Integral vs. Separable Attributes in Spatial Similarity Assessments

Computational similarity assessments over spatial objects are typically decomposed into similarity comparisons of geometric and nongeometric attribute values. Psychological findings have suggested that different types of aggregation functions—for the conversions from the attributes’ similarity values to the objects’ similarity values—should be used depending on whether the attributes are separable (which reflects perceptual independence) or whether they are integral (which reflects such dependencies among the attributes as typically captured in geometric similarity measures). Current computational spatial similarity methods have ignored the potential impact of such differences, however, treating all attributes and their values homogeneously. Through a comprehensive simulation of spatial similarity queries the impact of psychologically compliant vs. deviant methods have been studied, comparing the top-10 items of the compliant and deviant ranked lists. We found that only for objects with very small numbers of attributes—no more than two or three attributes for the objects—the explicit recognition of integral attributes is negligible, but the differences between compliant and deviant methods become progressively worse as the percentage of integral attributes increases and the number of groups in which these integral attributes are distributed decreases.