Kazutoshi Sumiya

Discovery of user behavior patterns from geo-tagged micro-blogs

The explosive growth of micro-blogging sites such as Twitter has enabled folks to share their personal up-to-dates. Compared to conventional blog sites, through the short length of messages, micro-blogging sites help users easily express their experiences, thoughts and feelings and share them instantly and globally. In addition, mobile devices based micro-blogging applications are ensuring the usefulness in a variety of our daily activities without spatial or temporal restriction. Especially, the most significant characteristics chiefly possible in such mobile micro-blogging is on the fact that the cutting-edge smartphones can utilize location sensing information that make it clear to analyze where the published messages are made almost in real time. In the respect of the diversity and the quantity of crowds writing the micro-blogs, we are sure that the micro-blogging sites can be a very important social media platform where a lot of valuable knowledge such as geographic social phenomena can be extracted. In this paper, we endeavor to find geographic social patterns from user movement histories made by mass mobile micro-bloggers. We particularly propose fundamental models based on aggregation and dispersion about movements of micro-bloggers in geographic regions. We also performed experiments to discover geographic characteristics from the micro-blog data actually gathered from Twitter.

Urban area characterization based on semantics of crowd activities in Twitter

It is essential to characterize geographic regions in order to make various geographic decisions. These regions can be characterized from various perspectives such as the physical appearance of a town. In this paper, as a novel approach to characterize geographic regions, we focus on the daily lifestyle patterns of crowds via location-based social networking sites in urban areas. For this purpose, we propose a novel method to characterize urban areas using Twitter, the most representative microblogging site. In order to grasp images of a city by social network based crowds, we define the geographic regularity of the region using daily crowd activity patterns; for instance, the number of tweets, through the number of users, and the movement of the crowds. We also analyze the changing patterns of geographic regularity with time and categorize clustered urban types by tracking common patterns among the regions. Finally, we present examples of several urban types through the observation of experimentally extracted patterns of crowd behavior in actual urban areas.

Interoperable augmented web browsing for exploring virtual media in real space

Although location-based information systems can now be accessed from mobile devices, searching for information from an outdoor environment requires a significant amount of effort given the time-consuming manipulations that need to be made. Augmented-Reality (AR) based systems can help resolve these issues by reducing the physical effort required to input keywords onto a small keyboard, and the mental effort necessary to review the search results. But most AR systems do not consider media sharing environments where individual systems are not interoperable in respect of their media and browsing abilities. In this paper, we focus on a media environment for interoperable augmented-reality systems, making the assumption that such devices will become more wide spread and that many people will share individual media in a real-world space. Furthermore, we describe such emerging scenarios and new research issues. We also present a model to manage the futuristic data, making it similar to existing physical media, and show a practical example using the system we have developed.

Semantic ranking of lecture slides based on conceptual relationship and presentational structure

Abstract We describe a presentation content retrieval method involving the semantic ranking of target slides based on the relations between slides related to a user query. This method uses a keyword conceptual structure of the conceptual relationship implicitly existing between keywords extracted from the slide text of the presentation content and the presentational structure of indents in the slide text. At present, many presentation files are shared over the Web by many universities through their own public sites. Although these files are useful and valuable to many potential students, the fact that such files have to be retrieved for selflearning purposes means that there is still a lack of support for self-learners to find the desired slides on a priority basis, i.e, on the basis of importance and urgency of requirement of the desired file. Our noble semantic ranking method helps a user to easily learn through slides, focusing on either highly detailed slides or introductory slides in an order related to the user query. We also present a prototype system supported by our method for slide ranking and evaluate its effectiveness through experiments.

Detection of Unusually Crowded Places through Micro-Blogging Sites

Recently, micro-blogging sites such as Twitter have garnered a great deal of interests as an advanced form of blogging, where individuals can share their experiences, thoughts, feelings, etc, in real time. Additionally, mobile device based micro-blogging applications are now enabling the incorporation of extremely precise location information in the form of GPS-based coordinates. With the enormous number of micro-blogs being published all over the world, the resulting social media presents a novel dataset that can be used to survey our society on a global or local scale. In this paper, we propose an effective method for the detection of unusual crowding in physical locations. This method achieves the extraction of useful and interesting movement patterns reflecting the occurrence of critical events in a geographic region. In order to accomplish this, we analyze common patterns of occurrence in each region over a specified time period employing K-means based micro-blog clustering. Furthermore, we contrast unusual occurrence patterns with the movement patterns of micro-bloggers. Finally, we present an experimental evaluation of the proposed method using a real dataset collected from Twitter.

Layer-Based Media Integration for Mobile Mixed-Reality Applications

Rapidly evolving and widely-used smart phones provide many novel applications and services, making it possible to gather information from any location. Recent advances in technology have introduced many useful functions to assist users of outdoor mobile devices by sensing nearby conditions, such as the user’s current location and even the user’s slightest motions. Among these growing capabilities and their potential applications, the development of new mobile mixed-reality applications will need to consider various integration forms to be beneficial in both mobile device applications and services. In this paper, a layer-based media integration model for mobile mixed-reality applications is proposed to help developers gather diverse resources in a unified form, i.e., media, sensing, and even internal processing controls. In particular, to make it possible to identify and look for relevant information about real-world geospatial objects of interests in a very intuitive and direct manner by simply pointing towards an object, an object-identification layer is introduced. For each identified object, various types of services can be naturally integrated with the layer-based model; to take a snapshot or to jump to a related web page, or to collaborate with the popular touch control. The authors address an integration model that will significantly simplify mixed-reality application development work to establish a relationship between a geospatial object, its graphic image on a project screen, and user interaction. The improved simplicity, advantages, and new capabilities of the proposed model are also validated with two implemented applications, “Automatic Tagging Camera” and “Touch-based Mixed-Reality Web Search” based on the layer model.

Automatic Generation of Learning Channels by Using Semantic Relations among Lecture Slides and Recorded Videos for Self-Learning Systems

We present a method of automatically generating learning channels by using the semantic relations that implicitly exist in slides of a lecture that has accompanying recorded video. These days, many lecture videos with presentation files are shared over the Web from many universities through their own public sites. Although these materials are useful and valuable to many potential students, their use of sequential static media for self-learning purposes means there is still a lack of support for self-learners seeking learning channels suitable for various levels of understanding. Our newly generated learning channels let users easily focus on either highly detailed slides or introductory slides without needing to examine all of the data. We describe a prototype system supported by this learning-channel construction method.

A Scene Extracting Method Based on Structural and Semantic Analysis of Presentation Content Archives

It has become easier for users to view presentation content consisting of synchronized heterogeneous media such as videos and slides. Users simply retrieve the stored content they wish to view. In most cases users are interested in viewing a specific portion of the content. Scenes extracted from a presentation, however, often do not help the user understand the content, even when those scenes contain the query key word. A method is needed for extracting viewing intervals from content which are meaningful to the user. In this paper, we propose a method to extract scenes according to their relation. These relations are decided at the point of change between scenes. In this paper we describe our method and subject it to evaluation.

Scene extraction system for video clips using attached comment interval and pointing region

A method was developed to enable users of video sharing websites to easily retrieve video scenes relevant to their interests. The system analyzes both text and non-text aspects of a user’s comment and then retrieves and displays relevant scenes along with attached comments. The text analysis works in tandem with non-text features, namely, the selected area and temporal duration associated with user comments. In this way, our system supports a better-organized retrieval of scenes that have been commented on with a higher degree of relevancy than conventional methods, such as using matching keywords. We describe our method and the relation between the scenes and discuss a prototype system.

Geographical recommender system based on interaction between map operation and category selection

We propose a geographical information recommender system based on interaction between users map operation and category selection. The system has three interfaces, the layered category interface, the geographical object interface and the digital map interface. Our system interactively updates each interface based on the category interest model and the region interest model. This paper describes each interface and each model, and how to update them by our system.