New data modeling techniques with applications to two data mining problems

This dissertation explores two graph-based models for image mining and stream mining purposes. A multi-layer structured image mining framework is first introduced. From the bottom to top, the layers are pixel layer, segmentation layer, object layer, semantic layer and pattern layer. This dissertation proposes two independent algorithms for this image mining framework. An image segment algorithm Visual-hint Boundary to Segments (VHBS) is proposed at pixel and segmentation layers. VHBS abides by two visual hint rules based on human perceptions. Then a graph-based fuzzy linguistic metadata schema named Snowflake at object layer is introduced to describe the spatial relationships of the segments of images. VHBS partitions an image into segments and Snowflake generates a graph-based model to describe the spatial relationships between these segments. These two algorithms form the low-level of the image mining and the high-level of image mining such as content-based analysis is built based on the low-level layers.;The second part of this dissertation introduces a novel stream mining model for prediction purposes. Currently, this model named Prediction Intensity Interval model for Hurricanes (PIIH) is successfully used for hurricane intensity prediction. PIIH models a hurricane's life cycle as a sequence of states. States are discovered automatically from a set of historic hurricanes via clustering and the temporal relationship between states is learned as a dynamic Markov Chain. Novel techniques damped and weighted features are introduced in PIIH to improve the prediction accuracy, where damped features assume that different features use different normalization strategies to fit the distributions of feature values and weighted features assume that features do not contribute equally to the creation of states which result in superior prediction of intensity. The experiments demonstrate that PIIH provides compatible accurate prediction compared with the hurricane intensity prediction models used by National Hurricane Center currently.