This package contains several methods on conducting multivariate DTW-NN.
====== Content of Folders ===============
Data: the pickled datasets for processing. Results: the folder to store the produced results. Source: the source code folder.
====== List of Methods ==================
LB_MV_ws: min and max are precomputed on references (ws stands for with setup). LB_MV: online version. No pre-setup.
LB_TI_ws: LBMV followed by online LB_TI with pre-setup. LB_TI: online version. No pre-setup.
LB_PC_ws: point clustering-based method with pre-setup. LB_PC: point clustering-based method. No pre-setup.
===== How to run ========================
Example 1:
Step 1) Go to Source folder.
Step 2) Run test_onesetting.py 2
It will invoke all the methods on the second dataset in 'Data/' folder, produce the performance measurement results, and put them into the folder 'Results'.
Step 3) Run peakResults.py
It will generate the following files in folder ../Results/tryonesetting/ 0X0_All_speedups.txt: contains the speedups from each of the methods. 0X0_All_skips.txt: contains the number of skips from each of the methods.
Example 2:
Step 1) Go to Source folder.
Step 2) Run test_tryall.py 2
It will invoke all the methods on the second dataset in 'Data/' folder and try every setting for a method, produce the performance measurement results, and put them into the folder 'Results'.
Step 3) Run peakResults.py
It will generate the following files in folder ../Results/tryallsettings/
0X0_All_speedups.txt: contains the speedups from each of the methods.
0X0_All_skips.txt: contains the number of skips from each of the methods.
Which includes the best results of each method for each datasest.
===== Notes =============================
-
In the real deployment, parameters of TI and PC methods would need to be selected through a separate process by experimenting with the candidate series. Similarly, for a given dataset, one can use a small set of samples to pick the better choice between LM_TI and LM_PC; this adaptive scheme leads to a method named TC-DTW.
-
The included datasets were part of the UCR multivariate datasets: http://www.timeseriesclassification.com/index.php More can be downloaded from that website. Picked files are used by this package.
-
If an error like module 'Source' was not found appears when running the commands, just add the root directory of this package (i.e., the path of NewMDTW) to your PYTHONPATH environment variable.