Queries for Bias Testing

Selecting a good bias prior to concept learning can be difficult. Therefore, dynamic bias adjustment is becoming increasingly popular. Current dynamic bias adjustment systems, however, are limited in their ability to identify erroneous assumptions about the relationship between the bias and the target concept. Without proper diagnosis, it is difficult to identify and then remedy faulty assumptions. We have developed an approach that makes these assumptions (e.g., about the irrelevance of features or feature values) explicit, actively tests them with queries to an oracle, and adjusts the bias based on the test results. 1 Introduction Bias is a fundamental aspect of any supervised concept learner. Numerous papers have noted this importance (e.g., Mitchell 1980; Haussler 1988). The type of bias that we discuss here is the choice of a hypothesis language. The hypothesis language defines the space of hypotheses. A strong bias defines a small hypothesis space; a weak bias defines a large h..