Feature Selection for Case-Based Learning: A Cognitive Bias Approach

Experimental research in psychology, psycholinguistics, and cognitive science has discovered and examined numerous psychological constraints on human information processing. The tendency to concentrate on (a) present, rather than missing cues, (b) the current focus of attention, and (c) recent information are three examples. Short term memory limitations provide a fourth constraint. This paper shows that psychological constraints such as these can be used effectively as domain-independent sources of bias to improve learning. We first show that cognitive biases can be automatically and explicitly encoded into a baseline training instance representation. We then investigate the related problems of cognitive bias interaction and cognitive bias selection, and compare two selection methods that make varying assumptions about the independence of individual component biases. Finally, the paper shows that performance of a nearest-neighbor case-based learning algorithm on a natural language task improves as more cognitive biases are explicitly encoded into the baseline instance representation.