Machine Learning
Lazy Learning of Bayesian Rules
Machine Learning
| Hi-index | 0.00 |
LBR is a highly accurate classification algorithm, which lazily constructs a single Bayesian rule for each test instance at classification time. However, its computational complexity of attribute-value pair selection is quadratic to the number of attributes. This fact incurs high computational costs, especially for datasets of high dimensionality. To solve the problem, this paper proposes an efficient algorithm LBR-Meta to construct lazy Bayesian rules in a heuristic way. It starts with the global classifier trained on the whole instance space. At each step, the attribute-value pair that best differentiates the performance of the current local classifier is selected and used to reduce the current subspace to a further smaller subspace for the next step. The selection strategy used has a linear computational complexity with respect to the number of attributes, in contrast to the quadratic complexity in LBR. Experimental results manifest that LBR-Meta has achieved comparable accuracy with LBR, but at a much lower computational cost.