| A recognized group of large temporal databases consist of a collection of sequential data. Some of these databases have some unique characteristics, which ask for a radically different approach for mining useful knowledge. The huge size and uncertainty embedded in these collections make it difficult and impractical to mine the whole database at once.; This research proposes a domain independent novel technique to mine temporal databases through layered phases. We introduce a novel method to mine sequential data by layered phases for prediction and forecasting in both short term and long term. The concept of phase and hierarchy plays a fundamental role in many real world application domains. We apply such a point of view to mine sequential data. To construct model of a pool of sequence we project the model to a layer of phases interacting with each other. The whole model will be built gradually in a top-down fashion and each level might be investigated for more details. This technique consists of tow major steps: phase recognition and model construction. We provide a novel iterative phase recognition technique to detect major phases in data. In addition, we explore phase transition for a layered model construction as an extension to Hidden Markov Model (HMM).; In the second part of this dissertation, we introduce and analyze a special but powerful and popular case of Layered HMM as Self-Similar Layered HMM, for a certain group of complex problems, which show self-similar property, and exploit this property to reduce the complexity of model construction. In addition, phase recognition through monitoring fractal dimension will be introduced and analyzed. The last chapter of this thesis is devoted to exploiting self-similarity to discovering interesting rules by scaling the previously recognized sequential association rules.; Mining by layered phases has investigated in the context of four different domains: Synthetic data which we used for test and tuning the algorithms; Critical Care, a large databases of patients in critical care condition; Network database and World Wide Web. We illustrate data mining by layered phases has several advantages in terms of accuracy in predictions, interpretability, optimization and speed. |
