The generalization challenge

Binary sensors report events such as a door opening or a motion detector firing. These signals can support activity recognition, but data collected in different homes often varies in layout, sensor placement, and format.

That variation makes it important to evaluate models beyond a single dataset or environment. My PhD work brought 13 public sensor datasets into a standardized machine-learning and evaluation pipeline.

Representing events as text

I explored a semantic feature approach: sensor-event sequences were transcribed into text blocks using templates, then embedded with language models. The resulting representations could be compared with more traditional features in activity-recognition tasks.

The evaluation compared seven transcription strategies and three embedding models, using cross-validation and metrics designed to account for class imbalance.

What the results suggest

In specific cross-domain evaluations, the textual representation approach improved recognition by 30%. This is a result for the evaluated settings, not a universal guarantee; the value of the representation depends on the data and task.

The broader question is how to build activity-recognition systems that remain useful when sensors, environments, and data sources change. Reproducible evaluation across datasets is one step toward answering it.