Generalization Error Bounds for Threshold Decision Lists

Martin Anthony; 5(Feb):189--217, 2004.


In this paper we consider the generalization accuracy of classification methods based on the iterative use of linear classifiers. The resulting classifiers, which we call threshold decision lists act as follows. Some points of the data set to be classified are given a particular classification according to a linear threshold function (or hyperplane). These are then removed from consideration, and the procedure is iterated until all points are classified. Geometrically, we can imagine that at each stage, points of the same classification are successively chopped off from the data set by a hyperplane. We analyse theoretically the generalization properties of data classification techniques that are based on the use of threshold decision lists and on the special subclass of multilevel threshold functions. We present bounds on the generalization error in a standard probabilistic learning framework. The primary focus in this paper is on obtaining generalization error bounds that depend on the levels of separation---or margins---achieved by the successive linear classifiers. We also improve and extend previously published theoretical bounds on the generalization ability of perceptron decision trees.