Master's Thesis: Visual Analysis of Big Time Series Datasets

This repository contains all source codes for my master thesis, Visual Analysis of Big Time Series Datasets, including our implementation of the Feature DTW transformation and the Multi-Component Feature DTW transformation.

The structure of this repository is:

• notebooks - Jupyter notebooks with all of our tests and implementation of our analytic pipeline
• src/feature_dtw - source codes for Feature DTW package
• results - Results from our tests on UCR datasets in CSV format
• thesis - LaTeX source codes of the thesis's text

Feature DTW Package

This package contains Python implementation of Feature DTW transformation [1] and its prototyped version [2].

Our implementation is following the well-known Scikit-learn transformer API.

Installation

Firstly we clone our repository:

$ git clone https://github.com/H00N24/visual-analysis-of-big-time-series-datasets.git visual-analysis
$ cd visual-analysis

Our package is working for Python 3.8 and above. To install our package simply use

$ pythom3.8 -m pip install .

We highly recommend to use a virtual environment for installing and testing our package.

Once our package is installed try:

$ python -c "import feature_dtw; print(feature_dtw.__version__)"
1.0.0

Example Usage

import numpy as np
import pandas as pd
from dtaidistance import dtw

from feature_dtw import FeatureDTWTransformer, MultiComponentFDTWTransformer

Time Series Dataset

ts_dataset = np.array(
    [
        [17, 21, 30, 2, 10],
        [16, 15, 4, 3, 9],
        [19, 43, 14, 41, 43],
        [46, 3, 38, 29, 19],
        [18, 10, 16, 36, 6],
        [29, 19, 2, 26, 9],
        [8, 41, 10, 13, 44],
        [7, 38, 34, 20, 20],
    ]
)

Feature DTW transformation

fdtw = FeatureDTWTransformer(
    n_components=2,
    metric=dtw.distance,
    metric_param=dict(window=2, use_pruning=True),
    random_state=42,
)

fdtw.fit_transform(ts_dataset)

Out:

array([[16.03121954, 22.93468988],
       [ 0.        , 21.88606863],
       [56.0357029 , 42.96510212],
       [50.53711507, 28.26658805],
       [30.91924967, 19.87460691],
       [21.88606863,  0.        ],
       [45.16635916, 46.08687449],
       [39.89987469, 35.65108694]])

Multi-Component Time Series Dataset

comp_1 = np.array(
    [
        [40, 21, 9, 41, 45],
        [27, 35, 10, 23, 35],
        [5, 28, 0, 8, 43],
        [47, 14, 38, 14, 12],
        [38, 25, 20, 36, 20],
        [27, 41, 30, 47, 26],
        [20, 25, 30, 8, 13],
        [33, 24, 24, 16, 40],
    ]
)

comp_2 = np.array(
    [
        [46, 40, 7, 27, 3, 44, 38],
        [3, 17, 1, 44, 23, 17, 11],
        [40, 15, 47, 26, 14, 26, 9],
        [4, 46, 25, 3, 4, 40, 13],
        [43, 0, 28, 24, 36, 46, 27],
        [18, 41, 5, 19, 6, 33, 42],
        [9, 4, 2, 37, 39, 18, 15],
        [41, 26, 35, 26, 35, 38, 16],
    ]
)

comp_3 = np.array(
    [
        [18, 27, 12],
        [11, 30, 25],
        [9, 36, 14],
        [33, 1, 9],
        [7, 24, 47],
        [48, 2, 18],
        [24, 31, 48],
        [15, 46, 12],
    ]
)

Multi-Component Feature DTW Transformation

mcfdtw = MultiComponentFDTWTransformer(
    n_transformers=[
        FeatureDTWTransformer(n_components=3, metric="euclidean", random_state=42),
        FeatureDTWTransformer(
            n_components=2,
            metric=dtw.distance,
            metric_param=dict(window=2, use_pruning=True),
            random_state=42,
        ),
        FeatureDTWTransformer(n_components=3, metric="chebyshev", random_state=42),
    ]
)

mcfdtw.fit_transform([comp_1, comp_2, comp_3])

Out:

array([[28.10693865, 37.50999867,  0.        , 66.12866247,         inf,
        13.        , 30.        ,  0.        ],
       [ 0.        , 33.06055051, 28.10693865,  0.        , 52.87721627,
         0.        , 37.        , 13.        ],
       [30.3644529 , 57.99137867, 49.47726751, 41.        , 46.87216658,
        11.        , 39.        ,  9.        ],
       [47.27578661, 49.77951386, 52.50714237, 53.09425581, 39.49683532,
        29.        , 15.        , 26.        ],
       [26.73948391, 25.17935662, 28.12472222, 56.57738064, 36.18010503,
        22.        , 41.        , 35.        ],
       [33.06055051,  0.        , 37.50999867, 52.87721627,  0.        ,
        37.        ,  0.        , 30.        ],
       [35.4682957 , 44.66542287, 54.49770637, 15.03329638, 54.19409562,
        23.        , 30.        , 36.        ],
       [20.66397832, 38.96151948, 30.5450487 , 51.33225107, 47.0637865 ,
        16.        , 44.        , 19.        ]])

Set-up the Work Environment

In this section we will provide guide to set-up your work environment.

We are using Nix, direnv, and Poetry, to maintain a fully reproducible environment. We highly recommend to use these tools.

With Nix, direnv, and Poetry

If you have both Nix and direnv available you can simply use:

$ direnv allow

direnv uses the .envrc file and prepares the full work environment.

or without direnv:

$ nix-shell
$ poetry install

Without Recommended Tools

$ python3.8 -m venv .venv
$ pip install -r requirements.txt -r requirements-dev.txt
$ pip install .

Jupyter Kernel

We are using Jupyter for all of our testing and data science work. As we are using python virtual environment we have to set-up new IPython kernel pointing towards our virtual environment.

python -m ipykernel install --user --name masters-thesis --display-name "Master's Thesis"

References

[1] Kate, Rohit. (2015). Using dynamic time warping distances as features for improved time series classification. Data Mining and Knowledge Discovery. 30. 10.1007/s10618-015-0418-x.

[2] Brian Kenji Iwana, Volkmar Frinken, Kaspar Riesen, Seiichi Uchida, Efficient temporal pattern recognition by means of dissimilarity space embedding with discriminative prototypes, Pattern Recognition, Volume 64, 2017, Pages 268-276, ISSN 0031-3203, https://doi.org/10.1016/j.patcog.2016.11.013. (https://www.sciencedirect.com/science/article/pii/S0031320316303739)

The complete bibliography of my master's thesis is listed in thesis/bib/.

BibTeX

@MastersThesis{Kurákthesis,
    AUTHOR = "KURÁK, Ondrej",
    TITLE = "Visual Analysis of Big Time Series Datasets [online]",
    YEAR = " [cit. 2021-05-17]",
    TYPE = "Master's thesis",
    SCHOOL = "Masaryk University, Faculty of Informatics, Brno",
    SUPERVISOR = "Barbora Kozlíková",
    URL = "Available from WWW <https://is.muni.cz/th/zd4lj/>",
}