Ordinal regression for preference learning in wearables using sensor data

Wearable technology makes it increasingly common to obtain successive measurements of a variable that changes over time. A key challenge in various fields is understanding the relationship between a time-dependent variable and a scalar response. In this context, we focus on active frames equipped with electrochromic lenses, currently in development. These lenses allow users to adjust the tint at will, choosing from four different levels of darkness. Our goal to train an ordinal regression model to predict the preferred tint level using ambient light data collected by an Ambient Light Sensor (ALS) to deploy it on electrochromic frames, which would control the tint in a personalized way. We approach this as an ordinal regression problem with a functional predictor. To tackle the complexities of the task, we use an adaptation of the classical ordinal model to include functional covariates. We explore the use of wavelets and B-splines functional basis, as well as regularization techniques such as lasso or roughness penalty. In two datasets with data issued from wearable devices, these functional ordinal regression models outperform LSTM and FCN networks, being up to 10% more accurate while remaining interpretable.