Running a bunch of data on a low power system with limited disc space is an exercise in balance. When I first started building the machine learning component of SwitchBoard, I had made the assumption that four time categories (dawn, morning, afternoon, evening) would be sufficient to find some usefulness. And this was true. But, in an effort to make it more precise (thus, more useful), I’ve decided to experiment with upping the time categories to six (midnight, dawn, morning, afternoon, evening, night). Since behaviors in your home are tied closely to time - you likely don’t have your kitchen light on as much at 3am than you would at 5pm - I’m going to try and see if the predictions are more accurate.