[in Japanese]

View full abstract

[in Japanese]

View full abstract

[in Japanese]

View full abstract

Having emotions is essential for robots to understand and sympathize with the feelings of people. In addition, it may allow the robots to be accepted into human society. The role of emotions in decision-making is another important perspective. In this paper, a model of emotions based on various neurological and psychological findings that are related to empathic communication between humans and robots is proposed. Subsequently, a mechanism of decision-making that is based on affects using convolutional LSTM and Deep Deterministic Policy Gradient is examined.

View full abstract

Recently, educational-support robot have been attracting increasing attention as learning-support devices. Previous research has proposed the sympathy expression method that robot expresses emotions which sympathize with learner. In previous research, the sympathy expression method has been used for the robot which expresses its emotion by face. However, no research for sympathy expression method using the robot which expresses its emotion by using body or unifying body and face. Therefore, this paper examines the learning effect and the impression about three types robots which have different emotion expression methods in long term experiment.

View full abstract

In this paper, we propose a method for dividing continuous time-series data into segments in unsupervised manner. Humans recognize perceived continuous information by dividing it into signicant segments such as words and unit motions. To this end, we have been proposed a method based on hidden semi-Markov model with Gaussian process (GP-HSMM). However, it has a big drawback that it requires the number of classes into which time-series data is segmented. To overcome this problem, in this paper, we extend GP-HSMM to nonparametric Bayesian model by introducing hierarchical Dirichlet processes (HDP), and propose hierarchical Dirichlet processes-Gaussian process-hidden semi-Markov model (HDP-GP-HSMM). Hence, the infinite number of classes is assumed and the number of classes are estimated by applying slice sampling. In the experiment, we used the various time-series data and showed that our proposed model can estimate more correct number of classes and achieve more accurate segmentation than baseline methods.

View full abstract

The action decomposition trees show intended goals of actions and their procedures. In order to improve readability of large-scale action decomposition trees, this research aims to develop the software’s function to display the portions of the trees which are related with user-specified information from various viewpoints. We define the general and domain-specific types of the viewpoints, which can be used by the users to specify the kinds of the information that they need. For users’ understanding of relationships between actions, the software also has a function to show the reason why the specific portions of the trees are displayed.

View full abstract

[in Japanese]

View full abstract

This project aims to develop a software system for knowledge sharing among care workers in elderly care facilities. In order to provide a controlled vocabulary set for common description of the knowledge of care processes, the authors are currently building an elderly care ontology. It mainly defines the physical and cognitive activities in the care processes as temporal changes between states of target objects. This article also discusses the usages of the ontology in the knowledge sharing system, which include adaptive generation of hierarchical structures of the vocabulary.

View full abstract

This collaborative research with a manufacturing company focuses on the adjustment process of design specifications of industrial machinery. The adjustment process consists of two steps; (1) to extract such specifications that need to be adjusted from the required specifications given by a customer and (2) to propose changes of those specifications for negotiation with the customer. In order to help engineers perform such processes efficiently, this research aims at knowledge modeling of past cases of adjusted specifications and ontology building for defining concepts in the knowledge models. The similarity calculation using the semantic structure of the ontology can improve the recall ratio of the extraction of the specifications to be adjusted. The semantic relations in the knowledge models could also contribute to higher accuracy ratio. Furthermore, the pattern knowledge of the past changes will realize semi-automatic generation of proposals of adjusted changes.

View full abstract

Many kind of Internet of Things (IoT) devices are recently developed. The IoT devices connect to an open platform and communicate each other with message exchange. Since the order of the message exchange is not deterministic, developers tests IoT devices in exhaustive order. The tests are usually conducted manually, which brings two issues: p1) testers possibly miss some of the order; p2) testers possibly overlook some evaluation items. As a solution, we developed a testing environment, which has a model checker that communicates with software on actual devices or an emulator. The testing environment solves p1) by automatic evaluation. The testing environment solves also p2) by automatic and simultaneous evaluation of evaluation items expressed as a Linear Temporal Logic (LTL) expression by integrating multiple LTL expressions with and-operator. We applied the test execution environment to three market defects and confirmed that the test execution environment tested the defects correctly.

View full abstract

In this paper, we propose pattern recognition for tennis tactics using ball trajectory data from motion capture system. The purpose of the study is to adapt machine learning in order to implement feature extraction of rallies in tennis game using positions of ball bounce. We modeled this task as time-series data statistical modeling based on the Hidden Markov Model. We also conducted experiments and we verified the dispersion of the mixture component and the centroid, corresponding to four types of tennis court area division. Moreover, we implemented feature extraction of rally according to the initial state probability and the state transition probability.

View full abstract

Data analytics is used in various field including business, science and sports. The evaluation of players and teams affects tactics, training and scouting in football teams. Players and teams are often evaluated by data such as shots and goals in game results. However it is not enough to fully understand potential of the players and teams. This paper describes a new analysis method using football passing data. In order to evaluate performance of players and teams, we employ graph mining. There is an indicator called centrality that evaluates individual contribution within an organization and it is used to evaluate the players and the teams. In calculation of the centrality for a given player pair, we consider not only the shortest path of passing but also longer ones. As a result, we found our method to be more consistent with game results than conventional methods.

View full abstract

In shield construction, it is concerned that the shortage of skilled workers causes serious accident due to oversight of fluctuation and delay in taking action. Therefore, revealing the cognitive processes of experienced shield machine operators is important. In this report, a part of the cognitive processes was revealed by analyzing eye tracking data and control logs of shield machine. Two machine operators were conducted eye tracking test. There were differences in viewpoint between the two operators as a result of comparison of the gaze movements. From Eye tracking and data of control machine, we revealed time lags of recognizing fluctuation and clarified operator’s sequences of estimating manipulation. We will develop operating support system and foster advanced operator in the future.

View full abstract

Automatic analysis of electrocardiograms (ECG) has been attempted. Although almost all methods pay attention only to short-term waveforms, their changes over time are said often important in diagnosis. In this paper, we focus on such long-term waveform change and propose a method to extract it as a pattern. The proposed method combines a method of expressing time series data as a trajectory in a feature space and feature extraction by an autoencoder. Evaluation experiments suggested the existence of regularity in the pattern and its association with the disease.

View full abstract

To improve the performance of location-based advertising, a model of ``sense of locality'' is estimated, where the output variable is the ``label'' of each location and inputs are geographical and demographic information associated with the location. As the input variables are all taken from the Internet, the output is unique dataset that we collects from people who know well the location. The model is estimated with three methods and XGBoost outperforms over logistic regression and SVM. The results show fairly predictive power with f1 score of 0.66.

View full abstract

[in Japanese]

View full abstract

The authors study the social knowledge centered on values. "Super aged society" indicates that social structure will change. We want to obtain the problem consciousness of the people from the free description reply, consider the connected values and explore the countermeasure.

View full abstract

[in Japanese]

View full abstract

Communication based on understanding of human is very important. As shifts to two-way communication to consumers progress, people's understanding with sensuous sensibility values is necessary. In this paper, we present the design of the mind to express sensory sensibility values.

View full abstract

In this paper, I aim to expand my research on the multiplicity of persons in kabuki. In particular, I propose a method for representing hierarchically the information of kabuki actors, which is divided into geinojin resource and life resource, using text data (Japanese Wikipedia) and I translate the text data into the information based on a property representation form to use it in my two types of narrative generation systems called geino information system and integrated narrative generation system. Last, I discuss various effects for expanding narrative multiplicities from kabuki person information that was divided into geinojin resource and life resource.

View full abstract

[in Japanese]

View full abstract

The purpose of this study is to analyze inbound tourists come to Japan. Segmenting the tourists by their values and comparing the satisfaction and the revisit intention among the segments. There are some useful studies of tourist segmentation by the motivation, however, the motivation items were defined by each study. The motivation might be changed by the time or the destinations. The values can be considered that are more stable than the motivation which will change by the situation. This study will attempt the values to inbound tourists’ segmentation analysis. Additionally, satisfaction and revisit intention are compared by the segments. Based on these analysis, repeater segment’s characteristics will be induced.

View full abstract

[in Japanese]

View full abstract

This study presents a validated recommendation on how to shorten the surveys while still obtaining segmentation-based insights that are consistent with the analysis of the full length version of the same survey. We use latent class analysis to cluster respondents based on their responses to a survey on human values. We first define the clustering performance based on stability and similarity measures for ten random subsamples relative to the complete set. We find foremost that the use of true binary scale can potentially reduce survey completion time while still providing sufficient response information to derive clusters with characteristics that resemble those obtained with the full Likert scale version. The main motivation for this study is to provide a baseline performance of a standard clustering tool for cases when it is preferable or necessary to limit survey scope, in consideration of issues like respondent fatigue or resource constraints.

View full abstract

Recently, some empirical studies argue that ``horizontal shareholding'' (or ``common ownership'') lessen competition among companies and prevent growth of an industry. Especially, horizontal shareholding by index funds occupy more weights of lists of shareholders and became a subject of greater debate. In this study, I built an artificial market mode, a kind of agent based model, and investigated an effect of increasing of horizontal shareholding by index funds to competitions and market prices. Our result shows that even when holding ratio of index funds is not so much they lessen competition. Moreover, when value of a company beating competition grow, its market price grow more than the company value (overshoot) and become over valued, then shareholders who encourage competition decrease and the company lose competition power. On the other hand, when value of a company losing competition drop, its market price fall deeper than the company value (overshoot) and become under valued, then shareholders who encourage competition increase and the company gains competition power. My simulation result indicated such mechanism balances competition powers among corporates. Growing index funds may possibly weaken the balancing mechanize.

View full abstract

[in Japanese]

View full abstract

[in Japanese]

View full abstract

If potential M&A cases can be detected automatically, this technology will improve the efficiency of M&A target recommendation and effectiveness of in-process M&A cases. However, in the past, M&A recommendation was impossible due to insufficient data and complexity of M&A. In this research, we provided a clustering method with cash flow features and company relationship features. From M&A clustering, we observed that M&A tend to concentrate in specific clusters. In order to improve the precision of M&A recommendation, we also analyzed the relationships between features from financial items and we extracted important features for identifying company relationships. The result of this research shows feasibility of recommending M&A from big data. In the future, we will design and select more features for analyzing M&A and we will associate results from AI with Management Science.

View full abstract

Back in the history of Japanese management mechanization, how has people changed and has achieved adaptation of organizational structure? In the organization, I would like to explore what kind of human resources they have educated the young people who entrusted the future and seek the adaptation of the organization structure in the next generation. For the possibility of the introduction of artificial intelligence, I want to find the intrinsic role entrusted to humans that robots can not substitute, so that a bright future can be found.

View full abstract

In this paper, we analyze newspaper articles using deep learning to forecast market trends. We have proposed a method to forecast market trends based on time-series text analysis using deep learning. This method works very well for forecasting TOPIX from The Nikkei (Nihon Keizai Shinbun) between 2008 and 2014, but the prediction accuracy falls after 2015. In this paper, we propose to reduce the duration of the training data in order to improve the prediction accuracy after 2015. As a result of the period of training data over the past three years, the prediction accuracy has been improved by 12.2%, from 55.1% to 67.3%.

View full abstract

Nowadays various kinds of data about customers behavior on online websites or applications can be collected for the marketing analysis. However, especially for the physical retailing stores, it is still difficult to acquire their behaviors on competing firms. In this research, we develop a novel approach of considering the usage of competitors and present the way to predict customers future purchase on retailing stores with ID-POS data combined with the third-party large-scale geofence location data collected from the smartphone apps.

View full abstract

We propose a time-series gradient boosting decision tree for a data set with time-series and cross-sectional attributes. Our time-series gradient boosting tree has weak learners with time-series and cross-sectional attribute in its internal node, and split examples based on dissimilarity between a pair of time-series or impurity between cross-sectional attributes. Dissimilarity between a pair of time-series is defined by dynamic time warping method or in financial time-seires by indexing dynamic time warping method. Experimental results with stock price prediction confirm that our method constructs interpretable and accurate decision trees.

View full abstract

In recent year, many researchers pay attention to stock trading using technical analysis. In the investment using technical analysis, we predict future price movements from patterns of past price movements and trade by using technical indicators to judge turning point and trends of the market. However, it is necessary to have expert knowledge and there is a problem that it is difficult to make a profit. Therefore, we construct investment strategy that can make a good profit using technical indicators.

View full abstract

This paper proposes an objective evaluation method for cystoscopic diagnosis of bladder cancer based on transfer learning using pre-trained DCNN (Deep Convolutional Neural Network) model. In the proposed method, lesion detection with comparatively fewer cystoscope images was realized by using the DCNN model pre-trained with a large amount of general images as fixed feature extractor and by learning the extracted feature vectors with subsequent classifiers. In this paper, in order to verify the effectiveness of the proposed method, experiments using actual bladder cystoscopic images were performed. As a result of the experiment, the proposed method achieved 95.7% in sensitivity and 93.3% in specificity in the two-class classification of normal and flat lesions which are difficult to distinguish, and showed the effectiveness for the cystoscopic diagnosis of bladder cancer.

View full abstract

This paper investigates the possibility of predicting depressive tendency by lifestyle data. Previous studies analyzed few aspects of lifestyle, whereas the current study utilizes multivariate analysis to examine multiple perspectives of lifestyle, such as social, sleeping, and dietary habits. We created a questionnaire including depressive tendency score (K6) as well as lifestyle, and recruited 987 participants, who answered it using a crowdsourcing service. Classification models were obtained using machine learning to classify the participants as high or low depressive tendency. Random forest classifier achieved 0.97 accuracy. Particularly effective features were chosen from Chinese medicine, and personality describing neuroticism.

View full abstract

Pelvic angle is one of the important indicators to assess standing or walking posture. Conventional method to measure pelvic angle are using motion capture or inertial sensor that require to put markers or sensors on the body. This may impose a burden to subjects. In this research, we proposed a contactless pelvic angle measurement method using point cloud obtained from depth camera. We applied ICP algorithm to extracted point cloud of pelvis region from depth camera, and estimated the angular displacement caused by motion. As a result, proposed method was effective in measuring pelvis angle in upright position. Although, there was not much correlation with the true value in the results of estimation while walking, it was suggested that this method could capture the feature of walking postures in different pelvic angles.

View full abstract

The fluctuation of an RR interval (RRI) on an electrocardiogram (ECG) is called heart rate variability (HRV). Since HRV reflects the activities of the autonomous nervous system, HRV has been used for many kinds of health monitoring systems. However, HRV is easily influenced by arrhythmia, which prevents the precise health monitoring. The present work focuses on premature ventricular contraction (PVC) which is common arrhythmia. To modify RRI data with PVC, the present work proposes a new method based on denoising autoencoder (DAE), referred to as DAE-based RRI modification (DAE-RM). The performance of DAE-RM was evaluated by its application to clinical RRI data which contains artificial PVC (PVC-RRI). The root mean squared error (RMSE) of modified RRI was improved by 84% from PVC-RRI. The result showed that DAE-RM could modify PVC-RRI data appropriately. The proposed DAE-RM has the potential for realizing precise health monitoring systems which use HRV analysis.

View full abstract

[in Japanese]

View full abstract

Segmentation of medical images with high precision and speed is an important task in many medical scenes. One such method for this task is GraphCut based on energy minimization problem. However, in GraphCut, it is difficult to perform segmentation completely and automatically if adjacent pixel values are similar. There are many methods for this problem, but most of them are not suitable in speed. In deep learning methods, automatic segmentation is possible because of its capability of capturing complicated features. In this research, we propose a model incorporating 3D U-Net extended with Residual Unit and 3DCNN for correcting segmentation results.

View full abstract

In this study, we propose a method called auxiliary weight (AW) for neural networks in which each input value is weighted according to its contribution to the input dimension. AW is similar to Lasso regularization in the sense that it can extract features; however, AW is faster than Lasso in processing data that contains a several contributing dimensions and massive non-contributing dimensions, such as the data of medical mass spectrometry. (Code:https://bitbucket.org/akira_you/awexperiment)

View full abstract

In mental disorder diagnosis based on fMRI images, conventional studies perform preprocessing such as feature-extraction using correlation analysis since the fMRI dataset is composed of a small number of high-dimensional samples. However, this preprocessing could miss features necessary for diagnosis. On the other hand, deep generative models achieved good accuracy even with a small dataset with limited preprocessing. In this paper, we model fMRI brain images using a deep generative model with a subject-wise variable. The proposed model explicitly separates individual differences from the mental disorder and noise in fMRI images. The proposed model achieves accuracy higher than conventional methods.

View full abstract

We report a segmentation process of multiple anatomical structures in chest X-ray pictures, which is a key element of computer-aided diagnosis (CAD) systems, and its evaluation results. The segmentation process utilizes U-Net, that is, a type of fully convolutional network. The segmentation targets are a small region like a first thoracic vertebra and a line structure which is boundary between anatomical structures. In the evaluation, we achieve Dice index 0.91 as the segmentation accuracy for the first thoracic vertebra, and Dice index 0.71 through 0.81 as the segmentation accuracy for the line structures.

View full abstract

Neural machine translation (NMT) has significantly improved quality over traditional statistical-based machine translation (SMT). However, it is known that it is difficult for NMT to translate sentences containing rare terms. Therefore, in this research, we propose a method of replacing rare terms with synonyms and translating it, and replacing the translated synonyms with translations of the rare terms in the bilingual dictionary. This approach has two technical issues: acquisition of synonyms from small scale corpus, and selection of the synonyms. The proposed method shows the better result than the result of the existing technics in sign test.

View full abstract

Contact centers have massive amount of dialog records to be used to improve the quality of their service. This paper describes a method of summarizing a large number of dialogs on the same topic into a tree structure. The method consists of two steps: 1) summarization of each dialog with phrase structure rules, and 2) organizing dialogs into a tree structure using text clustering. A common oddness was observed among summarized dialogs, and this problem was mitigated by applying Multi-Sequence Alignment (MSA). With the proposed method, we were able to summarize real-life dialogs into a reasonably small tree with only two hours of rule writing labor. Also, applying MSA helped to reduce the number of nodes and led to a higher purity score.

View full abstract

In sightseeing information navigation systems, the information presented by natural language has a potential to improve usability. Several systems tried to embed the informing contents in a prepared template for generating sentences which are useful for tourists, which is called a slot filling based method. However, it is difficult for the systems to generate diverse expressions and unseen patterns. To solve this problem, we propose a neural network based sentence generation method instead of using a slot filling based method. In this research, we construct the contents as a one-hot vector representation and construct the neural network based language generator and the one-hot content vectors for generating natural and understandable sentences. We collected a tourist information corpus via crowdsourcing. Existing language generation systems used word classes. However, these systems often connect words unnaturally. In this research, we also proposed a re-ranking system based on a neural language model to solve the problem. In our experiments, we confirmed the naturalness and validity of the sightseeing guidance sentences generated by our proposed method.

View full abstract

[in Japanese]

View full abstract

[in Japanese]

View full abstract

Learning systems with haptic or Pseudo-haptic feedback has developed. However, a study have not been conducted about effect of haptic and Pseudo-haptic feedback in learning system. Therefore in order to investigating effect of haptic and Pseudo feedback in learning system, we developed Learning system with feedback based on haptic and Pseudo-haptic. This paper describes the experiment using Kanji learning support system and the experiment using Pulley learning support system.

View full abstract

Sacred light is commonly used in a residence in Heian period, however it is not appropriate to use in museum now, because of fire prevention. Virtual antique folding screen with optical effects of sacred light enables us to appreciate the antique folding screen as appreciated by Heian noble people.

View full abstract