The Integration of Data Streams with Probabilities and Relational Database using Bayesian Networks

Abstract

As sensor devices develop, not only the amount of uncertain sensor data streams is dramatically increasing, but also the streams are processed in a variety of ways. We believe one of important ways is to reason contexts from them, and the integration of dynamic reasoning result and static data in databases. This paper proposes the integration of probabilistic data streams and relational database by using Bayesian networks which is one of the most useful techniques for reasoning uncertain contexts in the physical world. And this paper has three concrete contributions. For the first contribution, we model the Bayesian networks as an abstract data type in the object relational database. Bayesian networks are stored as objects, and we define new operator to integrate Bayesian networks and relational database. Since Bayesian networks has the graphical model, it does not directly fit relational database that is constituted of relations. Our new operators allows to extract a part of data from Bayesian networks in the form of relations. For the second contribution, to allow continuous queries over data streams generated from the Bayesian networks, our proposed method introduces a new concept, lifetime, into the Bayesian networks. Although the Bayesian networks is a famous reasoning method, it is not yet treated in data stream systems. The lifespan allows a Bayesian networks to detect multiple events for each evaluation of a continuous query. For the third contribution, we proposed efficient methods for probability values propagations. The methods omits unnecessary update propagations for continuous queries. The result of experiments clearly showed that our proposed algorithm outperforms usual algorithms.