Abstract: There are two kinds of accounts for category-based feature inductions: the similarity accounts and the knowledge accounts. The similarity account claims that feature inductions are based on overall similarities between premise categories and conclusion categories, and inductive strengths increase with overall similarities. The knowledge account emphasizes the role of knowledge of feature relevance in feature inductions regardless of the role of similarities. The similarity account adapts to inductive reasoning in knowledge-poor domains, whereas the knowledge account adapts to inductive reasoning in knowledge-rich domains. The two accounts are separate from each other, and have no relation with each other.
However, in many cases, people conduct inductive reasoning based on between-categories similarities on relevances of induction features. This reasoning simultaneously involves similarity and knowledge. In the present study, we proposed a relevance similarity model of feature induction to describe human inductive reasoning based on relevance similarities. The model integrates similarity with knowledge of feature relevance, claims that size orders of inductive strengths are determined by products of relevance similarities times relevance strengths between induction features and relevance features of induction features, and size orders of inductive confidences (subjective confidences for judgment of inductive strengths) are mainly determined by relevance strengths. Thus, the model dissociates inductive confidences and inductive strengths. According to the model, feature inductions based on relevance similarities include the two kinds of strong and weak relevance inductions, which are respectively based on strong and weak relevances between induction features and relevance features. In strong relevance inductions, induction features have a few definite relevant features, thus feature inductions are based on similarities on relevant features, and have strong relevance strengths and strong inductive confidences. When there are no features which are especially relevant to an induction feature, the whole of a premise category’s all features will become the weak relevance of the induction feature, and feature inductions will be based on the overall similarity across categories. This is the weak relevance induction which has weak relevance strength and weak inductive confidence. Therefore, feature induction based on overall similarities is a special case of the relevance similarity model.
The model makes the following predictions. (1) When relevance strengths remain constant, inductive strengths will increase with relevance similarities, whereas inductive confidences will remain constant. (2) When relevance similarities remain constant, inductive strengths and confidences both will increase with relevance strengths. (3) When relevance strengths and relevance similarities vary simultaneously, and vary in the opposite directions, size orders of inductive strengths will be determined by products of relevance strengths and relevance similarities. Size orders of inductive confidences will be mainly determined by relevance strengths. (4) Regardless of relevance similarities, strong relevance inductions have higher inductive confidences than weak relevance inductions do. (5) There will be dissociation between inductive confidences and inductive strengths, because inductive strengths are determined by products of relevance strengths times relevance similarities, whereas inductive confidences are mainly determined by relevance strengths. The results of two experiments with college students supported the above main predictions. However, in weak relevance inductions, the results showed that inductive confidences were also affected by overall similarities, and increased with overall similarities. This finding shows that overall similarities form a base of weak relevance inductions, and inductive strengths and inductive confidences all increase with overall similarities. But on the whole, strong relevance inductions have higher inductive confidences than weak relevance inductions do, regardless of relevance similarities.
In summary, the relevance similarity model can describe and explain main inductive reasoning based on relevance similarities, and has a greater explanatory power and scope than the previous relevant theories do.

Key words: feature induction, the relevance similarity model, relevance, similarity, inductive strength, inductive confidence