Leveraging Interactive Preference Learning to Guide Participatory Design Sessions

Abstract

Participatory design is valuable in early product development to quickly gauge sentiments towards product prototypes by directly integrating end-users in the design process. However, it can be difficult for some users to naturally contribute when presented with a set of complicated tools to manipulate candidate designs. Therefore, we examined the use of active preference learning as assistance for design space exploration in a sample use case (Head-Up-Display-Design in cars) to quickly estimate the overall preferences of participants by sequentially presenting key design variants for evaluation from a continuous parameter space. The goal of the study was to determine whether we are able to learn valid, meaningful preferences in short one-off design sessions. Hence, we compared the quality of the predictions and evaluated the subjective experience of the interaction when employing models based on active learning and Bayesian optimization against a random search baseline. Results indicate that active design assistance, compared to random search variants, is able to significantly better predict latent preferences (Cohen's kappa: 0.63 vs 0.12) and are perceived as significantly more engaging and rewarding (user engagement score: 3.68 vs 3.32). This highlights the potential of algorithmic design space exploration as an alternative interaction strategy for participatory design sessions.