment. In this experiment, the MT method and OCSVM were used with voltage values and CPU temperature as features. Results of the experiment, the MT method was not able to find all the points that the experts wanted to identify as predictive signs of anomalies, but the OCSVM was able to detect all the points.

In recent years, more and more machine learning algorithms have been developed to detect anomalies in time series data. One of the approaches is the generative adversarial networks. Among these approaches, this study uses a method that learns from normal data and produces a high anomaly score when anomalous data is input. This method has advantages such as unsupervised learning and the ability to capture high-dimensional data distribution compared to conventional anomaly detection methods. On the other hand, the Attention mechanism has been widely used mainly in NLP. By using this attention mechanism, it is possible to capture the characteristics of the entire time series more directly than RNN. Although there is a possibility that the accuracy can be improved by applying this mechanism to the time series task, there are few studies that use the Attention mechanism in the research of anomaly detection by generative adversarial networks. In this study, we propose a new approach using adversarial networks with the Attention mechanism and show that our method improves the performance of time series anomaly detection compared with conventional methods.

Cyber attacks against industrial control systems (ICS) are increasing, and in particular, physical damage caused by the tampering with control commands and sensor data contained in control traffic is a social threat. Since the normal range of data contained in control commands varies by environment, anomaly detection using a self-coder such as Auto Encoder has attracted attention. However, the communication data in an ICS environment contains a huge amount of control packet, and the monitoring target includes the parameters of control commands included in the payload, making it difficult to detect anomalous data in the training data. It is not realistic for system operators to check each packet in the training data one by one to eliminate anomalous data. In this paper, we propose a method to efficiently eliminate anomalous communication data in the training data by semi-supervised learning by feature vectorization of control communication packets using BERT, and confirmed its effectiveness through experiments.

This paper reports on the construction of a value structure model and an AI application system that utilizes the model for the purpose of value co-creation that seamlessly connects the online and real worlds without distinguishing between them in a large-scale visitor attraction service. In this paper, we collect data on the changes in visitors' awareness and their migration and measurements that occur in real and online events, and furthermore, we conduct probability modeling that includes the actions of event staff. This allows us to model the value structure that is commonly recognized as a service system, and by using this model, we can make the value of the experience at the real event and the value of the service provided online common for each visitor. It will also be possible to provide additional experiences and post-event information online for each visitor after their visit. This will promote the creation of value together with visitors, on-site staff, and event planners involved in the event.

As many companies are promoting digital transformation (DX), the active use of AI is being considered. However, the gap between the expected accuracy and the introduction of AI into corporate processes, which has improved remarkably through kaizen activities, is a major barrier. To solve these issues, this study proposes an approach that applies Human in the Loop Machine Learning, defines a system architecture that incorporates "Collaboration between human and AI " into the operational phase at the project conception stage, designs and implements it, and spirals up the system during the operational phase.

A shift schedule is essential for managing the work of each employee. However, it is difficult to manually create a shift schedule while taking into account various constraints such as the days employees are available to work and the congestion in the store. In Individualized teaching addressed in our study also, it is necessary to take into account many constraints that include not only available workdays of instructors and subjects they can teach but also days and subjects students can take. In other words, scheduling for Individualized teaching is necessary to find both of a shift schedule for instructors and a subject timetable for students, being useful to develop a system to automatically create a proper schedule satisfying their constraints. In this paper, we propose a two phase optimization method that creates a shift schedule using genetic algorithms and a timetable for students using simulated annealing.

It is difficult to grasp the purchasing behavior, purchase consciousness and decision making of customers in previous studies on customer relationship management. In this research, the RF-PACV method has been proposed by increasing the number of purchases (R), cumulative purchase number (F), number of clicks (P), and number of items (A), item type (C), and number of favorite (V) based on the RFM model and the features of net business data. The Calinski-Hadabasz (CH) criterion is lead into the K-means clustering analysis method in this paper. As a result, the improved K-means clustering algorithmthe and the RF-PACV model are able to obtain five customer data, and the RF-PACV analysis model can analyze the characteristics of each customer group,and distinguish the different consumer habits and preferences of customers.

In this research, we propose a novel method for generating time-series data using a neural network. Time-series data, especially data in real financial markets such as stock prices, is often sampled irregularly, and its noise structure is more complex than the standard Brownian motion. In order to generate time-series data with such characteristics, we extend and generalize the Neural Stochastic Differential Equation (SDE) model based on Brownian motion and propose Neural Fractional SDE-Net based on fractional Brownian motion. More specifically, we propose here the Neural Fractional SDE-Net (fSDE-Net) based on fractional Brownian motion whose the Hurst index is larger than half, which shows the long-term memory characteristics. We theoretically establish a numerical analysis method for fSDE-Net and show the existence and uniqueness of the solution for fSDE-Net. Furthermore, our experiments with various time series dataset demonstrate that the fSDE-Net model can replicate the long-term memory property well.

Data science techniques are indispensable in seismology, which bases large-scale numerical simulations and big data analyses. A new national project "Seismology TowArd Research innovation with data of Earthquake" (STAR-E) launched in 2021 by Ministry of Education, Culture, Sports, Science and Technology, Japan, is accelerating the integration of information science and seismology. In this paper, we introduce two examples of data science techniques applied in seismology, (1) data assimilation to integrate numerical simulations and observational data based on Bayesian statistics, and (2) deep learning to detect evidence of ordinary earthquakes and slow earthquakes from seismic observational data. We also discuss the importance of integrating data assimilation and artificial intelligence in future to reduce the computational cost required in numerical simulations of seismological issues such as seismic wave propagation.

In this paper, we estimated the evacuation behavior, computed by Multi-Agent Simulation (MAS), using Graph Neural Network (GNN). The results computed by MAS are useful for determining guidance strategies and building design in the event of a disaster. However, the computational time strongly depends on the number of agents and the complexity of the agents model. Therefore, We perform simulations by replacing the computation of MAS with machine learning models. Specifically, we construct a GNN that represents the building structure as a graph and estimate the subsequent agent positions using the agent positions until a certain point in the MAS as input. We evaluated GNN with evacuation behavior data computed by MAS and found that GNN express the unique properties of evacuation behavior.

GNNs are the neural networks for the representation learning of graph-structured data, most of which are constructed by stacking graph convolutional layers. As stacking n-layers of ones is equivalent to propagating n-hop of neighbor nodes’ information, GNNs require enough large number of layers to learn large graphs. However, it tends to degrade the model performance due to the problem called over-smoothing. In this paper, by presenting a novel GNN model, based on stacking feedforward neural networks with gating structures using GCNs, I tried to solve the over-smoothing problem and thereby overcome the difficulty of GNNs learning large graphs. The experimental results showed that the proposed method monotonically improved the prediction accuracy up to 20 layers without over-smoothing, whereas the conventional method caused it at 4 to 8 layers. In two experiments on large graphs, the PPI dataset, a benchmark for inductive node classification, and the application to the surrogate model for finite element methods, the proposed method achieved the highest accuracy of the existing methods compared, especially with a state-of-the-art accuracy of 99.71% on the PPI dataset.

The automation and optimization of planning tasks, such as resource allocation and operational planning of various systems such as transportation systems and production facilities have been addressed mainly in operations research and each specific field. Conventionally, planning problems i.e. scheduling problems are reduced to combinatorial optimization problems and addressed using their solvers. In such cases, scheduling complex systems for a long period might incur combinatorial explosions and would be difficult to obtain the solution. Several scheduling problems can also be regarded as optimal control problems. Optimal control problems include several problems concerning sequential decision making such as board games. Reinforcement learning is a method to address them, and its recent advancement is significant. If the complex scheduling problems are reduced to optimal control problems i.e. deciding resource allocation at each time step, not as a whole, recent powerful reinforcement learning can be leveraged to obtain solutions in a short period after the training. In this paper, we introduce these perspectives and their practical applications including railway scheduling and chemical plant operation.

In general, the design space of materials research is quite huge. The goal of materials informatics (MI) is to discover new materials or design parameters that exhibit innovative properties from such a vast search space. The basic workflow of MI consists of forward and inverse problems. The objective of the forward problem is to obtain a statistical model that forwardly predicts physicochemical properties of input materials. The inverse problem, on the other hand, predicts candidate materials with a given set of desired properties by finding the inverse map of the forward model. Here, we focus on the integration of machine learning and simulation in materials science. The biggest barrier to the implementation of data-driven materials science is the lack of a sufficient amount of data. In addition, the goal of materials research is to discover innovative materials that exist in unexplored areas where less or no data exists. Therefore, interpolative predictions based only on conventional data-driven approaches are generally insufficient to achieve this goal. This talk will present some case studies of materials exploration based on transfer learning and adaptive design of computer experiments on a high-dimensional design space, which will be the key to bridging the computer experiments and the machine learning workflow in materials science.

In the uncertain and unstable era of recent years, it is crucial to building a digital twin using queue simulation for factories, logistics, and supply chains in the manufacturing and logistics industry and perform optimal operations. However, it has been difficult to optimize the parameters of discrete queuing simulation to reproduce the data as an actual situation. Therefore, we solved this problem by using a data assimilation method that extends kernel ABC. While the parameters of the simulation are usually determined as deterministic parameters, this method can determine stochastic parameters. This method can be interpretable as determining the parameters of the deductive simulation model by an inductive machine learning method. We will introduce the mathematical background of this method and examples of applying it to actual manufacturing and logistics companies.

Efforts to realize a carbon-free society are becoming the minimum requirement for business continuity. With regard to ESG investment, which is an investment that evaluates corporate environmental efforts that are attracting worldwide attention these days. Research on evaluation methods for benchmarks that evaluate opportunities for new profit generation is becoming active. With this activation of ESG investment, information disclosure regarding decarbonization has led to the trust of the market and has begun to be directly linked to corporate value.Therefore, in this study, we visualized the relative progress of each decarbonized topic in each country. Specifically, articles related to decarbonization were extracted from about 40 million articles over the past three years, BERT embedded vectors were clustered for the articles, and characteristic words for each cluster were extracted using c-TF-IDF. From this study, it was found that the efforts for decarbonization are in contrast in Europe, the United States, China and Japan.

In this paper, we investigated relationships between the diversity of news articles and various user activity by analyzing user logs of a popular news app. Following previous study, we quantitatively define the diversity of consumed news articles for each user by using vector representation for each article. We found that users with high diversity are more likely to continue using the app and consume a higher number of articles and visit the app more often. We also found that the use of various logics with different degree of personalization which are responsible for placing articles in the app play a role in determining diversity for each user. We conclude that the diversity of consumed news articles becomes higher when the users consume news articles demonstrated by multiple logics in a well-balanced way. This study implies that encouraging users to consume diverse contents may result in increase of active users and user activity in the long run for a web service.

Global supply chains and global ownership networks are complex structures. Raw materials are transformed into finished products through many companies in various countries and industries. Therefore, on these pathways it is not easy to manually find companies that are complicit in human rights violations, environmental destruction, or have geopolitical risks. In this study, we estimate the battery procurement pathway of Tesla, Inc. as a case study and identify the national economic security risks latent in the estimated procurement pathway as (1) traditional geopolitical risk, (2) networked geopolitical risk, and (3) hidden dominance geopolitical risks, which are classified into three categories. Each risk can be automatically identified from network analysis (Betweenness centrality, Network power index, and Network power flow) on publicly available inter-firm transaction and shareholder information.

In “The Image of the City” by Kevin Lynch describes how individuals perceive and recall features in urban spaces and represent a city by sharing their opinions. However, due to constraints such as time and space, the traditional forms of public participation to describe and shape citizens' mental representation of the city are untenable. Therefore, Lynch’s approach has stimulated research into digital participatory technologies, mainly artificial intelligence to study the perception and analysis of urban dynamics. In this paper, we propose to study this phenomenon using a digital participatory platform with a combination of “AI-mediated” and “non-mediated” support involving a large number of citizens using open call in five major cities of Afghanistan. The aims were to (1) examine whether AI can elicit city image through extracting and visualizing the insights of citizens in real-time, and (2) evaluate how AI-mediated discussion can elicit innovative elements of city image in comparison to non-mediated discussion. The results show that AI-mediated discussion improved the responsiveness of the participants and increased the number of identified ideas compared to non-mediated discussion.

While the Internet has made it possible for people all over the world to be connected, we have faced the wall of the diversity of values. Although we often have to consider or respect other people's values, it is not easy to sense them due to the limitations of our knowledge, experience, and imagination. We have worked on how to mutually consider unfamiliar values and obtain a synergetic effect among people by developing a discussion support system called AIR-VAS. AIR-VAS is a system aiming to support being aware of other people's values in group discussions. AIR-VAS functions to recognize characteristic opinions of a group and share them among all people engaging in the discussion. Through the sharing of opinions, people can obtain different viewpoints on the issue of the current discussion so that AIR-VAS can stimulate people to generate/evaluate/analyze ideas. On the other hand, AIR-VAS extracts information about other groups based on the frequency of word and co-occurrence information, and we found a problem in the quality of the stimulation that system presented.In this paper, we propose a method that shares information based on distributed representation of language model. We evaluated the method by discussion data, and the result shows the effectiveness of the method.

Automatic content generation based on natural language processing is an active research area, especially for story generation. Research on story generation has focused on generating consistent text pertaining to characters' actions and events; however, there have been few studies on generating characters' lines (utterances) and dialogues. Story plots are not created to stand alone, but are instead used as the basis for the next step in the creation process, such as creating the scripts of movies or plays, the storyboards for comics, and the main body of novels. In this study, we construct a dataset for supporting content creation based on natural language processing. We have annotated about 8,000 four-panel mangas with panel information such as lines, onomatopoeia, character situations, and locations in each panel, as well as a summary of the entire four panels. This paper describes the details of the annotations and the results of the annotations. We also report the results of an experiment using this dataset to generate character lines from summaries.