12 Self study - Session 3

In this session, we will revisit the chapters Chapter 7, Chapter 8, and Chapter 9.

We have seen that the algorithms for all three tasks more or less can be categorized into the same categories:

Based on distances or metrics: E.g. trees and forests
Based on distributions, kernels: E.g. Gaussian Mixture Models, SVMs
Based on subsequences, shapelets: E.g. bags of words, shapelet-based methods
Based on deep learning: E.g. recurrent neural networks, or treating the time series as images using CNNs
Based on random convolutions: E.g. Random Convolutional Kernel Transform (ROCKET)
Based on ensembles of the above.

We have implemented nearest neighbors (distance-based) for clustering, time series forests (distance-based) for classification, and convolutional neural networks for regression (deep learning-based).

Let us have a look at one additional algorithm for each task.

Note

You can use sktime’s load_UCR_UEA_dataset function to load the data sets mentioned below, e.g.

from sktime.datasets import load_UCR_UEA_dataset

ds_name = "UWaveGestureLibrary"
ret_type = "numpy3d"

X_train, y_train = load_UCR_UEA_dataset(
    name=ds_name,
    split="train",
    return_type=ret_type,
)

X_test, y_test = load_UCR_UEA_dataset(
    name=ds_name,
    split="test",
    return_type=ret_type,
)

# For classification, convert y to categorical
y_train = pd.Series(y_train).astype("category")
y_test = pd.Series(y_test).astype("category")

sktime is not the sole source of time series based algorithms. Besides sktime, there are numerous others, e.g. aeon.

Note

aeon is quite strict about dependencies. In case you have problems installing it, consider to initialize a new project and first install aeon before adding further packages.

Exercise 12.1 (Clustering of univariate time series)

Load the GesturePebbleZ1 data set. Note that it is a variable-length time series dataset and you need to either use ret_type="df-list" or ret_type="nested_univ" to load it (using the example above).
Bring the data into a suitable format and explore the data set visually (using plots).
Use the Time Series \(k\) means algorithm to cluster the data. Do you need to reshape the data for this algorithm?
Use the KASBA Holder and Bagnall (2026) algorithm to cluster the data (e.g. implemented in aeon).
Include useful (!) visualizations.
How do the results compare?

After the exercise, you should be able to answer the following questions:

How does KASBA differ from the commonly used \(k\) means algorithm?
In you will find MSM defined. You should be able to illustrate how MSM is calculated.

In Chapter 8 and Chapter 9, our input was univariate, meaning each sample consisted of exactly one time series. Now, we will work with multivariate time series, which are more common in real-world scenarios since typically multiple sensors are available.

Exercise 12.2 (Classification of multivariate time series)

Load the UWaveGestureLibrary data set.
Use the time series \(k\) nearest neighbour algorithm to classify the time series.
Use the CNN classifier to classify the time series.
How do the results compare?

Note that not all of the algorithms have implementations for multivariate time series, e.g. sktime’s TimeSeriesForestClassifier only works for univariate time series (and raises an exception when used with a multivariate series). Luckily, there are other implementations available, e.g. aeon offers a TimeSeriesForestClassifier that is compatible with multivariate time series.

Exercise 12.3 (Classification using aeon’s Time Series Forest Classifier)

Load the UWaveGestureLibrary data set.
Use aeon’s TimeSeriesForestClassifier to classify the data set.
Are you able to achieve better results than with the previous algorithms?

Note

You can use aeon’s load_regression function to load the data set mentioned below, e.g.

from aeon.datasets import load_regression

ds_name = "AppliancesEnergy"

X_train, y_train = load_regression(ds_name, split="train")
X_test, y_test = load_regression(ds_name, split="test")

Exercise 12.4 (Regression of multivariate time series)

Load the AppliancesEnergy data set.
Use the time series \(k\) nearest neighbour algorithm to predict the energy consumption of the appliances.
Use the time series random forest tree algorithm to predict the energy consumption of the appliances.
Use the CNN regressor to predict the energy consumption of the appliances.
How do the results compare?