Vol.2 No.1 2009

10/88

Research paper : Predictive modeling of everyday behavior from large-scale data (Y. Motomura)−7−Synthesiology - English edition Vol.2 No.1 (2009) It is also necessary to implement an applied system that can be embedded as infrastructure within society, as a product that can be tolerated as an actual service.7 Bayesian Network Applied SystemThe applied system of the Bayesian network can be developed by implementing a probabilistic inference algorithm or model construction algorithm as a computer program. Building on work in the Real World Computing Project, the IPA Unexplored Software Project, and other such projects prior to 2001, and by searching Bayesian networks from large amounts of data in 2002, the author developed the software BayoNet that can perform probabilistic inference based on this work [17][18]. This software has been licensed to private enterprise and commercialized; however, due to the fact that a high degree of specialized knowledge is necessary in order to apply it to the resolution of particular problems and the fact that the utilization procedure is not self-explanatory, it has been somewhat difficult to train users who can fully utilize the software. If it is software developed for highly specified purposes, it is not necessary; however, software featuring a Bayesian network emerging as purely fundamental research on mathematical models can be applied for extremely broad purposes; and at the point in time when it can be utilized in practice, new investigations resulting in even more valuable purposes can occur. Therefore, a taskforce of the strategic center for venture development was started in 2003, and researchers personally had the opportunity to begin a search for business models that use this technology. At this point in time, many research results have remained as essential technologies, such as algorithm refinement and acceleration or inference precision improvements; however, we have felt resistance to further development of technologies in circumstances in which the outcomes were not obvious. Therefore, we decided to prioritize the search for outcomes by problem resolutions that had the possibility of being adequately treated, given the efficiency at the time.The advantage of using Bayesian networks is that by performing probabilistic inferences, we can determine the probability distribution of arbitrary variables and conduct quantitative evaluations in various situations. In many conventional multivariate analysis procedures, quantitative relationships are often modeled based on a covariance relationship that assumes linearity among variables (linear independence). In the Bayesian network model, quantitative relationships are represented by a conditional probability table. In a conditional probability table, a family of conditional probability distributions are not hypothesized, but rather, the table forms a model in which non-linear, non-normal relation interactions can be represented with great freedom. In addition, predictor variables and objective variables are not clearly distinguished; therefore, introduction of latent [implicit] variables is also straightforward. In other words, even variables that cannot be observed can be treated as latent variables. Therefore, latent variables are introduced that become categories, and when analyzing the statistical data of a user or customer, we can extract attributes of groups that perform the same activities, classify constituencies, and that can even be utilized in customer segmentation.It is extremely important that these characteristics respond by recommending information or products that are acceptable matches, depending upon the user or customer activities (Web browsing history, etc.), attributes, or circumstances. In collaborative filtering, information or products desired by customers or users cannot reflect situation dependence when displayed by a portable telephone or car navigation system. Information recommendation technology for such activities that change depending upon the environment is important, even in ubiquitous computing, in which a variety of situation changes in actual space are imagined.7.1 User- and Situation-Dependent Information Recommendation in a Car Navigation SystemIt sometimes happens that the driver of a car wants to stop somewhere while driving. For example, while driving for some purpose, the driver decides to stop to eat at a restaurant. In conventional car navigation systems, a category is specified, and all corresponding restaurants are listed in order by distance. The user must find the appropriate restaurant from within the list; however, the user has to operate a touch switch or remote control in order to see detailed information about restaurants, so it is not easy for the driver to locate the desired restaurant.Therefore, if a car navigation system were to model the driver’s preference of various restaurants, given various situations and criteria using a Bayesian network, and, using a probabilistic inference from this, if the system replaced the driver while driving, and automatically selected the appropriate destination, it would be an extremely practical function. A person’s taste depends largely upon their personality and upon the situation while driving. While driving, it is necessary to select the most appropriate choice at the time, among conditions that change moment by moment.To illustrate this dependence on situations and personal differences, a Bayesian network can be efficiently applied that can model complex relationships among variables and uncertainty. Therefore, we test and evaluate a car navigation system that suggests content appropriate for the user [6]. This system possesses, as a Bayesian network, a user taste model within the vehicle information system. Content, such as restaurants or music, is suggested by content providers, and a score showing how appropriate it is for the user and conditions at the time is calculated as a conditional probability when the situation and user attributes are given.

※このページを正しく表示するにはFlashPlayer9以上が必要です