Feature selection: the effect of cluster separation, attribute variability, and presentation format upon visual pattern recognition

Abstract

The present research series was designed to investigate variables hypothesized to affect feature selection and to determine the effect of these experimental manipulations upon pattern classification, pattern discrimination, and pattern recognition. Among the variables examined in the three-experiment series were (1) the distance between the class centroids of two clusters to be discriminated, (2) the method and duration of presentation of members of the pattern pairs, (3) the physical proximity of pattern components comprising stimulus units, and (4) the type and amount of variability introduced about the prototypical segments of tri-component patterns. For all three studies, the fifteen-element patterns used as stimuli were produced by the VARGUS 9 computer program (Evans & Mueller, 1966). During Experiments I and II, these stimuli were presented as contiguous polygons. Experiment III, however, required that stimuli be mapped into tri-component patterns and presented as contiguous polygons or as segmented stimulus units. These tri-component patterns, moreover, were constructed to have five levels of variability about the pattern thirds. Sixty-four Ss, randomly assigned to treatments, served in each of the first two experiments and eighty Ss served in the third experiment. During Phase 1 of each experiment, Ss made classification judgments and similarity assessments to successively presented stimulus pairs. During Phase 2, Ss were required to recognize members of familiar pattern classes from a set of unfamiliar pattern classes. In all treatments, stimuli were presented for comparison on slides via an automated Carousel projector system. For all experiments, three-way and four-way analyses of variance and additional a posteriori tests were performed on selected dependent measures. With regard to Experiments I and II, the physical distance between class centroids (cluster distance) was found to have its most pronounced effect during the Phase 2 pattern recognition task. In contrast, stimulus presentation and level of pattern variability were associated with significant interactions only during pattern acquisition (Phase 1). Results of Experiment III revealed that, while contiguous patterns were classified more accurately than segmented patterns, segmented patterns were subsequently recognized more accurately. In addition, both the type of pattern component variability and the location of component variability within a pattern affected schematic classification and pattern recognition. Results are discussed in terms of alternative strategies used by Ss in perceptual processing. Evidence suggests that, although Ss tend to weight the left side of the pattern most heavily, this strategy is modified by the spatial separation of pattern components. In addition, the present findings indicate that pattern processing during pattern classification is less extensive than that required for subsequent pattern recognition.