Vol.2 No.1 2009

7/88

Research paper : Predictive modeling of everyday behavior from large-scale data (Y. Motomura)−4−Synthesiology - English edition Vol.2 No.1 (2009) (i) for each node, limit the candidates that can become new nodes, (ii) select a child node, add and graph the new candidate nodes one by one, (iii) decide on and evaluate the parameters on which the graph is based, (iv) only when evaluated highly, use as a new node, (v) when there are no more new node candidates to add, or when the evaluation does not increase even if a new candidate is added, move to another child node, (vi) repeat (i) – (v) for all child nodes.In general, new search spaces increase combinatorically; therefore, a device is needed to avoid an increase in computational load by limiting combinations of new nodes that become ranked from the beginning to be candidates. Furthermore, we consider independently the search portion of the graph (ii), (v), and the mode evaluation portion (iii) and think about various study methods.One can expect that the use of a Bayesian network would be an effective approach to construct a non-deterministic model from large amounts of data by means of statistical learning. However, obtaining a causal structure from only statistical data is fundamentally difficult, and the task of searching the graph structure is NP hard. In such cases, it is actually necessary to skillfully implement the variable candidates or search range limitation, or to introduce appropriate latent variables.3.4 Probabilistic InferenceThere are other types of models that possess a graph structure; however, many of them are often used to visualize graph structures that explain data. On the other hand, by constructing them with discrete random variables and conditional probability tables, Bayesian networks very efficiently implement the probabilistic inference algorithm, which estimates the probability distribution of arbitrary random variables in the model. This is a significant advantage over other graphical models and is a crucial feature in operating an intelligent learning system with a realistic computational load.Probabilistic inference on Bayesian networks is implemented by the following procedure: i) assign the value of an observed variable (e) to a node, ii) assign a prior probability distribution to both new nodes and nodes having no observed value, iii) calculate the posterior probability distribution P(X|e) of the desired object value (X).In order to find the posterior probability in item ii), a probability propagation method is implemented which renews the probability distribution of each variable according to the dependence between variables.When all paths within a graph structure that does not consider the direction of links of the Bayesian network do not possess loops, the Bayesian network is called a singly connected network. In this case, even networks with structures in which multiple new nodes and child nodes exist, the calculation is completed by utilizing its conditionally independent character by performing, for each node, probability propagation calculations of 4 types: propagation upstream, downstream, from upstream, and from downstream (Figure 2).These computations are completed in order of network size (number of links), and the calculation efficiency is extremely high. When the network is viewed without considering the direction of links and there is a portion in which even one path is somehow looping, this Bayesian network is called multiply connected. In this case, there is no guarantee of an exact solution; however, probability propagation can be applied as an approximate solution method; this is called the Loopy BP method.4 User ModelingBased on the fact that the information system and the user advance processing conversationally, and that this information system is a portion of the entire system, which is subject to operations, the information system, the user, and even the environment and surrounding circumstances, must be considered. Consequently, viewing the entire system as a control object, human behavior and response should be thought of as one component of the object of calculation. In this case, we want to evaluate what the user is requesting under given circumstances and how to react to the system output results obtained. In the system, it is necessary to describe the cognitive state of such users in terms of computable user models.Machines (programs) learn from data by the development of machine learning. In other words, the approach wherein models are constructed from data and revisions are made sequentially [iteratively] is feasible. In the construction of machine learning models, statistical tests are repeatedly implemented to act as an automated model selection process, based on information criteria; and the result is considered to be the appropriate model. In other words, a statistically meaningful model is chosen through machine learning from within an extensive search space.Input to XOutput from XInput to XOutput from XY1YjUiU1Xπ( )Xλ ( )YXXλ ( )XUuπ ( )XYXπ ( )UXuλ( )X.)()|()()(,)()()(,)()(,)()|()().()()Pr(ikkUkXikxXUijkYkXXYjYjYjXUiUiXuuUxPxuxxxxxUiuUXPxxxxXFig. 2 Belief propagation.

※このページを正しく表示するにはFlashPlayer9以上が必要です