Download

Codes (MATLAB scripts) to replicate the experiments in this blog post are available in the zip file. 

Download

This file is outdated. Please check the help of my R package @ http://cran.r-project.org/web/packages/LPStimeSeries/index.html

R code to classify time series. The example dataset is GunPoint from UCR time series database. 

Download

This file is outdated. Please use the one @ http://cran.r-project.org/web/packages/LPStimeSeries/index.html

R package for Learned Pattern Similarity (LPS). Only tested in Linux environement. 

Install it from the local file by 

'R CMD INSTALL yourfoldergoeshere/LPS_1.0.tar.gz'

The documentation still needs some work. After finishing it, the package will be submitted to The Comprehensive R Archive Network (CRAN).

Contact me if you have any questions.

Download

LPSBest computes the similarity based on a preset segment length factor and depth. The setting of segment length and the depth is determined based on a leave-one-out (LOO) cross-validation (CV) accuracy on the training data. Attached zip file contains 75 *.Rdata files that contains a list called 'tuned'. 'tuned' has the following information (recall that our experiments allowed for 21 model evaluations).

• $ params: Evaluated parameters. Segment length setting as the factor of time series length and depth  (a 21 by 2 matrix)  
• $ errors: A 21 by number of training time series matrix where ijth entry is 0 if jth time series is classified correctly by the parameter combination in row i of $params, 1 otherwise.
• $ best.error  : best LOO CV error rate
• $ best.seg    : segment length factor that provides the best LOO CV
• $ best.depth  : depth that provides the best LOO CV

Download

Full set of results for LPS submitted to Data Mining and Knowledge Discovery. The excel file has 9 worksheets named as:

1. ProblemChar: Summary of the problem characteristics (75 datasets)
2. Comparison: Test error rates for LPS and competitor algorithms (Tables 1 and 2 in the manuscript)
3. ComparisonGroups: Test error rates and ranks for LPS and competitor algorithms based on the problem category
4. ComparisonGroupsReduced: Test error rates and ranks for LPS and selected competitor algorithms based on the problem category
5. LPSBest: Test error rates for each replication (10 replications, ntree=200) with selected parameters and leave-one-out (LOO) cross-validation (CV) accuracy over training datasets with selected parameters (segment length and depth settings)
6. LPSFullRegr.Splits: Test error rates for each replication (10 replications, ntree=200) for LPS with regression splits and fixed parameters (random segment length, depth=6)
7. LPSFullRandomSplits: Test error rates for each replication (10 replications, ntree=200) for LPS with radom splits and fixed parameters (random segment length, depth=6)
8. CVAccuracyOther: LOO CV accuracies of all classifiers (to draw Texas Sharpshooter Plot)
9. LinesBagnallTest: Error rates from J. Lines and A. Bagnall. Time series classification with ensembles of elastic distance measures. Data Mining and Knowledge Discovery, pages 1–28, 2014

Download

This is Matlab implementation of multivariate version of LPS. Currently R package does not support the multivariate time series.

This version requires each attribute of multivariate time series as seperate files. We also provide the multivariate example dataset (uWaveGesture dataset) here. Please extract the data in the same folder with the matlab scripts and run.