ELIS is an interpretable shapelet-based classifier for time series classification. It provides the efficient discovery for shapelet candidates which are the most discriminating time series subsequence for each class. It also supports the fine-grained shapelet adjustment to achieve accurate time series classification.
Our novel approach yields significant improvements to all three of the important metrics the industry is now using for performance evaluation: (1) accuracy, (2)computational efficiency, and (3) interpretability. Moreover, unlike previous methods, ELIS has fewer parameters to tune.
ELIS is built by CMake. To build ELIS, execute following commands in the src directory:
mkdir run
mkdir run/data
cmake .
make
1.discover shapelet candidates from training time series:
run/discover ts_length training_file ts_number
2.adjust shapelets and train classifier:
run/adjust train ts_length training_file ts_number alpha lambda regularization learning_rate
3.predict for test time series
run/adjust test ts_length testing_file ts_number
We show an example of training and testing with Beef dataset in UCR Benchmark.
cd run
mkdir beef
cd beef
pwd
../discover 470 ../data/Beef/Beef_TRAIN 30 > dout
../adjust train 470 ../data/Beef/Beef_TRAIN 30 -25 0.01 7600 0.01 > tout
../adjust test 470 ../data/Beef/Beef_TEST 30 -25 > rout
cd ..
Our approach ourputs 3 txt files:
init.txt - the discovered shapelet candidates for each class
learned.txt - the shapelets for each class
result.txt - testing results
We provide a script to run tests on 15 datasets
./script