Modulating Users' Involvement in Interactive Machine Learning Solutions: A Model Cascade Strategy

Adapting intelligent systems to the end-user goals and their desire for involvement is essential when designing trustworthy interactive solutions. In intelligent environments, where sensitive information must be preserved, the challenge becomes two-fold: i) approaching the critical personal data to the user to promote privacy (i.e., Edge Computing); and ii) adaptatively modulating users' participation throughout the time. For this reason, this work proposes an interactive approach based on a cascade of Machine Learning models that makes optimized decisions related to classifying individual data and labelling it. For the evaluated use-case of a Human Activity Recognition system, the initial quantitative results of the proposed strategy show that an interactive cascade of simpler models can improve the non-interactive approach used as a benchmark and, at the same time, modulate the degree of participation of the user, measured as the number of times they would be inquired to provide a new label for newly obtained data. Thus, this paper provides insights into how this approach may be used in designing intelligent systems to adapt to the role of users in the personalization of intelligent models and how to build flexible experiences and learning systems where the user feels involved. All this while maintaining the privacy requirements that apply to Edge Intelligence and Edge Computing concepts.