And we use the true positive rate (T PR), the false positive
rate (FPR), and Accuracy as follows:
Accuracy := (T P + T N)/(T P + FN + FP + T N) (2)
T PR := T P/(FN + T P) (3)
FPR := FP/(FP + T N) (4)
Table. V shows the event detection accuracy results for each
of the 12 subjects. The smallest witnessed Accuracy is 81.7%
and the overall Accuracy is 86.6%.
To further explore the accuracy w.r.t different FPR, we construct
the receiver operating characteristic curve (ROC) [35],
which illustrates the performance of a binary classifier system
as its discrimination threshold is varied. ROC is created by
plotting the T PR against the FPR, as shown in Fig. 7.
The red curve is constructed from the empirical data.
However, this curve may be imprecise to describe the receiver
operating characteristics, due to inadequate data or uneven
data distribution. Therefore, we fit the smooth estimation
(the green curve in Fig. 7) out of the empirical data by the
binormal model which assumes that the detection results
follow two independent Gaussian distributions. Basically, the
ROC exhibits the accuracy of event detection with different
levels of tolerance to false. For example, given a false positive
threshold of 0.1, the T PR achieves around 80%. It can be
computed that the area under the curve (AUC) is around 0.9,
which indicates that our approach can accurately detect