Table 5.

Activity recognition accuracies, using semi-supervised learning with a varying amount of real data and one month of synthetic data. * = difference from accuracy trained on real data is statistically significant (p < 0.05). The list of activities learned is mentioned in the introduction.

Number of Real Sensor Events used for Training	Accuracy When Trained on Real Data Only	Accuracy When Trained on Real Data Plus 200,000 Synthetically-Generated Sensor Events
5000	35.54% ± 5.10%	47.90% * ± 3.60%
10,000	45.50% ± 5.75%	61.19% * ± 4.96%
15,000	41.68% ± 5.49%	52.83% * ± 3.23%
20,000	47.61% ± 5.91%	57.57% ± 2.99%