In this paper, we propose a method for quickly finding a given number
of instances of a target class from a fixed data set. We assume that
we have a noisy query consisting of both useful and useless features
(e.g., keywords). Our method finds target instances and trains a
classifier simultaneously in a greedy strategy: it selects an instance
most likely to be of the target class, manually label it, and add it
to the training set to retrain the classifier, which is used for
selecting the next item. In order to quickly inactivate useless query
features, our method compares discriminative power of features, and if
a feature is inferior to any other feature, the weight 0 is assigned
to the inferior one. The weight is 1 otherwise. The greedy strategy
explained above has a problem of bias: the classifier is biased toward
target instances found earlier, and deteriorate after running out of
similar target instances. To avoid it, when we run out of items that
have the superior features, we re-activate the inactivated inferior
features. By this mechanism, our method adaptively shifts to new
regions in the data space. Our experiment shows that our binary and
adaptive feature weighting method outperforms existing methods.
/ Fax: +81(Japan) 75-753-5978 / Office: Research Bldg. #7, room 404