Experiments on UCR time series datasets with LPStimeSeries package

I started experimenting on UCR time series database with LPStimeSeries package (version 1.0). The aim of this experimentation is to see the sensitivity of error rates, training and testing times to the parameter levels.

As discussed earlier, we proposed a randomized version of LPS which selects the segment length randomly for each tree between certain thresholds. That left us two parameters to tune: the number of trees and the depth of the trees. Our submission (to PAMI) discusses that LPS is not affected by these parameters if they are set in certain range. In order to check if this is true, we conduct a new experiment. Here are the experiment settings:

1- We train 200 trees for each dataset.

2- Segment length is selected randomly for each tree between 0.1 and 0.9 (as the factor of the time series length).

3- The depth of the trees is set to 3,4,...,10

4- For each depth setting, we replicate the runs for 25 times

An Ubuntu 13.10 laptop with i7-3540M Processor (4M Cache 3.00GHz) and 16 GB DDR3 memory is used for the experiments. The experiments are still running, this page will be updated regularly. For each depth setting, the results are shown on a boxplot (25 replications). The aim is to show the robustness of LPS to depth paramater. LPS generates accurate results if the depth is set in a reasonable range (between 6-8). Here are the results for each dataset (in the same order as in our submission to PAMI):

 

 

 

Copyright © 2014 mustafa gokce baydogan

LinkedIn
Twitter
last.fm