Group technology must group similar parts into families. In classifying parts based on their global shapes, the similarity of parts has to be manually measured by performing pair comparison. The cost of exhaustively performing pair comparison is quite high when the number of parts to be grouped is large. This paper proposes interval intersection, a novel similarity inference method that effectively infers the pair-comparison data from a set of known data. Justified by empirical experiments, the proposed method outperforms the previous methods when 31% or more of data is known.