4.7 Instance-Based Learning 131 The multiplier needs to be greater than 1, and the w i are set to a constant at the start. The algorithm we have described doesn't allow for negative weights, which-- depending on the domain--can be a drawback. However, there is a version, called Balanced Winnow, which does allow them. This version maintains two weight vectors, one for each class. An instance is classified as belonging to class 1 if ( w 0 + - w 0 - ) a 0 + ( w 1 + - w 1 - ) a 1 + ...+ ( w k + - w k - ) a k > Figure 4.11(b) shows the balanced algorithm. Winnow is very effective in homing in on the relevant features in a dataset; therefore, it is called an attribute-efficient learner. That means that it may be a good candidate algorithm if a dataset has many (binary) features and most of them are irrelevant. Both Winnow and the perceptron algorithm can be used in an online setting in which new instances arrive continuously, because they can incrementally update their hypotheses as new instances arrive. 4.7 INSTANCE-BASEDLEARNING