Similarity refers to the psychological nearness or proximity of two mental representations. Research in cognitive psychology has taken a number of approaches to the concept of similarity. Each of them is related to a particular set of assumptions about knowledge representation.
Mental distance approaches (Shepard 1962) assume that mental representations can be conceptualized as some kind of mental space. Concepts are represented as points within the space. Similarity between concepts is a function of the distance between the concepts in space. Concepts represented by points that are near to each other are more psychologically similar than are points that are conceptually distant. A strength of this approach is there are many mathematical techniques for deriving spaces from data such as multidimensional scaling (Shepard 1962) and latent semantic analysis (Landauer & Dumais 1997).
Featural approaches (Tversky 1977) were developed to address limitations of the mental distance approaches. For example, spaces are symmetric. The distance between two points is the same regardless of which point you start from. However, psychological similarity is not symmetric. For example, we often prefer to state similarity in one direction. For example, it feels more natural to say that 101 is like 100 than to say that 100 is like 101. Furthermore, many metaphors are also directional. Saying "That surgeon is a butcher" means something quite different from saying "That butcher is a surgeon."
Featural approaches assumed that people represent concepts by lists of features that describe properties of the items. A similarity comparison involves comparing the feature lists that represent the concepts. Features that are shared in the feature lists are commonalities of the pair and features that are contained in one feature set but not the other are differences of the pair. It is possible to account for people's intuitions or ratings of the similarities between concepts by assuming that judgments of similarity increase with the number of commonalities (weighted by the salience of those commonalities) and decreases with the number of differences (weighted by the salience of the differences).
Structural approaches to similarity (Gentner & Markman 1997) were developed to address limitations of the featural account. In particular, featural approaches assume that the commonalities and differences are independent of each other. However, commonalities and differences are not psychologically independent. In fact, determining the differences between a pair requires finding the commonalities. Consider the comparison between a car and a motorcycle. Both have wheels. That is a commonality. However, cars have four wheels, while motorcycles have two wheels. That is a difference. Because this difference required first finding a commonality between the pair, it is called an alignable difference. Alignable differences contrast with nonalignable differences which are aspects of one concept that have no correspondence in the other. For example, cars have seatbelts and motorcycles do not. Research suggests that alignable differences have a larger impact on people's judgments of similarity than do nonalignable differences. Thus, the relationship between the commonalities of a pair and the differences is important for understanding people's assessments of similarity. Structural approaches to similarity emerged from research on analogy.