Exploring the Effectiveness of Interactive Preference Learning for Adapting Designs to Abstract Semantic Attributes

Abstract

Abstract. Abstract semantic attributes of designs (e.g., comfortable, luxurious, and durable) play a significant role in the assessment of user-facing products, capturing intangible factors that people may consider aside from performance requirements. However, due to the difficulty of mapping highly subjective and varying perceptions to specific design features, it remains a challenge to quickly and accurately translate these qualities into designs using computational design tools. Seeking to align computational and human representations of subjective design information, we investigate the utility of adapting representations of semantic attributes to designers’ perceptions through interactive models. A study is conducted in which users evaluate parameterized drinking mugs, indicating their perceptions of how comfortable each is to hold. Interactive Bayesian optimization is used to adaptively arrive at a design that optimizes this subjective quantity for each participant individually. Participants (N = 31) guide the model by providing their own decisions or building off of empirical data from a prior group of participants (N = 25). The resulting designs are evaluated across different scenarios, demonstrating the extent to which outputs of noninteractive models can be used to represent a subjective, semantic attribute and how interactive models may improve perceived alignment between human intent and computionally generated outputs.