In order to revitalize the shopping street, it is necessary to urge customers to visit each store in the shopping street and to know the attractiveness of each store. However, regarding the current situation, there is a problem that the visits of most customers tend to concentrate in certain stores such as big chain stores. Therefore, in our study, we actually implemented an event using city walk application that we developed. Also, in the event, in order to prevent customers from visiting such stores, we created an incentive design to determine the point allocation of each store, with the aim of stabilizing the unique index of the research called “store appeal value”. And then, by analyzing the log data of the event, we were able to confirm the visit to each store in the shopping street and show the effectiveness of the proposed method.

In natural scenes such as mountains, to realize precise aerial shootings, large scale terrain modeling with multi- drone is proceeded. The SfM (Structure from Motion) is applied for this modeling based on the pictures produced by the automatic flying multi-drone. Using the terrain modeling, the procedure of precise automatic aerial shootings is proposed.

Many travel surveys have been carried out in many tourist destinations to investigate tourist's needs. However, traditional survey methods can be time and money consuming. Meanwhile, the analysis of the massive and up-to-date travel reviews may be able to provide a low-cost and real-time substitute. There are many tasks before we can develop a method to perform the investigation automatically. This paper will explain the progresses we have made through manual analysis and show the remaining tasks for developing an automated method.

Estimating the number of people traveling by route is an important task in transportation research. We propose a new method that utilizes the number of people staying and passing to estimate the number of people traveling by route with high accuracy. By using artificial data, we demonstrate that the error rate of the proposed method is smaller than the conventional method. In addition, the analysis of read-world data are shown.

The development of information science in recent years has greatly contributed to studies of animal behavior. We are collaborating with Maruyama Zoo in Sapporo to reduce the load on zookeepers taking care of animals using artificial intelligence. One of our targets is to automatically create chimpanzee’s Ethogram for health management and maintenance of breeding environments. Ethogram is an inventory of all behavior patterns of specific individuals and species. In order to create an Ethogram, it is necessary to identify individuals. Therefore, in this research, we examined whether each individual chimpanzee can be recognized using Convolutional Neural Networks ,which has high accuracy in image recognition field. Experimental results showed that it is possible for our system to identify individual chimpanzees.

The creation of art using deep learning has been paid attention to in recent years. Also, there is a haiku as an art that has long been popular in Japan. In this research, we demonstrate the usefulness of deep learning as art creation by making haiku from motifs, which is a general method of creating haikus, using deep learning. First, we train LSTM based on a large amount of past haiku, let it generate a stringa. Second, we extract the ones that satisfy the condition as a haiku from the generated character string and calculate the evaluation value as to whether it fits the motif image or not. If the evaluation value is high, it is assumed that the generated haiku matches the motif image. In this process, we conducted an experiment to confirm whether LSTM was able to learn rules as a haiku.

The overfishing and depletion of marine resources including tuna have become problems in Japan. Managing to marine resource is necessary to increase catch amount. However, set-net is difficult to separate fish speceis. Therefore, this research used a sonar image obtained by an echo sounder installed in a set-net. We verify of fish species estimation model based on echo sounder image.

In the activities of people on WEB such as SNS, the burst phenomenon is observed. Recently, Hawkes Process is used as a method to analyze the burst phenomenon. It is known that when the branching ratio, which is an index representing the internal dynamics of Hawkes Process, exceeds a certain threshold, the event time series transits from steady state to nonstationary state where the burst phenomena is likely to occur. In this paper, we focus on social tagging system among data obtained from SNS, and analyze how branching ratio changes timewise with service growth.

With its huge spread in daily life, social media is now a big platform to make new human relationships. While recent research identified the relationship between users' personality traits and how individual users behave on social media, knowledge about how personality traits affect users' mutual interactions is limited. This research proposes a method to relate human personality to community formation on social media. By applying this method to Twitter data, it was inferred that personality factors such as gregariousness, conscientiousness, cheerfulness shows a "homophilous" tendency. This means that similarity in these personality traits breeds user connection. Furthermore, we investigated the relationship between community's composition of personality and community dynamics(size, activeness). Analysis by the proposed method may contribute to creating a social matching system with better user satisfaction.

This study aims to construct a system to support evaluating credibility of fake news. We proposed a credibility evaluating model using dynamic relationship network which dynamically determines credibility by relative consistency. This credibility evaluation model consists of a conclusion node representing the truth and false of the article and a fact node having information of 5W1H. For evaluation of the crudities model, we evaluated the network using non-fake news and fake news. As a result, it is possible to identify unreliable information from the fact node, and the conclusion node. It was also known found that the trend of the article can be inferred.

Anonymous online communications more easy to self-disclosure than offline communications. However, Japanese young people's online social relationships tend to include offline social relationships. Therefore, it is important to complement offline social relationships by communication tools which have few offline relationships. For this complementing, we discuss online social support in an avatar chat service "Pigg Party." We marshal user-interview results of application developers of this service from a viewpoint of online social support.

We investigate a method that assigns Basic Subject Headings (BSH) 4th edition to the results of web people searches to help users select and understand people on the web. By assigning BSH headings to people, well-formed keywords can be assigned. In this paper, we examine the following combination of factors: (a) web-page rank, (b) position inside HTML, (c) synonyms, and (d) document frequency. We report our results of experiment using an 80-person dataset.

Sharing application window among users may improve cooperative works. However, configurations of sharing multiple application windows among multiple PCs are complicated. We implemented a management system for the configuration. The main idea of the system is to introduce an analogy of classrooms, which helps users to understand and make the configurations. The system provides a novel interface for management of groups in a classroom. This paper shows the implementation and evaluation of the system. Specifically, we realized a practical application window sharing system by optimizing streaming of windows.

PMF (Probabilistic Matrix Factorization) is a well-known approach of recommending systems. It has achieved wide use and excellent performance not only in research but also in industry. In recent years, this algorithm has been combined with some side information, and the accuracy of recommendation has been improved. In this paper, we propose a novel probabilistic model using image shape features that integrates convolutional neural network (CNN) into probabilistic matrix factorization (PMF).

The Cold-Start is one of the problems in web marketing. For example, when the system recommend items to users on an e-commerce site, a purchase log and a review log are handled. However, since web marketing data tends to be long tail, log data of most users and items are few to learn. In recommender system, the item is presented to the user based on this past log, thus long tail items are hardly presented than popular items. In this research, we propose the method combining semi-supervised learning and singular value decomposition against this Cold-Start problem. In addtion, we report the result of verifying our proposed method with the user rating score of the movie provided by MovieLens.

This paper aims to extend the context search engine so that it can retrieve both of open and private data. The context search engine has been developed for answering trend-related queries. While it is designed to store and retrieve publicity available data (open data), it is supposed that users often have data that can't be disclosed to the outside due to problems such as privacy protection. Making it possible to access both of such local data and open data in integrated manner would lead to effective utilization of data such as value creation from data. In order to examine the utility value of local data without sharing its contents, this paper introduces the concept of Data Jacket into the context search engine. Experimental results show its effectiveness.

In recent years, it has become increasingly important for many companies to utilize data obtained from consumers in their marketing. While various companies exchange and utilize marketing data with each other, the approach to extract co-creative value is still in the process of improvement. In this research, by using IMDJ (Innovators Marketplace on Data Jackets), we analyzed expectations, needs, and problems for data exchange, based on the characteristics of marketing data and players. We also proposed incentive design for marketing data exchange with other companies.

In this paper, we try to develop a method for extracting causal expressions from news articles. First, we investigated clue expressions that are included in news articles. Next, we focused on specific clue expressions and developed the method to extract causal expressions.

The potential expectations for discovering problems and solving them by combining data and knowledge in different domains have been increasing. However, there are many social barriers to its realization. Data is originally the property that expresses the observed phenomena or events in the world in symbols and character strings, which is Data 1.0 in this paper. Afterward, Data 2.0 has been developed by the privilege of personal devices or Internet of Things, which is the era of linked data stored in different domains. While the benefits of data 2.0 are immeasurable, the technologies and discussions to combine only data are reaching the limits. In this paper, we propose "Data 3.0" as a new stage in the cross-discipline data utilization and discuss the Data Landscape in the era of Data 3.0.

Demand for sensing data trading market (SDTM) is increasing as becoming more important of application developers who use sensing data gathered from various IoT devices. For them, metadata of sensing data is crucial information to analyze it especially with AI technology. This paper discusses metadata definition, generation and utilizing technology with implementation on prototype.

Team behaviors are regarded here as an essential component of cultures and lifestyles of people engaged in businesses, sports, etc. In order to detect team behaviors, pedestrians’ similarity of the acceleration time-series is proposed as the temporal evaluation index. Here, a redevelopment area is taken as the space for experiments with acceleration sensors, and the correspondence of acceleration similarity and the pedestrians’ mutual feeling of co-walking is evaluated.

Remote sensing using synthetic aperture radar (SAR) images has attracted attention as a method of disaster area detection. However, there is a problem that a lot of experts and time are required for wide-area SAR image interpretation. Therefore, in this research, we propose a method that automatically detects landslide disaster areas using convolutional neural network (CNN). The proposed method uses not only the SAR images after disaster occurs, but also the images before the disaster and altitude data (DEM). In the experiments, the accuracy of classification as disaster area and non-disaster area in the testing areas was 75.56%, and intersection over union (IoU) was 21.86% that showed the ratio of the areas classified as disaster to the actual disaster areas. From these results, it was clarified that the landslide disaster areas could be classified by CNN considering the features of SAR images and DEM data before and after the disaster.

Although artificial neural networks (ANN) is widely used for real-time flood prediction model, it is pointed out that the weak point of the model is poor applicability for the inexperienced magnitude of flood. In this study, the ANN models were applied to Abashiri River catchment. The training period of the ANN models were 1998-2015. The validation data was the 2016's largest flood since the river-stage observation had started. The main component of the model was the four-layer feed-forward network. As a network training method, the deep learning based on the denoising autoencoder was applied. The river-stage prediction up to 6 hours showed very good accuracy, and proved it can nicely predict the such a large flood.

The covered route search problem for graphs with deadline time has applications such as aerial photographing route search problem at natural disaster. However, information on the deadline time of a node is not necessarily fully known in advance, and it is not realistic to calculate the optimum route in advance. In this paper, we propose a method to calculate a route as efficient as possible while giving priority to nodes with efficient covering under these conditions. The feature of this method is to prevent extreme deterioration of performance against dynamic change of information by using heuristics that makes the remaining nodes as one unit as possible. As a result of experiment using 100 patterns graph randomly allocated urgency, this method was able to reduce flight time of average 36.52% compared with simple method . And, in many graphs, the proposed method reduces the case that does not meet the deadline.

Providing stable and high-quality service is a critical issue for mobile network service providers. However, due to an unexpectedly huge amount of data traffic exceeding network capacity of a provider, a mobile network service experiences severe failures such as network troubles, performance deterioration, and slow throughput. Then, the service users often detect service outages before the service provider detects them. They can immediately publish their impressions on the service through social media and search for failure information on the web. In this paper, we propose a machine learning approach that incorporates multiple user behavior data into detecting and forecasting failure events. The approach is based on novel feature extraction methods and a model ensemble method that combines outputs of supervised and unsupervised learning models from multiple user behavior datasets. We demonstrate the effectiveness of the approach by extensive experiments with real-world failure events.

This paper demonstrated a significance of policy decision making with multiple stepwise citizen participation programs with introducing a case of revising the basic environmental plan Sapporo. Sapporo city implemented participatory programs involving both applicants who were willing to join the discussion and residents who were randomly chosen. The participation programs were designed from the beginning stage to final stage of the planning. Back Cast Scenario Workshops (BCSW) were conducted in the participation program, in which participants begin with the discussion mapping out an ideal future and consider backward to present situation. We discussed the different role of participants in each step of the participatory program. Furthermore, we stated the advantages and points to note as for installing BCSW. Finally, we mentioned the limitations of our approach and remarked expectations to AI.

Conducting workshop for consensus building is restricted by human, time and spatial. So, it is difficult for many citizens to participate. We have developed a consensus building support system that supports discussion. However, since it is difficult to conduct a large-scale social experiment exceeding 500 people, the support system on that scale has not been verified. Therefore, in this study, we conduct a large-scale social experiment using support system for Place Branding in Nagoya city and examine effect and problem. Through this verification, the discussion using the support system proved effective for satisfying the participants.

We aim to develop autonomous facilitator agents as a novel technology for large-scale consensus building. In this paper, we consider a model for understanding discussion context that is needed for implementing autonomous facilitator agents. Our consideration consists of two viewpoints: (1) content and (2) process. (1) Content: Facilitator agents need to understand structure of argumentation (e.g., issues, ideas, evidences, and participants' interests). (2) Process: they also need to grasp the change of atmosphere among discussion participants (e.g., dynamics of group interaction and group sentiment). We formulate a model of discussion context from these viewpoints and discuss how to utilize the model for determining facilitators' actions and utterances.

In this paper, we identify a new class of distributional constraints defined as a union of symmetric M-convex sets, which can represent a variety of real-life constraints in two-sided matching settings. Since M-convexity is not closed under union, a union of symmetric M-convex sets does not belong to this well-behaved class of constraints in general. Thus, developing a fair and strategyproof mechanism that can handle this class is challenging. We present a novel mechanism called Quota Reduction Deferred Acceptance (QRDA), which repeatedly applies the standard DA mechanism by sequentially reducing artificially introduced maximum quotas. We show that QRDA is fair and strategyproof when handling a union of symmetric M-convex sets. Furthermore, in comparison to a baseline mechanism called Artificial Cap Deferred Acceptance (ACDA), QRDA always obtains a weakly better matching for students, and, experimentally, performs better in terms of nonwastefulness.

It is widely known that herd behavior, a cognitive bias in humans, causes irrational or inappropriate behaviors in evacuation situations. In this paper, we propose the evacuation decision model for herding based on the biological response threshold model. We also show that the proposed model can be applied for multi task problems in which the agents must choose two or more tasks to perform such as evacuation and fire fighting.

This report describes the conversion from a logic programming language PROLEG to a Bipolar Argumentation Framework(BAF),aiming at a support for judicial judgment using Artificial Intelligence. Although reasoning process of a judgment is clealy shown in PROLEG, it is difficult to understand the relationship between inference rules and the entire structure of a judgment. Here, we give a semantics to a BAF and convert the descrition in PROLEG to BAF, so that the semantics of PROLEG is reserved. As a result, it is easier to understand the relationship between inference rules and the entire structure of a judgment in the obtained BAF which consists of a set of arguments and relations of attack and support between arguments.

This report proposes a qualitative spatial representation focused on the change of shapes, and describes temporal spatial reasoning using this representation. It aims at an application to the developmental biology by symbolically representing a process of development such as cell division and organ formation. In the process, we can find some crucial shape-changes such as bending, getting a hole, division and so on. We propose a language to represent such a characteristics and show the conditions to find possible state transitions from a given state. It enables qualitative simulation and backward reasoning for an observation of an unusual situation.

In the domain of tourism navigation, tourist spots as well as route visiting them are important information to be presented to tourists. However, it is difficult to determine tourist spots to visit and the route all at once. This is a challenging problem. While an existing study has tried to solve such a tourism navigation problem by extending Traveling Salesman Problem (TSP) with introduction of weight for nodes (spots), its formulation became complicated. This paper proposes more simple formulation of the tourism navigation problem, which assigns all factors needed for a solution to only edges. The solution using Simulated Annealing method is also proposed.

The number of companies using subscription business model is increasing, and churn prediction is getting a more important task. In existing research, various type of machine learning models have already been used, but churn prediction has to be trained by combining various data such as time series data and non-time series data, which has not been fully studied. On the other hand, the technique of deep learning is still being developed, and one of its characteristics is that it can learn various data and models from end-to-end. In this research, we propose a churn prediction model with deep learning using data of WealthNavi inc. which manages the service of Robo-adviser. Specifically, we propose a method to learn time series data and non-time series data with one model. In the experiment, the effectiveness of this method was demonstrated by obtaining the result exceeding the accuracy of the classifier of the existing research.

In this paper, we tackled the recommendation of the M\&A candidate considering the change in business performance after M\&A. By incorporating the multitask learning framework into the Neural Collaborative Filtering method which is one of recommendation method using Deep Learning, we aimed to propose recommendation method considering the post-conversion change. Experimental results show the similar accuracy as the simple logistic regression method. By using this method, it will be possible to not only recommend M\&A targets but also to show to acquirers what kind of benefits they can obtain by acquiring.

Mathematical model to analyze search action on the Internet is presented on the sense of sociophysics. As an extension of the mathematical model for hit phenomena, in the present model, we assume that the search action of ordinary people in society is affected not only by mass media campaign, direct communications and indirect communications, but also by Twitter and blog posting. The present theory can reproduce the search action very well and it seems to work better than the previous mathematical model for hit phenomena.

Demand for souvenirs that can be purchased only at the particular location or area is increasing, because anyone can purchase various souvenirs on online shopping sites. Such souvenirs are called local limited souvenirs in this paper. However, there are no Web sites and services that collected information about local limited souvenirs. The purpose of this study is to construct a system that presents useful information about local limited souvenirs, in order to provide support for selecting souvenirs. This paper proposes a method for extracting names of souvenirs and shops from blog articles using CRF (Conditional Random Fields). The effectiveness of the proposed method was confirmed by evaluation experiments.

We created a corpus for prefectural assembly minutes. Each record in the corpus has a field of "speaker's name." This field is manually checked to improve credibility of the corpus. The corpus also has attributes of speakers (namely, electoral district, birth year and gender) in the case of assembly members. We try to analyze the corpus and illustrate activities of all assembly members in Japan. This paper describes: 1) the number of assembly members in Japan, 2) age and gender composition of assembly members, 3) difference in the amount of speech depending on age and gender, 4) difference in the contents of speech depending on gender, 5) difference in the contents of speech depending on age.