PyChange - Change detection with Python

Quickstart

Example1: Differences in two sequences.

python3 PyChange.py --filename=random.csv --cell=B --values=A --time=T --method=CUSUM --preprocessing=diff

Computes the changes in trend in two sequences stored in random.csv with the MaChaMP method. One of the sequences has a significant trend change, the other does not.

Example2: Comparsion of methods.

python3 Example.py

Creates Example.pdf, where all available methods are applied to the same step function.

Use as a script

Store your time series in a .csv file and run PyChange.

--filename: Name of .csv File (e.g. from standard QtFy output).
--cell: Name of unique time series identification column.
--values: Name of column of time-series values
--time: Name of column of timepoints.

--method: Type of change detection method.

• CUSUM: Pages Cummulative sum control chart. (h=8.01, k=0.25).    
• EWMA: Expoentially weighted moving average control chart. (burn-in period 30, lambda = 0.05, L=3.7).
• QChart: For start up processes and short runs (3o3, maxlike=2.).

--preprocessing: Transformation of sequence before applying change detection.

• none: default Raw sequence, for change in mean.  
• diff: Change detection on gradient sequence, to remove linear dependencies 
• logdiff: To remove multiplicative dependencies in diff.
• percdiff: Percentile differences for changes in intensities.
• logpercdiff: To remove multiplicative dependiencies of percdiff.

If --filename=Name.csv, the output is a ChangesName.csv file. The columns are: CellID, Changepoint location, and timepoint of change.

Use as a module

import numpy as np
from PyChange import PyChange

r = np.random.RandomState(42)
seq = list(r.randn(200)) + list(r.randn(200) + 1.) + list(r.randn(200)) + list(r.randn(200) + 1.)

cp = PyChange(seq)
print cp

The function PyChange has three input variables:

• seq: required The time seires as a list.
• transform: the preprocessing as above.
• method: the method as above.  

Remarks

Be carful:  

• r and the desired package needs to be installed.
• E-Divise is slow.
• The Bayesian methods return posterior probabilities for each point being a changepoint.
• All methdos other are set to the settings suggested in repective the documentation.

Any questions?

Come to my office BSB 2.01.
Or email: [email protected]