Proceedings of the Annual Conference of JSAI
Current issue
Showing 1-50 articles out of 514 articles from the selected issue
  • Masanori TAKANO, Yuki OGAWA, Fumiaki TAKA, Soichiro MORISHITA
    Session ID: 1D2-OS-3a-01
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    People can watch television news not only on terrestrial television (linear) but also on online websites (on-demand). On linear, people watch the news on news channels that broadcast programs according to certain time schedules. On the other hand, on-demand, they can choose news programs to watch. We study news-program repertories in a mixed environment of linear and on-demand.

    Download PDF (302K)
  • Tsubasa SHINTOU, Yuuki OGAWA, Hiromitsu HATTORI, Fumiaki TAKA, Fujio T ...
    Session ID: 1D2-OS-3a-02
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    In recent years, with the spread of the Internet, the media has become more diverse, and the number of information sources that individuals can choose has increased explosively. Among them, selective exposure that selects only the information that one likes is treated as a problem. If many people make selective contact, correct information sharing will not be possible, and there is a risk that the division of communities and political ideas will further progress. In this study, we analyzed the selective exposure by focusing on the questionnaire survey at the time of the 2019 election and the Twitter usage data of the subjects. Specifically, we found out the tendency of information contact and the community from Twitter followers and retweet information, and analyzed it based on the answers about political attitudes and consciousness in the questionnaire survey. As a result, we were able to extract a characteristic community from the bias of information contact. It was also found that there are differences in political attitudes among the communities. We were also able to extract the commonalities between selective exposure tendencies and political indifference, and the communities that apply to them.

    Download PDF (383K)
  • Shotaro ISHIHARA, Norihiko SAWA
    Session ID: 1D2-OS-3a-03
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    This paper proposes a method to summarize news articles by sentence selection and compression. We can extract N texts which represent the article, and enumerate summary candidates by compressing each text through syntactic analysis. MMR (Maximal Marginal Relevance) and TF-IDF (Term Frequency - Inverse Document Frequency) are used as metrics. Experiments showed that the proposed method was able to extract the same topics as the human editor's summary in the rate of 26%. Even though the rate wasn't high enough, most of the achievements couldn't be described as incorrect as one of the summary candidates. This methodology has a potential to reduce the burden on editors and generate some collaboration.

    Download PDF (1301K)
  • Shohei HISADA, Taichi MURAYAMA, Shuntaro YADA, Shoko WAKAMIYA, Eiji AR ...
    Session ID: 1D2-OS-3a-04
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    In the era of social media, we are often exposed to unintentionally biased information due to filter bubbles. Such biased information amplifies opinion fragmentation and political polarization. To cope with this problem, we analyze the bias of the news media and help people understand the news correctly. The existing survey on the bias is conducted by expert analysis and crowdsourced evaluation to the media. In this study, by using a topic model for Twitter comments on news, the distance between news media is calculated from hierarchical clustering's estimated probability of the topic. The fundamental idea is to measure the bias by analyzing the content of the topic and the news media's similarity. When we applied this method to tweets about the Science Council of Japan issue, we found that the results for mainstream media such as Asahi Shinbun were roughly consistent with other studies on political bias. We found that it is possible to capture bias in media that have not been surveyed before.

    Download PDF (746K)
  • Ryosuke SAWA, Takeo KIBAYASHI
    Session ID: 1D2-OS-3a-05
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    In 2020 Japan, a lot of news was reported that would make the automotive industry lively. Most of the content is urging the automotive industry to change, such as the electrification policy for carbon neutrality in 2050. However, the reality is that the workers in the automotive industry are not aware that change is necessary, I saw them as colleague. It is presumed that the cause of not being aware is due to social media and low literacy of the masses. Social media displays news with user's preference by AI, masses cannot understand the fact from news. As a result of investigation the world situation with the aim of making the workers of the automotive industry aware of the need for business transformation. it is easy to speculate that the automotive industry will collapse. Discuss the need to convey the facts of the collapse through the social media.

    Download PDF (694K)
  • Yoshiki FUJIKANE, Kazuhiro KAZAMA, Mitsuo YOSHIDA, Yoshinori HIJIKATA
    Session ID: 1D3-OS-3b-02
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    In this paper, we propose a method to automatically find controversial news articles in order to analyze the bias of opinion in mass media or social media. First, we define the controversy measure of a news article using the number of users who mentioned it and the number of days that it were mentioned, assuming that news that causes controversy and debate in social media is mentioned by a limited but certain number of users for a relatively long period. In addition, we analyze the polarization and cluster structure of media graphs and user graphs of specified news topics and the context, and verify whether we can find controversial news articles.

    Download PDF (780K)
  • Atom SONODA, Hiroto NAKAJIMA, Fujio TORIUMI
    Session ID: 1D3-OS-3b-03
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    The phenomenon of filter bubbles and echo chambers has become a social issue. Our goal is to quantitatively evaluate these behaviors from log data. So far, we have discussed users' behavioral changes based on the diversity of article categories or vectors obtained from linguistic information of the titles. However, we believe that it is important to understand the characteristics of the behavior of users with specific interests in order to essentially understand issues such as filter bubbles. In this paper, we assessed the magnitude and change in user interest in the covid-19 against long-term data and analyzed the impact on engagement. We also showed that using the title information, which changes with time, as well as click logs allows for more accurate analysis of the news articles.

    Download PDF (449K)
  • An Application of J-LIWC, J-MFD, and Word-coocurrence Networks
    Kazutoshi SASAHARA, Shimpei OKUDA, Tasuku IGARASHI
    Session ID: 1D3-OS-3b-04
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    On social media, a variety of information is posted and shared in real time on a daily basis. Many of these posts relate to the impact of the COVID-19 pandemic and consumer life. Such spontaneous “real voices” (texts) from consumers contain potential signals for quantitatively understanding consumer psychology and behavior in the pandemic. In this study, we quantify consumer psychology and behavior from large-scale social data on Twitter, and discuss analytical methods and cases to gain insights into the resale phenomenon in the pandemic. Specifically, by applying our psychological category dictionary “J-LIWC” and moral foundation dictionary “J-MFD” to the above data, we were able to visualize that different consumer emotions were evoked depending on the types of product resold, such as toilet paper or masks, and what kind of moral violation consumers perceived about the resale. Furthermore, by analyzing the above data using a word co-occurrence network, we observed trends in products resoled and changes in purchasing behavior linked to the pandemic’s progression. These findings provide a hint for reconstructing the society and economy toward the post-corona era.

    Download PDF (883K)
  • Shogo MATSUNO, Saeyor SANTI, Takeshi SAKAKI, Yasuhiro HINO
    Session ID: 1D4-OS-3c-01
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    The impact of social media on the diffusion of information is becoming increasingly difficult to ignore in marketing communications and news dissemination. In particular, the diffusion of information through social media is said to play a major role in the spread of echo chambers and fake news. In this research, we would like to clarify what factors affect the scale of information diffusion on social media in corporate PR and news dissemination. In this paper, we define the characteristics of influencers as: 1) users who have many followers who spread their posts, and 2) users who post many tweets (≅. users who spread posts without hesitation), and examined the influence on post spread (retweet) using a social graph constructed from Twitter records. As a result, we found that the probability of a post being spread is higher for users with either of these characteristics than for randomly selected users. And in particular, the probability of a post being spread is highest for users with many followers who have a small number of followers/followers in their private graphs.

    Download PDF (820K)
  • Keito ISHIHARA, Shotaro ISHIHARA, Hono SHIRAI
    Session ID: 1D4-OS-3c-02
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    In this study, we tackle abstract summarization of Japanese news articles using BERT, which is common in the field of natural language processing in recent years. Specifically, we use BertSum, a summarization method that is an extension of BERT. We trained BertSum using three types of BERT, and the experiment showed that Japanese pre-trained models worked better than multilingual model. There was no significant difference in the performance of the model pre-trained on Japanese news articles and Japanese Wikipedia. We also discussed tokenizers and unknown words, which are important in dealing with news articles in Japanese.

    Download PDF (382K)
  • Yuki OGAWA, Masanori TAKANO, Soichiro MORISHITA, Fumiaki TAKA
    Session ID: 1D4-OS-3c-03
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    In this study we analyze the relationship between news tweet viewing and news video viewing. Specifically, we determine whether news tweet viewing has an effect on ABEMA news video viewing. The results of the analysis show that viewing tweets on socially important topics such as international and corona are associated with continued viewing of news videos.

    Download PDF (810K)
  • Yutaro SATO, Godai SAITO, Kenri KODAKA
    Session ID: 1E2-OS-2-01
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    Sense of numbness arises when one person holds the palm of one hand against another person’s opposite palm and strokes with his other hand the two joined index fingers (double-touch). While the numbness is thought to be a side effect from ownership distortion, no study has investigated direct relationship between ownership and numbness. In this study, we examined the effect of angles between two fingers in the double-touch situation on sense of ownership, numbness and transform for the experimenter’s finger. The result showed that opening and closing both fingers enhances the sense of ownership, numbness and transform in a strongly correlated manner.

    Download PDF (1147K)
  • Discussion from viewpoint of cognitive bias
    Noritaka KANAIZUMI, Takeshi ITO
    Session ID: 1E2-OS-2-02
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    After the appearance of Alpha Zero, in two-player deterministic and perfect information game like Shogi and Go, AI stronger than human was invented. WereWolf Game has been regarded as the new theme of game study. The project named “Artificial Intelligence based Werewolf” has the ultimate goal: “Building an agent which can play werewolf with humans naturally”. Researchers make progress in developing werewolf AI, but it is still few that studies reveal how human plays werewolf. In this study, we pay attention to “Why are players deceived” in decision-making process. Concretely, we examined how human makes decision by showing participants similar stories which changed a part of remark a bit. As the result, we suggest the influence of cognitive biases on human by comparing the examples which the participants are deceived and those which are not. We found that human does not necessarily make rational decision while playing werewolf game and that he/she conducts decision-making under the influence of various forms of bias depending on the content of what they are focusing on.

    Download PDF (653K)
  • Daiki KONDOH
    Session ID: 1E2-OS-2-03
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    Since the outbreak of the new coronavirus, the assessment of the validity of information has become important from the national to the individual level. Belief bias is one of the cognitive biases that make them deviate from what is rational. In previous studies, it has been confirmed mainly by syllogism, but there is a big difference between syllogism and the information that people actually evaluate, and this difference may cause a difference in the individual tendency of belief bias. Therefore, in this study, we created a text task that is similar information people evaluate in real life, in order to examine whether there is a difference in the tendency toward belief bias. The results showed that there was no relationship between the tendency toward belief bias in syllogism and the text task, suggesting that the mechanism of belief bias in each task is different. Next, in order to explore the mechanism of belief bias in the written task, I developed a model that assumes that people evaluate the probability that the information in the text is correct, and make a judgment about its validity. The model predicted about 50% of the belief bias.

    Download PDF (411K)
  • Teruaki HAYASHI, Yumiko NAGOH, Kai ISHIKAWA, Hirohiko ITO, Kenichiro T ...
    Session ID: 1E3-OS-8a-01
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    The technologies using Artificial Intelligence (AI) have been implemented as services and have been expected to solve various social problems. However, AI’s contributions to people’s mentality and unknown/unobserved events have not been discussed. The imaginary part of the information which has not been converted into data involves a fundamental unexplored problem. Therefore, in this study, we focus on the trends in the society and the people’s mental changes in the Corona-related confusion and discuss the externalization of data sources of unexplored data using Marketing Research Online Communities and Variable Quest.

    Download PDF (875K)
  • Kai ISHIKAWA, Takuya MIZUKAMI, Soichiro TODA, Tomohiro INOKUCHI, Haruk ...
    Session ID: 1E3-OS-8a-02
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    The interaction between human mind and AI provides a foothold for understanding social acceptability. People's expectations for AI are influenced by how it is perceived, and whether AI should consider a certain attribute of data also depends on its meaning in society. In addition, the advent of mind-changing technology has enabled the means of "rapid rehabilitation" and is creating a new dilemma with the execution of punishment based on rehabilitation and social retribution. In this presentation, we will examine the applicability of AI to social implementation design through the above case studies.

    Download PDF (432K)
  • Kaira SEKIGUCHI, Koichi HORI
    Session ID: 1E3-OS-8a-03
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    AI technologies are making rapid progress and demanded to consider AI ethics. Actually, many AI ethical principles, guidelines, results of case studies, etc. have been provided by academic societies, governments, international organizations, etc. However, few AI projects in engineering positively introduce such AI ethics; there is a gap between AI technologies and AI ethics. Recent studies have shown that an AI-based creativity support tool can contribute to fill the gap between them. At the same time, these studies are clarifying that one creativity support tool alone is insufficient. One of the reasons is that ethics relates societal values and no system can be neutral on these values. Therefore, this research proposes to develop a grass-roots AI network to support ethical design practice including ethical AI and to realize complementary relationships among the systems. Then, this paper reports a result of a study that simulates to network two existing creativity support tools and confirms the feasibility of such complementary relationships.

    Download PDF (773K)
  • Takashi MATSUMOTO, Arisa EMA
    Session ID: 1E3-OS-8a-04
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    Various problems related to the trustworthyness of artificial intelligence (AI) services have been pointed out, and the necessity of third-party audits and guarantees has been raised in the world. However, the decision-making of AI are uncertain. And significant risks of AI depend on the purpose of AI services, learning data, algorithms, degree of human intervention, user characteristics, and etc. In addition, AI model alone can't mitigate the risks sufficiently and continually. In this study, by using a risk control model (risk chain model) , we classified "Values and objectives to be realized" and "risk scenario" from 30 use cases(AI services) through discussions with experts.

    Download PDF (750K)
  • Applying AHP for structuralization, visualization, sharing, and recognizing diversity of opinions and value systems inside individuals and groups
    Gensei ISHIMURA
    Session ID: 1E3-OS-8a-05
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    Several methods have been tried for handling trans-science issues by participation of citizens. However, they need much resources, and the discussions there include many options and evaluation criteria and participants find difficulty in visualizing and sharing important agendas. To solve the problem, we tried to develop novel workshop method based on AHP (Analytic Hierarchy Process) for collective decision making, consensus building, community building, and citizenship learning. The method developed in this study suggested that the value systems of the workshop participants and their diversity could be structurally visualized and shared, and that this would affect each participant's decision-making confidence.

    Download PDF (467K)
  • Takuichi NISHIMURA, Yasuyuki YOSHIDA, Chiaki OSHIYAMA, Koki IJUIN, Nam ...
    Session ID: 1E4-OS-8b-01
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    Expressions based on physical movements such as ballroom dancing and hip-hop have been incorporated into school education and are highly evaluated for their effectiveness in preventing dementia. One reason is that it moves the body comfortably to the music. The other one is that it includes brain activity while communicating with others. In such a physical expression, the motor learning process has many individual differences and is unknown. In addition, the power of expression varies greatly depending on the current mental state. In this paper, we will ground linguistic knowledge structuring to various nonverbal data. We will show an example of ballroom dancing and challenge the collection and analysis of these implicit and unexplored data.

    Download PDF (919K)
  • For knowledge search in collaboration using AI
    Chiaki OSHIYAMA, Takuichi NISHIMURA
    Session ID: 1E4-OS-8b-02
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    In the clinical site, interventions are made to solve the problems that the target person has an any trouble. Clinical professionals use their expertise with a purpose to effectively carry out the intervention. The intervention technique contains a lot of unformal tacit knowledge. However, its contents are not visualized present in detail. Therefore, it was difficult to share among experts with different degrees of proficiency or among experts with different fields. Information can be shared among various experts and supporters by structuring and visualizing knowledge for problem solving. We have been collected and visualized solutions as unexplored data based on the knowledge and implementation details of various people in various situations that we have not been able to retrieve. We would like to make a proposal to create a system that service providers can access enormous problem-solving knowledge, can send out themselves, and service recipients can easily obtain advanced services.

    Download PDF (1684K)
  • Hiroya MIURA, Nami IINO, Masatoshi HAMANAKA, Hideaki TAKEDA, Takuichi ...
    Session ID: 1E4-OS-8b-03
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    We are conducting research on the development of artificial intelligence (AI) applications to support decision making in the practice of skills based on the perspective of skilled users. In this paper, we analyze verbal and non-verbal information in a lesson of classical guitar, which is one of the skills, to clarify how the teaching process of instructors changes by structuring knowledge.

    Download PDF (470K)
  • Hiroki WATANABE, Eiichi OSAWA
    Session ID: 1F2-GS-10a-01
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    Infrastructure systems that exist in modern society are interdependent. In such a system, due to the interdependence, the influence of a failure that occurred in a certain system can be propagated to other systems, and there is a possibility that a cascade failure occurs. Improving the robustness of the network is an important issue, however, it is difficult to completely prevent failures in the network because of its nature. Resilience is a word that has both meanings of recovery and resistance. Improving resilience means improving the ability to cope with disabilities. In this paper, we propose and verify a network restoration order determination method using the characteristics of complex networks network interdependency for the purpose of improving the resilience of network. Experimental results show that the proposed method is effective for determining the recovery order of a damaged network.

    Download PDF (636K)
  • Non KAWANA, Ken NAGANUMA, Masayuki YOSHINO, Chiaki OTAHARA, Yumiko TOG ...
    Session ID: 1F2-GS-10a-02
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    e-KYC: electric-Know Your Costumer is a system in which user authentification is performed via images/videos over the Internet for opening an account at a bank. Also, its procedures are performed non-face-to-face. In this paper, we experimented to see if it is possible to spoof a user authentification such as e-KYC by using Deepfake. In this paper's experiment, we first made an original e-KYC application based on OSSs and applied Deepfake attack for this system. More precisely, our e-KYC application requests random actions, such as ‘tilting your face’, to the user and verifies it with the stored image. As a result of the experiment, the spoofing attack was successful; that means our Deepfake attack is turned out to be a practical thread for e-KYC. Also, we summarize some countermeasure technologies against this attack.

    Download PDF (536K)
  • Soichi ONOZUKA, Yuta HIGUCHI
    Session ID: 1F2-GS-10a-03
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    In the case of error detection of Web screens by the convolutional neural network of the previous research, it is necessary to label the training data of normal and error. In error detection on a Web screen, it is difficult to collect training data by assuming an error screen in advance because it is uncertain what kind of error will occur. In this study, we compared the images acquired in the past and calculated the similarity to detect errors probabilistically without labeling the training data. As a result, Few-shot learning is possible to emphasize the characteristics of the training data from less past data, and detect error candidates on the Web screen by link prediction of the graph neural network (GNN).

    Download PDF (665K)
  • Yuuki YAMANAKA, Masanori YAMADA, Tomokatsu TAKAHASHI, Tomohiro NAGAI
    Session ID: 1F2-GS-10a-04
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    Tampering with just one byte of traffic payloads used in industrial control systems (ICS) can cause serious physical accidents. Therefore, it is necessary to analyze the payload in a cyber attack detection system targeting ICS. However, since various protocols are used in ICS, a high level of expertise is required to manually extract the features from the payload. Therefore, in this paper, we propose a method for automatic payload analysis using Bidirectional Encoder Representations for Transformers (BERT). By treating each byte as a word and using BERT, we can obtain one fixed-length feature vector from the payload. The vector contains information such as the position of each byte and its relation to to nearby bytes. We experimentally show the effectiveness of the proposed method on several ICS datasets in the anomaly detection task.

    Download PDF (262K)
  • Tomohiro NAGAI, Yasuhiro TERAMOTO, Masanori YAMADA, Yuuki YAMANAKA, To ...
    Session ID: 1F2-GS-10a-05
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    The threat posed by unknown cyber attacks requires detection of intrusions and incident response, as the cyber attack may cause physical damage in Smart Factory. Recently, malicious attacks have rewritten parts of the payload to mimic normal payloads.Much recent research focuses on deep learning based anomaly detection. However, previous work on anomaly detection have not focused on the presentation for explainable decisions. In this paper, we propose methods for explanation of anomaly detection using decisiton tree. We evaluated using a dataset obtained on a factory simulator to demonstrate its ability to present the anomaly bytes of cyber attacks.

    Download PDF (404K)
  • To improve employees’ work engagement
    Una SHIMOKAWA, Minori FUJIOKA, Ryo ICHIYAMA, Aoi TAKAYA, Nanako NARIKI ...
    Session ID: 1F3-GS-10b-01
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    This study aims to confirm the factors and programs to improve work engagement of employees for the purpose of promoting Health and Productivity Management using AI. Although mental health care is promoted, there are still many employees with psychological problems. Moreover, it has been getting serious along with COVID-19. In recent years, many researchers have investigated mentalhealth care. And it has been reported that improving work engagement reduces stress for employees. However, it is said that one of difficulties of Health and Productivity Management is the lack of work force because of declining working population. Therefore, we focused on primary prevention and compared the effectiveness of the program by AI and human to improve work engagement. As a result, there was no significant difference between them. In addition, this paper refers to the promotion of Health and Productivity Management by analyzing correlations between personalities and work engagement.

    Download PDF (492K)
  • Kei SATO, Hiroyuki SAKAI, Kaito TAKANO, Daisuke INOUE, Kana FUJINO
    Session ID: 1F3-GS-10b-02
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    This paper proposes a method for estimating relationship between corporate philosophy and performance factors in companies extracted from securities reports. The corporate philosophy is a verbalization of the characteristics of the fundamental way of thinking and ideals such as the company's thought, mission, and aspirations. And, performance factors in companies is a sentence that describes the factors that led to the performance, such as "the sales of *** was good". Our method extracts sentences containing corporate philosophy and performance factors in companies from the securities report by using deep learning. Then, the extracted corporate philosophy sentences and performance factor sentences are expressed in a distributed expression for each sentence by using Word2Vec. Our method calculates the degree of similarity between corporate philosophy sentences and performance factor sentences, and that is used as the relevance. As a result of the evaluation, our method for extracting corporate philosophy attains 77% accuracy, and the relationship between corporate philosophy and performance factors in companies was estimated with relatively good accuracy.

    Download PDF (574K)
  • Akihide HIGUCHI, Masanao OCHI, Junichiro MORI, Ichiro SAKATA
    Session ID: 1F3-GS-10b-03
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    Mergers and acquisitions (M&A) are one of the most important management strategies. However, the execution of M&A does not always lead to post-acquisition success, and it is difficult for even experts to realize such success. Therefore, a research on the success and recommendation of M&A is needed. In this study, we proposed an index to evaluate M&A quantitatively from two perspectives, and showed the possibility to recommend M&A by analyzing the trend of the index. First, the expected synergy realization was evaluated using the existence of goodwill impairment, and a significant difference was found for the capital adequacy ratio. Next, as for the synergy of name recognition, the index was calculated by capturing the change in Google Trends before and after the acquisition, and the difference in the name recognition of the pair before the acquisition was shown to be a possible explanatory variable.

    Download PDF (354K)
  • Akio ITO, Kei NAKAGAWA
    Session ID: 1F3-GS-10b-04
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    The carry trade strategy has received much attention under the global low interest rate environment.The carry trade strategy refers to a strategy that maximizes income gains such as dividends and interest, and can generate profits if the price fluctuation is smaller than the income gain.Many empirical analysis shows that although carry trade strategies have positive returns on average, they are known to produce very large losses (tail risk) with small probability. Therefore, in this research, we construct a carry trade strategy using Conditional Value at Risk (CVaR), which is a risk measure for controlling tail risk, and aim to generate stable profits while suppressing tail risk. To that end, we focus on the Regularized Multiple CVaR portfolio, which solves the instability of CVaR portfolio, and propose to introduce a carry level as an expected return. We perform the empirical study and confirm that the risk/return and tail risk are better than the baseline method.

    Download PDF (354K)
  • Kohei KAWAMURA, Kaito TAKANO, Hiroyuki SAKAI
    Session ID: 1F3-GS-10b-05
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    Stock prices often move with future performance forecasts rather than past performance, and even if the current performance is in the red, the stock price may rise if the company shows that the company's performance will recover. Therefore, performance forecast information is important for investment decisions. In this paper, we propose a method to extract forecasted business performance (statements that describe future performance forecasts of companies) from summaries of financial statements as a support for individual investors' investment decisions. In particular, we focus on forecasted business performance containing performance factors. Our method is able to generate highly accurate training data automatically, and by using the pre-trained BERT as the classification model, a relatively good result was obtained with an F1-score of about 83.6.

    Download PDF (460K)
  • Koki KANIE, Takashi ONODA, Takahiro NISHIGAKI, Yukihiro SHINODA, Makot ...
    Session ID: 1F4-GS-10c-01
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    At present, a person is performing the determination of the ground fault appearance at the time of a distribution line failure, which requires time cost and human cost. In addition, there are changes that push up power transmission and distribution related costs, such as the introduction of renewable energy due to power system reforms, and the power transmission and distribution sector must reduce costs by improving efficiency. Therefore, there is a demand for automation by a machine for determining the appearance of a ground fault when a distribution line fails. Existing studies have reported the possibility of distinguishing between ground fault-like cable degradation and other ground fault-like features. Therefore, in this study, we will examine the discrimination of five other types of ground faults in the existing studies. The current waveforms of five types of ground faults were learned by a support vector machine, and their potential discrimination performance was evaluated by leave-one-out cross-validation. As a result, a high accuracy rate could be obtained in all classifications of ground faults.

    Download PDF (788K)
  • Yuta TAKAHASHI, Junichiro FUJII, Masazumi AMAKATA, Takayoshi YAMASHITA
    Session ID: 1F4-GS-10c-02
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    Despite the development of deep learning, the number of application cases in the civil engineering field has not increased so much. For the reason, it is difficult to define the boundary conditions of the problem to be solved, and although there are various abnormalities to be detected. Additionally, the anomaly data may be less or not exist. For example, illegal dumping and illegal occupation in river patrols can take various forms depending on the context. Considering river patrols using UAV / AI now, we can assume a situation where there are few aerial images at the start stage. In this study, we verified whether the learning data could be complemented by learning together with the images obtained on the ground registered in the river maintenance database RiMaDIS, using Faster R-CNN. For ground image selection, images close to the feature space of aerial images were selected based on several criteria and methods. Among the proposed methods, ground images selected by the occupancy rate of the Bounding Box and the Deep Network (ShuffleNet, Inception v3) improved the average Precision.

    Download PDF (1232K)
  • Junichiro FUJII, Ryuto YOSHIDA, Masahiro OKANO, Masazumi AMAKATA
    Session ID: 1F4-GS-10c-03
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    The inspection of civil engineering structures such as river revetments has traditionally been carried out visually by engineers. Visual inspection requires a great deal of labor and is subjective to the engineer's judgment, resulting in inconsistent inspection records. In order to solve these problems, an inspection method that applies image recognition technology based on deep learning is being researched. However, for the maintenance and management of structures, image recognition results alone are not sufficient. It is necessary to convert these results into physical quantities that can be used as indicators to determine the health of structures, and to monitor the changes over time. In this study, we propose a river revetment monitoring method based on the physical quantities of area, width, and length calculated for the crack detection results by deep learning. We also report the results of applying this method to actual rivers.

    Download PDF (1535K)
  • Naoki ISHITSU, Kenichi KUSUNOKI, Toru ADACHI, Hanako INOUE, Chusei FUJ ...
    Session ID: 1F4-GS-10c-04
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    We are developing a gust detection system using meteorological Doppler radar. When a gust such as a tornado occurs, a vortical airflow is generated in the lower level of the cumulonimbus cloud, and observation by a Doppler radar shows a pair of maximum and minimum Doppler velocities. So far, we have developed a mathematical model of this pattern, fitted it to observed data, and determined whether it is a vortex or not based on the calculated physical quantities. The problem with this method, however, is that it often results in many false detections and misses. In this study, we tried to distinguish vortices by using CNN and found that the performances were greatly improved by applying the CNN-based vortex determination.

    Download PDF (559K)
  • Katsuya KOSUKEGAWA, Yasukuni MORI, Hiroki SUYARI, Kazuhiko KAWAMOTO
    Session ID: 1F4-GS-10c-05
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    This paper addresses the problem of forecasting railway track irregularity, which is essential to secure the safety and to manage maintenance planning and scheduling. We propose a track degradation model using convolutional long short-term memory (ConvLSTM) for forecasting vertical irregularity of Shinkansen rail lines. The ConvLSTM architecture is designed to be trained on three types of data: spatio-temporal series data on the rail lines measured by a high-speed inspection train, static categorical data such as track structure and foundation, and binary time series data of maintenance workday records. We evaluate forecasting performance in terms of classification accuracy. Experimental results with real data show that the ConvLSTM model provides better forecasting performance when using static categorical data such as track structure and foundation, and binary time series data of maintenance workday records, at the critical points where the track degradation is progressing.

    Download PDF (620K)
  • Yuki NAKAGUCHI
    Session ID: 1G2-GS-2a-01
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    Recently, reinforcement learning (RL) has been showing increasingly high performance in a variety of complex tasks of decision making and control, but RL requires quite careful engineering of reward functions to solve real tasks. Inverse reinforcement learning (IRL) is a framework to construct reward functions by learning from demonstration, but there is no way to guarantee the performance of the learned reward functions in maximum entropy IRL, the mainstream of IRL. Therefore it is unclear how reliable the results can be. To provide a theoretical guarantee on the performance of maximum entropy IRL, we evaluate and discuss its performance theoretically.

    Download PDF (360K)
  • Kuratomo OYO, Takuma WADA, Takumi KAMIYA, Tatsuji TAKAHASHI
    Session ID: 1G2-GS-2a-02
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    Multi-armed bandit problems, which are the most fundamental tasks in reinforcement learning, have been widely applied to a range of problems such as online advertisement delivery and game tree search. In contrast to these traditional bandit problems that require absolute rewards to be quantifiable, dueling bandit problems (DBP) can deal with relative rewards by pairwise comparisons. In DBP, one of the most effective solutions is Double Thompson Sampling (D-TS). However, due to the pairwise comparisons, solving DBP requires many trials and errors, and that causes D-TS to do a lot of computation. In this paper, we focus on the fact that “satisficing” action selection leads to quick search for an action that satisfies a certain target level. We propose an algorithm that is based on Risk-sensitive Satisficing (RS) model. The result showed that there are some datasets on which its performance was inferior to D-TS’s. However, we propose a new method combining RS and T-DS that improves the performance for weak regret in DBP.

    Download PDF (493K)
  • Akane MINAMI, Yuki YOSHII, Yu KONO, Tatsuji TAKAHASHI
    Session ID: 1G2-GS-2a-03
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    The development of deep reinforcement learning has enabled learning of continuous state-action space, and the results have been remarkable in such a way enabling computers to surpass humans in playing digital and analog games. However, the problem that it requires a huge number of trials and errors has not been solved. In order to reduce the number of explorative action selections, we focus on an adaptive method called satisficing, which is in stark contrast with optimization. Satisficing leads to quick search for an action that satisfies a certain target level. Risk-sensitive Satisficing (RS) model that was defined based on satisficing in addition to “risk attitudes” based on the selection ratio of actions (representing the uncertainty of the value of actions). RS has been shown to be able to learn the optimal action sequence with a small number of exploration and finitely bound regret in the multi-armed bandit problems with when given some optimal target level. The linear RS (LinRS) is a linear approximation method for the RS, but the approximation for selection ratio of each action has not been sufficiently discussed. In this study, we propose StableLinRS, that is a new way to approximate the selection rate in LinRS. We also show the usefulness of StableLinRS in the contextual bandit problems in comparison with existing methods.

    Download PDF (596K)
  • Toshikatsu KATO, Yu KOHNO, Tatsuji TAKAHASHI
    Session ID: 1G2-GS-2a-04
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    We humans tend to search for a satisfiable action above an acceptability threshold (satisficing). A value function that implements satisficing together with the prospect theory-like risk attitudes called “risk-sensitive satisficing” (RS) model shows superior results in the bandit problems. However, wider application and analysis of the behavior of the model is intractable in some ways, because of the deterministic nature of the policy. In this study, we introduce the stochastic version of RS (SRS). Through comparison of RS and SRS in stationary and non-stationary environments, we show the merits of SRS.

    Download PDF (517K)
  • Takumi AKIBA, Tatsuji TAKAHASHI, Daisuke URAGAMI
    Session ID: 1G2-GS-2a-05
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    The goal of “social reinforcement learning” would be to realize effective learning by introducing the social nature of humans and organisms, into the framework of reinforcement learning. The social nature includes sharing information with others. The purpose of this study is to reveal the effect of sharing the maximum value of profit as a global aspiration and converting it to the aspiration value of each state in the framework of social reinforcement learning. The results show that the policy that combines the above two mechanisms is more adaptable to the two important factors of the number of agents and reward setting, compared to the policy that shares the action value and aspiration value of all states. It suggests that the sparseness of the information sharing and the resulting diversity in each agent contributes to the optimal performance.

    Download PDF (792K)
  • Takumi FUKUNAGA, Hiroyuki KASAI
    Session ID: 1G3-GS-2b-01
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    Optimal transport, which expresses the distance between probability distributions, have been applied to various applications. In optimal transport, it is necessary to solve a linear programming problem which has the tight mass conversation constraints, but it is known that optimal transport is difficult to solve fast. To improve this problem, relaxed optimal transport loosing the constraints has been proposed. It develop the fast methods and, moreover, some papers report that there are applications (color transfer, etc) which relaxed constraints are effective for. In this paper, we focus on the convex relaxed optimal transport, propose our new fast method and analyze it. Concretely, we propose the fast optimization method using Frank-Wolfe algorithm and prove the upper-bound of the worst convergence iterations. Finally, numerical evaluations show that our proposed method converges more fast than other existing methods.

    Download PDF (386K)
  • Zhongxi FANG, Jianming HUANG, Hiroyuki KASAI
    Session ID: 1G3-GS-2b-02
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    There are two crucial points in comparing graph structures. One is the representation of node feature vectors, and the other is the extraction of essential substructures. The representation of node feature vectors in graphs has been actively studied in graph representation learning to construct Graph Neural Networks (GNNs) that outperform the Weisfeiler-Lehman (WL) test. On the other hand, the latter is often used as a criterion for graph classification tasks, and comparison between key substructures is significant when comparing structures. However, the extraction of key structures is a difficult task and has not been sufficiently studied. In this paper, instead of extracting the key structures directly, we compare the structures that are likely to be important by giving weights to similar structures. Furthermore, we define a new distance between graphs. The results of the graph classification task using k-NN show that the proposed method outperforms the traditional distance methods.

    Download PDF (409K)
  • Shohei SEKIGUCHI, Emiko TSUTSUMI, Maomi UENO
    Session ID: 1G3-GS-2b-03
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    Knowledge Tracing (KT), using educational data to predict learners' knowledge states during the learning process, has attracted much attention. The most advanced KT method is Attentive Knowledge Tracing(AKT), which has been reported to show high prediction accuracy by incorporating a forgetting function of the past data to attention mechanisms. However, since AKT does not completely forget the past data, It causes non-negligible noises for estimating the past items weights. To slove the problem, we propose a new method to optimize the degree of forgetting the past data in AKT. In evaluation experiments, we compared the prediction accuracy of the proposed method with that of existing methods.

    Download PDF (393K)
  • Hinata EDO, Naoki HAMADA, Kazuto FUKUCHI, Jun SAKUMA, Youhei AKIMOTO
    Session ID: 1G3-GS-2b-04
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    The multi-objective evolutionary algorithm approximates the Pareto solution set by a finite number of solutions. In such an approach, as the number of objective functions increases, it is difficult to obtain the outline drawing of the Pareto solutions set. In this study, we propose a method to approximate the entire weak Pareto solution set by using a deep generative model. Focusing on the correspondence between the weight space of the Chebyshev scalarization approach and the set of weakly Pareto optimal solutions, we train a deep generative model that outputs the optimal solution of the Chebyshev scalarization function when a point on the standard unit is taken as the input and this is used as the weight vector. Experiments show that the proposed method obtains a more accurate Pareto solution set than some conventional methods when the number of objective functions is large.

    Download PDF (402K)
  • Naoki SAKAMOTO, Rei SATO, Kazuto FUKUCHI, Jun SAKUMA, Youhei AKIMOTO
    Session ID: 1G3-GS-2b-05
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    In constrained black-box optimization, optimizing the objective function is extremely difficult if the feasible domain X is a set of discrete feasible regions and even obtaining a feasible solution is difficult. This paper proposes a technique to transform the search space S into a simple one with almost no constraints. In detail, we create a map from the input space Z to X, Decoder G: Z -> X, and use Z of G as the search space to achieve the above transformation. To perform mapping to discrete regions, we make Decoder G concatenated small neural network models (NNs) with a shortcut connection, and we define loss functions for each NN. This prevents mode collapse, which is a well-known problem in deep generative models. In the experiments, we demonstrate the usefulness of the proposed technique using a test problem where the volume ratio of X to S is less than 1e-7.

    Download PDF (1203K)
  • Ruixing CAO, Takuma TANAKA, Akifumi OKUNO, Hidetoshi SHIMODAIRA
    Session ID: 1G4-GS-2c-01
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    k-nearest neighbour (k-NN) takes label average over a query ball, whose radius r<sub>k</sub> increases with larger k, and the non-zero radius results in a bias of the k-NN estimator. To reduce the bias, multiscale k-NN (MS-k-NN) first solves ordinary least squares (OLS) to predict the k-NN estimator at some points k=k<sub>1</sub>, k<sub>2</sub>, ..., k<sub>V</sub> from even-degree polynomials of the radius r<sub>k</sub>, and extrapolates the estimator to r=0. However, there remain two practical problems: (i) The polynomial used for extrapolation is derived from asymptotic theory; in finite-sample situations, the MS-k-NN estimator with even-degree polynomials is not necessarily restricted to a proper range [0,1]. (ii) OLS implicitly assumes the independence of the k-NN estimators at k=k<sub>1</sub>, k<sub>2</sub>, ..., k<sub>V</sub>, whereas the estimators utilizing some same labels are dependent. To solve these problems, we propose employing sigmoid-based functions and generalized least squares. We also propose local radial logistic regression (LRLR), which is inspired by MS-k-NN.

    Download PDF (427K)
  • Shoki YAMAKAWA, Takashi WASHIO
    Session ID: 1G4-GS-2c-02
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    Regression using pairwise comparison data between given instances but without using their objective values is called “uncoupled regression.” In this study, we propose its extension to Gaussian process regression. For example, when developing a new product specification, more preferred by customers, based on customer questionnaire results on many ready-made products, we do not expect that the customers can evaluate their preferences by setting numeric scores consistent across the customers. This is due to the variety of individual customers’ value scales. On the other hand, the collection of the evaluations mutually consistent among the customers is easier, if we apply pairwise comparison based questionnaires regarding the preference of products A or B. While a previous study has proposed an uncoupled regression method that enables point estimation of the objective value for each instance, we propose a Gaussian process based uncoupled regression method, which can evaluate the uncertainty of the regression model and its objective values. We show that the accuracy of the proposed method is practical in comparison with the supervised kernel ridge regression results through numerical experiments. This enables wider application of uncoupled regression, such as the design and development of efficient product specifications that take into account the uncertainty of estimated preference.

    Download PDF (594K)
  • Kazuki KOYAMA, Keisuke KIRITOSHI, Tomomi OKAWACHI, Tomonori IZUMITANI
    Session ID: 1G4-GS-2c-03
    Published: 2021
    Released: June 14, 2021
    CONFERENCE PROCEEDINGS FREE ACCESS

    For capturing non-linear dependency, feature selection methods using regularized learning and independence criterion have attracted much attention. In particular, HSIC Lasso is one of the most effective sparse non-linear feature selection methods based on the Hilbert-Schmidt independence criterion. However, the previous feature selection based on a single basis kernel function tends to ambiguous results, and moreover, relevant features may be missed in certain problem settings. In this study, we propose a method for multi-task learning using multiple basis kernel functions and a non-negative constrained Group Lasso with a group structure for each feature, which can clearly select useful features based on multiple independence measures. We applied the method to several synthetic datasets and real-world datasets and verified its effectiveness regarding redundancy, sparsity, and classification and prediction accuracy using the selected features. The results indicate that the method can more drastically remove irrelevant features, leaving only relevant features.

    Download PDF (397K)
feedback
Top