User login

Weekly Report -- 20/07/2012

23

Jul

2012

Added a couple more time series to my event detection testing data, adding a whole new set of misclassifications to try and resolve. On Friday I noticed a large number of false positives were cases where the sample standard deviation only just crossed the standard deviation threshold. However, this relationship was logarithmic rather than linear which is why I didn't notice it earlier.

Adding a new rule (pseudocode):

if (sd / log(sd_limit, 2)) < SD_LOG_LIMIT
then
NOT AN EVENT
fi

has drastically reduced the number of false positives in my dataset. I've lost a few genuine events as well, so it is not perfect, and I need to determine what the correct value for SD_LOG_LIMIT really is.

My last test run produced 258 correctly identified events, 42 false negatives and 48 false positives (out of a total of 975,375 measurements).

Started teaching 301 - first lecture went fairly well with students being attentive and asking questions, which is good.

Groups: