Vol.2 No.1 2009

9/88

Research paper : Predictive modeling of everyday behavior from large-scale data (Y. Motomura)−6−Synthesiology - English edition Vol.2 No.1 (2009) space in which the data arises and the bias are reflected in ways that resonate with human activity and semantics. In the space in which this type of data arises, peculiar restrictions and deviations in the frequencies of occurrence can be treated in terms of a probability distribution. Enumerating all the causal structures established in the world, as is the case for physical laws, is difficult because of the quantity of such descriptions; however, expressing the important elements as probabilities is an effective approximation. Work has been done to model this type of probabilistic construction by a Bayesian network in the actual space and to utilize it in Bayesian inference [15].In Bayesian inference, multiple class labels are taken to be Ci, and a posterior distribution combining both the likelihood P(x|Ci) for a signal pattern x and prior distribution P(Ci) which determines the class label Ci such thatP(Ci|x)=P(x|Ci)P(Ci) / ∑jP(x|Cj)P(Cj) (1) is maximized. It is known that this makes possible optimal recognition that minimizes the Bayes error probability. The fit of the data is represented by the likelihood, and prior knowledge is represented by the prior probability distribution. Learning from data and prior knowledge are naturally integrated by considering the maximized prior probability, which is the product of both of these, to be the inferred result. In cases when the frequency of occurrence of class labels depends upon the observation time and place, the prior distribution P(Ci) depends on the situation S. In such cases, consider this to be the conditional probability P(Ci|S), replace this with P(Ci) of equation (1), and obtain a class which maximizes the posterior probability of (2):P(Ci|x, S) = P(x|Ci)P(Ci|S) / jP(x|Cj) P(Cj|S) (2)The second term in the denominator of the right-hand side of Eq. (2) is the prior probability of the activity label Ci, which is in situation S in the label space. Here, we will consider the stochastic causal structure in the label space. When we construct, as a Bayesian network, a causal structure among a series of places and actions, for example, when we introduce advance knowledge with a causal structure of the following form: “If activity Cit occurred in circumstance S at time t, it is easy for activity Cit+1 to occur at time t+1,” the probability of an action when a person enters domain S is expressed as P(Cit+1 | CSt, S) and can be modeled by a Bayesian network. Having constructed the model by means of statistical studies on the data set of observations of activities when children are playing in an experimental environment imitating a living room, aside from past activities, dependencies between the relative distance of, for example, the sofa or wall in the room, the speed of movement, etc. are confirmed. Having studied the Bayesian network and naïve Bayes by means of activity data of other children, and inferred the activities of other children through Bayesian inference via Eq. (2), the identification rate was found to be approximately 50 % or less according to the most likely inference of naïve Bayes only, and could be increased to approximately 60 %~80 % by Bayesian inference, using a Bayesian network [14]. By means of this behavior inference algorithm, it is possible to efficiently form action-labeled data from observational images of everyday activities.6 Research as a ServiceNow that large-scale data can be measured in daily life, complicated problems can be handled through statistical learning. However, a characteristic problem of statistical learning is that as models become complicated at a high level, the amount of data necessary for learning increases. Sensor data observable superficially can be dealt with comparatively easily. However, the internal state of human behavior is a psychological aspect; therefore, a questionnaire survey used on test subjects is a necessity, and this entails a high cost. In addition, when acquiring data, practical problems exist, such as the problem of privacy and the fact that cooperation for the purpose of the research is simply difficult to obtain. Furthermore, even if a phenomenon is easily observable in terms of external factors, in order to completely collect predictor variables with high environmental dependence at the scene where they will actually be used, it is necessary that the environment wherein data is observed be controlled, so as to simulate the everyday environment as accurately as possible. Therefore, for this type of problem, the author considers it obligatory to unify actual service, investigation, and research. In this connection, the author lectures on the concept of “Research as a service” [22]. This clarifies the “means-end chain” as behavioral contingencies of humans in the context of behavior analysis, making it easy to make comprehensive models while including environmental dependence. Consequently, the results of the observations, evaluation questionnaires, and user feedback (psychological investigation), obtained while implementing the information service in society, are collected without separating the investigation and modeling procedure from the applications that use the model. This is known historically in cybernetics and in reliability engineering as the Deming cycle: PDCA (Plan, Do, Check, Action), in which a model is continuously corrected while cycling through actual problems.For an essential resolution of the uncertainty issue, an approach is necessary in which a cycle is permanently continued that collects additional data while using and controlling the model, with modeling based primarily on actual data. This is not limited to simply collecting actual data, but from the standpoint of research, implies that the researcher is imbedded in the field, which leads the way to new research that will bring about new values and evaluations [16].

※このページを正しく表示するにはFlashPlayer9以上が必要です