tsclassification provides a solution for quickly and accurately classifying time series data. It leverages shapelet-based learning techniques to identify the most informative subsequences within time series data. The algorithm was originally built for human activity recognition, but can generalize to many applications in time series pattern matching and classification.
Utilizes various methods to extract class-specific shapelets that best represent the patterns in the class' time series data. Shapelet learning is hardware accelerated.
Scores incoming time series against class shapelets, passes the results through a multilayer perceptron, and provides a multiclass classification label.
The methodology in tsclassification is detailed in the paper "Real-Time Human Activity Classification Using Gait Cycle Averaging and Biometric Heuristics" by Grant Ellison, M.P. Markovic, and Delaram Yazdansepas, presented at the 22nd International Conference on Machine Learning and Applications (ICMLA).
@inproceedings{gellison23,
title={Real-Time Human Activity Classification Using Gait Cycle Averaging and Biometric Heuristics},
author={Grant Ellison and M.P. Markovic and Delaram Yazdansepas},
journal={22nd International Conference on Machine Learning and Applications (ICMLA)},
year={2023},
publisher={IEEE},
DOI = {10.1109/ICMLA58977.2023.00056}
}$ pip install tsclassification
# ShapeletClassifier
clf = ShapeletClassifier(metric = 'dtw', # 'dtw' or 'euclidean'
classification_window = 300, # the discrete time steps to classify
w = 0.5 # warping constraint for the dtw measure
)'''
Fits the model to the provided dataset by extracting shapelets, compiling the data for training,
and then training a neural network classifier based on the extracted shapelets and compiled data.
Parameters
- X (array-like): A 1D time series dataset.
- y (array-like): The target values (class labels) for the samples in X.
- shapelet_method (str): The method used to extract shapelets from the dataset.
- 'barycenter' : Performs peak-analysis to extract subsequences.
Averages subsequences to a barycenter shapelet.
- 'random' : Chooses a random qty of subsequences. Chooses the one
with the minimum cumulative distance to each other subsequence.
- 'exhaustive : Chooses subsequences using a stepped, sliding
window. Chooses the one with the minimum cumulative distance to
each other subsequence.
- compiler_method (str): The method used to compile the dataset for training the neural network after
shapelet extraction.
- 'bootstrapped' : slides a stepped window, randomly placed, for
set number of samples. Repeats until the total qty of samples have been extracted.
- 'slide' : A simple, stepped, sliding window.
- 'random' : Randomly positioned, fixed windows are extracted until
the total qty of samples have been extracted.
- **kwargs: Additional keyword arguments that are passed to the shapelet extraction method and
neural network training procedure.
Note
The choice of `shapelet_method` and `compiler_method`, along with specific parameters passed
through `**kwargs`, significantly affects the model's performance and training time.
It's essential to choose these parameters carefully based on the dataset characteristics and the
analysis goals.
'''
clf.fit(X, # 1d time series array
y, # 1d list of classes
shapelet_method = 'barycenter',
sampling_method = 'bootstrapped',
qty = 100, # number of samples to extract for training the MLP
window = 100, # window size of 'slide' method
step = 30, # the step size for the sliding window technique
min_dist = 70, # the minimum peak distance for barycenter shapelet method
max_dist = 120, # the maximum peak distance for barycenter shapelet method
verbose = False,
max_iter = 2000, # **kwargs pass through to SKLearn's MLP class
hidden_layer_sizes = (10,)
)# takes in a list of time series arrays and returns a list of predicted classes
clf.predict([[]]) # Saving and loading the model
clf.save('path/to/file') # pickles the model
with open('path/to/file', 'rb') as f:
clf = pickle.load(f)
clf.predict([[]])