Classification using P-tree regarding yeast gene deletion data

Genomics data has many properties that make it different from "typical" relational data. The presence of multi-valued attributes as well as the large number of null values led us to a P-tree-based bit-vector representation in which matching 1- values were counted to evaluate similarity between genes. Quantitative information such as the number of interactions was also included in the classifier. Interaction information allowed us to extend the known properties of one protein with information on its interacting neighbors. Different feature attributes were weighted independently. Relevance of different attributes was systematically evaluated through optimization of weights using a genetic algorithm. The AROC value for the classified list was used as the fitness function for the genetic algorithm.