Design Creativity

Triplet Embedding

With an effective way at describing the similarity between engineering desgin configurations, metrics like novelty and diversity--which at their core are similarity-based--can be formulaically described. Our work has proposed crowdsourcing humans to perform triplet comparison tasks between designs (Is Design A closer to Design B or Design C?). Unlike traditional methods for similarity evaluation, these triplet queries do not require identification or metric evaluation of shared features. With these labeled triplet queries serving as constraints on the acceptable placement of designs, a low-dimensional embedding can be constructed and similarity can be measured as a Euclidean distance between designs. Then relative diversity can be described by the spread of the embedded space, and novelty by the centroidal distances in that space.