Proximity Forest (Using only DTW measure)

An effective and scalable distance-based classifier for time series classification. This repostitory contains the source code for the time series classification algorithm Proximity Forest, published in the paper [https://arxiv.org/abs/1808.10594]

Check out the original project developed in Java: https://github.com/fpetitjean/ProximityForest

Proximity Forests in Python using dtai DTW distance measure

This is the Proximity Forest algorithm implemented to classify Time Series.

Requirements:

• You can only run this project if you have a UNIX-based OS (Linux or OSX)
• You must have python3.7
• Give full permission to your project directory and files
• We recommend you locate your project in your $HOME directory.

Packages to Install Previously:

• pip install dtaidistance or pip3 install dtaidistance
• pip install -v --force-reinstall --no-deps --no-binary dtaidistance dtaidistance
• pip install scipy or pip3 install scipy

How to run a simple execution:

• Download the project

• Give full permissions to your project

• Go to the directory launchers

• Execute the following command:

  • sh TermLauncherParam.sh [DATASET_NAME], where [DATASET_NAME] is the name of the dataset you have stored in the project. The Names available are:
    • ItalyPowerDemand
    • MoteStrain
    • Plane
    • GesturePebbleZ2

• Try this example: - sh TermLauncherParam.sh ItalyPowerDemand

Parameters of python execution:

• Run the sh file application/TermLauncher.sh setting the following parameters:
• name: The name of the experiment
• -train: The training dataset path. It must have a .ts or .arff extension.
• -test: The testing dataset path. It must have a .ts or .arff extension.
• The training and testing datasets can be found in the folder datasets/
• -repeat: Number of repeats of the experiment
• -trees: Number of trees of the PForest
• -candidates: Number of candidates per tree
• -targetlast: It indicates if the last column of the dataset is the class of the serie.
  • P.e. if the line is 0.32, 0.45, ..., -0.12, X_AXIS, setting -targetlast=True means the X_AXIS is the series class and it's located in the last column

Comparison with Sktime.ProximityForest

sktime is a Python machine learning toolbox for time series with a unified interface for multiple learning tasks. You can find the package through this URL: https://sktime.org

If you want to compare this Proximity Forest Project to the sktime project:

• Run the sh file launchers/launchersktime.sh, which can be found in the launchers directory, setting the following arguments:
  • name: The name of the dataset. The Names available are:
    • ItalyPowerDemand
    • MoteStrain
    • Plane
    • GesturePebbleZ2
    • Chinatown
    • ERing
    • FacesUCR
    • FreezerRegularTrain
    • SmoothSubspace
    • ElectricDevices
  • trees: The number of trees for the algorithm
  • candidates: Number of candidates per split
• Example: sh launchersktime.sh Plane 20 3

Installing Package

You can install the package using the command:

pip install Pforests-dtw

Nevertheless, you must have installed the following packages:

• numpy
• dtaidistance
• pytest
• scipy

Example

from trees import ProximityForest
from core import FileReader
import random
train_dataset = FileReader.FileReader.load_arff_data("/Users/moradisten/Projects/PForests/datasets/Plane/Plane_TRAIN.arff")
test_dataset = FileReader.FileReader.load_arff_data("/Users/moradisten/Projects/PForests/datasets/Plane/Plane_TEST.arff")
Pforest = ProximityForest.ProximityForest(1, n_trees=100, n_candidates=5)
Pforest.train(train_dataset)
results = Pforest.test(test_dataset)
print(results.accuracy)

Data Structure: ListDataset

Normally, in Data Science, we are used to handle datasets with Pandas. But for these case, as researchers did, a data structure has been developed, which is the ListDataset. This data structure contains:

• series_data: Contains a list of the time series (X).
• classes: Contains a list of the labels (y).
• class_counter: Dictionary which contains the count of each label -> <label, label count>
• series_map: Map which indicates the number of series per label -> <label, nº series>

File Format

This project reads csv and arff files.

My Bachelor Thesis

If you speak Spanish, I invite you to read my bachelor's thesis about this matter

https://figshare.com/articles/thesis/Distance-based_Time_Series_Classifiers/13269005