Vol.2 No.1 2009

11/88

Research paper : Predictive modeling of everyday behavior from large-scale data (Y. Motomura)−8−Synthesiology - English edition Vol.2 No.1 (2009) ChineseHome styleWestern foodSteak XY RestaurantContent DataCategoryBistro ABUser Data CB CurrySteak XY RestaurantAttitude Budget Measurement of preferred data Situation Data Suggestion Candidates/Attitude ValueAttitude Calculation Bayesian Network 100040002000Ai = Σj log p(cj = Cij )S1C1S2CjC2U1U2p(Cj)Cij603040123Ci1=Western food Ci2= $20“Reason for recommendation:In the mood to watch an inspirational movie? It’s okay once in a while…”History InformationPersonal ProfileContentDatabaseUserDatabasePerson ABayesian NetworkTheater Request RecommendationContent Recommendation SystemOne person (2) Compute recommended content(1) User inputs situation SuspenseHorrorHorrorRecommend(3)Recommended Contentand displayed reason for itTimeHave Plans Traffic Level SituationRestaurantCategoryFranchiseRestaurantHigh-classImpressionCustomerClassAverageBudgetRestaurant AttributesAge RangeDrivingHistoryDisposableIncomeCar type Individual AttributesMain Dish It then recommends items with a high score, limiting them to superior content. For 182 actual restaurants in the Shinagawa neighborhood, a questionnaire was conducted among 300 test subjects, causing desired store locations to be selected in six situations (scenarios). A model was constructed from the gathered data. Restaurants desired in six situations (scenarios) were selected from the 182 restaurants in the Shinagawa neighborhood. Concerning the selection procedure, firstly, the user was queried about desired categories, and stores corresponding to those categories were displayed. If disliked, the next genre was chosen by the same selection method as in currently existing car navigation systems. There were multiple answers for selected restaurants, and ultimately 3778 records were obtained. There were 12 situation attributes, 17 restaurant attributes, and 12 user attributes. The model in Fig. 3 was constructed as a result. There are four attribute nodes representing users, three representing situations, and six representing restaurants. The model consists of all 13 of these random variables, and the probability distribution of restaurant attributes favored by specific users in a given situation is calculated by probabilistic inference.In the model of Fig. 3, for drivers with a light driving history, the probability is high that franchise restaurants such as family restaurants and fast food chains will be chosen; conversely, for extensive driving histories, the probability that these restaurants will be chosen is low. Franchise restaurants often provide parking areas and show a tendency to be favored by young or beginning drivers. In addition to “driving history”, there is a “have plans” interaction. This reflects the tendency that even in cases wherein the driving history may be long, in situations when the driver has plans and must hurry, there is a high probability that a franchise restaurant will be used. The proper tendency is obtained intuitively for other relationships, such as that between budget level and vehicle type.Using the model depicted in Fig. 3, a prototype of a restaurant recommendation system was also designed (Fig. 4). Favored content attributes are forecast as probability distributions from user variables and situation variables.Ai = Σj log p( cj = Cij ) (3)By recommending content for which the value of this score is high, a car navigation system appropriate for the situation and the user can be implemented. Upon comparing this prototype system and a conventional car navigation system, its effectiveness was confirmed by the fact that prediction results for restaurants matched the users' preferences and the situation. 7.2 Information Recommendation Appropriate for User and Situation with a Portable PhoneInformation recommendation technology appropriate for various users and situations is important in next generation portable phone services. Examples of application of Bayesian networks in a movie recommendation service in portable phone services have been introduced [19][20]. For approximately 1600 test subjects, their content evaluation history, user and content attributes were collected via a questionnaire that suggested movie content. Other than demographic attributes such as age, gender, employment, etc., questions regarding lifestyle, appreciation frequency as attitude attributes concerning movie viewing, concern over movie selection time, the primary purpose for watching movies (seven questions about wanting to be emotionally moved), evaluation of content (good/bad), one’s mood at the Fig. 3 Restaurant preference Bayesian network model.Fig. 4 Outline of restaurant recommendation system [6].Fig. 5 Mobile information service system that recommends movies depending upon the user and situation [19][20].j=1n

※このページを正しく表示するにはFlashPlayer9以上が必要です