RSAST: Sampling Shapelets for Interpretable Time Series Classification

Random SAST (RSAST) is a method based on STC and SAST, that generates shapelets randomly, guided by statistical criteria, reducing the search space of shapelets.

Results RSAST

• Results Default Split

• Results 10 Resamplings

• Results Comparison RSAST

• Execution time regarding the number of series

• Execution time regarding series length

RSAST, SAST and STC

Critical difference diagram

Pairwise accuracy comparison

Shapelet Approaches

Critical difference diagram

Pairwise accuracy comparison

Alternative Length Methods

In order to explore another alternatives for the default length method of the shapelets (ACF&PACF) some supplementary length methods are examined: Max PACF and None.

Critical difference diagram per length method

• The default behaviour implies chose all significant values from ACF and PACF tests.

• Max ACF, makes reference to the generation of subsequences considering solely the highest significant value from the Partial Autocorrelation Function (PACF).

• "None" variation, involves generating subsequences with a single random length chosen from the range between 3 and the size of the time series for each randomly selected instance.

Critical difference diagram: Best Performance Comparison

Scalability

• Regarding the length of time series:

• Regarding the number of time series in the dataset:

Usage

import os, numpy as np
from utils_sast import load_dataset, format_dataset
from sast import RSAST
from sklearn.linear_model import RidgeClassifierCV

ds='Coffee' # Chosing a dataset 
path=os.getcwd()+"/data"

ds_train , ds_test = load_dataset(ds_folder=path,ds_name=ds,shuffle=False)
X_test, y_test = format_dataset(ds_test)
X_train, y_train = format_dataset(ds_train)
clf = RidgeClassifierCV(alphas=np.logspace(-3, 3, 10))
rsast_ridge = RSAST(n_random_points=10, nb_inst_per_class=10, len_method="both", classifier=clf)
rsast_ridge.fit(X_train, y_train)
print('rsast score :', rsast_ridge.score(X_test, y_test))