Efficient Parallel Classification Using Dimensional Aggregates

Multidimensional aggregates are frequently computed to improve query performance in Online Analytical Processing applications. We present a new method for decision tree based classification trees using the aggregates computed in the multidimensional data model. The structure imposed on data in a explicit multidimensional storage mechanism leads to efficient dimensional operations. Decision tree based classification algorithms perform computations to find the best split point at each node of the tree. Efficient computation of the split in the decision tree can be done by using the one-dimensional aggregates if the cell values are the class-id values, and counts are maintained for each class. This is used repeatedly at the nodes of the decision tree to calculate splits and manage data. Previous parallel approaches for decision-tree based classification use sorted attribute lists and hash tables to compute the split point and split the data appropriately. The amount of data communicated is proportional to the product of number of records in the training set, and the number of dimensions, at each level of the tree, in the worst case. Parallel formulation of our approach uses data communication proportional to the product of the sum of cardinality of all dimensions and the number of non-classified nodes at each level of the tree. Communication volume is greatly reduced in our approach and is done in one phase of communication at each level of the tree, by coalescing messages. Preliminary results from our experiments on a coarse-grained, distributed memory parallel machine (IBM-SP2) show good performance.