When Unique Is a Challenge — Opening Human Activity Recognition for Design Opportunities

Interactive systems rely on clear models of interaction, where actions and feedback can be mapped and designed. Human Activity Recognition (HAR) offers ways to identify users’ activities, but it is less commonly used in interaction design. Studies have shown that activities can be recognised in ideal, lab-like settings; however, individual variation challenges real-world classification. Consequently, HAR remains difficult to use in design due to unreliability. We collected a variable-sensor-placement dataset of HAR activities. Machine learning classifiers were trained and evaluated, and visualisations were examined through five semi-structured interviews. The visualisations exposed activity groupings, overlap, temporal progression, and ambiguity, but their interpretation depended on contextual and technical information. Contextual pitting showed that unclear activities could become more reliable when evaluated against relevant alternatives. These results suggest that HAR can support interaction design when activity data, ambiguity, contextual constraints, and interpretation of visualisations are treated as design material.