The CMG’12 paper “A Note on Knee Detection“ (J. Ferrandiz, A. Gilgur) presented a method of “system phase change” detection by using the piece‐wise linear model against data with any supply‐demand relationship, e.g. CPU vs. transactions, load vs. traffic.
The weakness of the approach is the following underlying assumption in the methodology: the most of the data points are in the low load region. But all in all that is a relatively simple and effective way to capture the fact of constantly exceeding some threshold (confidence level e.g. 95%) of the data beyond detecting “knee”.
I see some similarity in my method of detecting the system phase changes (trends detection implemented by SETDS).
Based on my paper CMG'08 “Exception based Modeling and Forecasting” I use EV (Exception Value) meta-metric to detect pattern change in the data. The phases in the data should be separated by roots of the EV = 0 equation because for EV > 0 the data mostly exceeds the upper control limit and for EV < 0 – it is mostly below low control limit, and data is stable where EV=0.
But I used my way only for time series data (EV= f(t)), so detected phases are separated by points in time. Is that possible to apply my EV based approach to non-time series data? Not sure. But knowing that EV is just the deference between actual data and control limits (e.g. percentile based), the above mentioned knee detection algorithm could be a some kind of EV- based approach applied to non-time series data…
And I believe EV based approach is free from the assumption that the data should be somewhat misbalanced and it can also detect multiple “knees” with both directions (up and down). Only one caviar still exists (the same with knee detection algorithm): too many phases detecting, but that could be tuned by limits change, e.g. from 95 percentile to 99 and by grouping data (like aggregating minutes to hours for time-series data).
The weakness of the approach is the following underlying assumption in the methodology: the most of the data points are in the low load region. But all in all that is a relatively simple and effective way to capture the fact of constantly exceeding some threshold (confidence level e.g. 95%) of the data beyond detecting “knee”.
I see some similarity in my method of detecting the system phase changes (trends detection implemented by SETDS).
Based on my paper CMG'08 “Exception based Modeling and Forecasting” I use EV (Exception Value) meta-metric to detect pattern change in the data. The phases in the data should be separated by roots of the EV = 0 equation because for EV > 0 the data mostly exceeds the upper control limit and for EV < 0 – it is mostly below low control limit, and data is stable where EV=0.
But I used my way only for time series data (EV= f(t)), so detected phases are separated by points in time. Is that possible to apply my EV based approach to non-time series data? Not sure. But knowing that EV is just the deference between actual data and control limits (e.g. percentile based), the above mentioned knee detection algorithm could be a some kind of EV- based approach applied to non-time series data…
And I believe EV based approach is free from the assumption that the data should be somewhat misbalanced and it can also detect multiple “knees” with both directions (up and down). Only one caviar still exists (the same with knee detection algorithm): too many phases detecting, but that could be tuned by limits change, e.g. from 95 percentile to 99 and by grouping data (like aggregating minutes to hours for time-series data).
No comments:
Post a Comment