Vol.2 No.1 2009

12/88

Research paper : Predictive modeling of everyday behavior from large-scale data (Y. Motomura)−9−Synthesiology - English edition Vol.2 No.1 (2009) For use in customer Analysis Marketing ResearchBayesian Network Model Construction, InferenceHorizontal Developmentbased on statistical learningAutomatic construction of modelUser modelshowing tastesLarge-scale databaseFeedback dataMeasure user preferences,recommend optimal contentOperation HistoryVarious situations, usersWeb, Mobile, Navigation System, etcMovie, Shopping,RestaurantRecommendationApplication ServiceReusable Cognitive Model Implementation of ServiceComfortable /Uncomfortable New Services Design Object is… therefore…Data observationduring actual serviceIn need of /Want to buy Service Offerings Laughter /Anger Situation is…therefore… Data integrationSafe /DangerousBehavior Analysisthrough services Place is…therefore…Sensor integrationLike / Dislike Action is…therefore…Statistical learningof Human behavior model Situation dependenceEveryday Activities(Actual Field)FunHuman Cognitive Evaluation Structuretime (seven questions about being emotionally moved) were collected. Furthermore, for approximately 1000 people, all of the following were collected separately as free-form text: the content of each movie, what kind of feeling or situation, (theater, DVD, etc.), with who, with how many people, what time of day, was the movie appreciated. This data was input into BayoNet [17][18], the Bayesian network construction software developed by the author, and a Bayesian network model was constructed automatically. Through the Bayesian network constructed in this way, a prototype of the portable information system was developed that makes movie recommendations, based on situation and user tastes. If the user sends requests to services from the portable phone, together with information about the situation, the system implements the probability inference using registered user attribute and situation information from the database. Content whose probability of being selected is judged to be high is recommended as superior (Fig. 5). This movie recommendation system was also developed into an Internet service at auOne lab (http://labs.auone.jp) and released generally in 2007 with approximately 7000 recommendations implemented. Further, the model is being restudied from this recommendation history, and experiments are being conducted to improve recommendation precision. Using the calculation model for movie selection constructed in this way, we also proposed cooperation with a movie distribution company to optimize sales strategies for DVD content for which the movie release period has passed [21].As this information service spreads and multiple users utilize the system, the history of selection content accumulates ever-larger amounts of statistical data. Improvements in the Bayesian network model resulting from that data will increase the appropriateness and inference precision of the model, create a self-supporting feedback loop, and allow horizontal development of other services to be realized. Data obtained from the market through actual services becomes reusable knowledge for the calculation model; this knowledge cycle, reflected in the next service, can be called “Research as a Service” as noted earlier (Fig. 6). Research activities through this type of actual service can even be put into practice in a service engineering research center through construction of a calculation model from large-scale data and through implementation of an optimization design loop in the field. Such research activities are proposed as a business to improve the productivity of the service industry [22].8 ConclusionIn the present research, the development of software could be categorized as pure or basic research; however, software development which excludes the initial step, which is obviously outcome oriented, could be considered applied fundamental research. It seems that there were several conditions that implied that we rethink the criterion for application selection intuitively recommended in that process.1. There are unresolved problems in existing procedures.2. There were problems actualized by user requests.3. There are stakeholders that profit from resolutions of these problems and bear the corresponding cost and risk.In these types of conditions, using Bayesian networks to model human behavior, forecasting customer or user activities, and achieving improvements in value and efficiency by optimization of associated services is thought to be an appropriate outcome. Client enterprises that can realize these outcomes exist in industry types that possess contact points (channels) with various customers. Selecting the outcomes mentioned above, the appropriate fields become channels that can collect large amounts of data from customers such as the Internet, portable telephones, car navigation systems, and call centers. However, among these choices, two necessary types are: being able to Fig. 6 Knowledge cycle service due to Bayesian network.Fig. 7 Reusable model of human cognitive and evaluative structure.

※このページを正しく表示するにはFlashPlayer9以上が必要です