MetaTS | Meta-Learning for Global Time Series Forecasting

Features:

• Generating meta features
  • Statistical features : TsFresh, User defined features
  • Automated feature extraction using Deep Unsupervised Learning : Deep AutoEncoder (MLP, LSTM, GRU, ot custom model)
• Supporting sktime and darts libraries for base-forecasters
• Providing a Meta-Learning pipeline

Quick Start

Installing the package

pip install metats

Generating a toy dataset by sampling from two different processes

from metats.datasets import ETSDataset

ets_generator = ETSDataset({'A,N,N': 512,
                            'M,M,M': 512}, length=30, freq=4)

data, labels = ets_generator.load(return_family=True)
colors = list(map(lambda x: (x=='A,N,N')*1, labels))

Normalizing the time series

from sklearn.preprocessing import StandardScaler

scaled_data = StandardScaler().fit_transform(data.T)
data = scaled_data.T[:, :, None]

Checking How data looks like

import matplotlib.pyplot as plt
_ = plt.plot(data[10, :, 0])

Generating the meta-features

Statistical features using TsFresh

from metats.features.statistical import TsFresh

stat_features = TsFresh().transform(data)

Deep Unsupervised Features

Training an AutoEncoder

from metats.features.unsupervised import DeepAutoEncoder
from metats.features.deep import AutoEncoder, MLPEncoder, MLPDecoder

enc = MLPEncoder(input_size=1, input_length=30, latent_size=8, hidden_layers=(16,))
dec = MLPDecoder(input_size=1, input_length=30, latent_size=8, hidden_layers=(16,))

ae = AutoEncoder(encoder=enc, decoder=dec)
ae_feature = DeepAutoEncoder(auto_encoder=ae, epochs=150, verbose=True)

ae_feature.fit(data)

Generating features using the auto-encoder

deep_features = ae_feature.transform(data)

Visualizing both statistical and deep meta-features

Dimensionality reduction using UMAP for visualization

from umap import UMAP
deep_reduced = UMAP().fit_transform(deep_features)
stat_reduced = UMAP().fit_transform(stat_features)

Visualizing the statistical features:

plt.scatter(stat_reduced[:512, 0], stat_reduced[:512, 1], c='#e74c3c', label='ANN')
plt.scatter(stat_reduced[512:, 0], stat_reduced[512:, 1], c='#9b59b6', label='MMM')
plt.legend()
plt.title('TsFresh Meta-Features')
_ = plt.show()

And similarly the auto encoder's features

plt.scatter(deep_reduced[:512, 0], deep_reduced[:512, 1], c='#e74c3c', label='ANN')
plt.scatter(deep_reduced[512:, 0], deep_reduced[512:, 1], c='#9b59b6', label='MMM')
plt.legend()
plt.title('Deep Unsupervised Meta-Features')
_ = plt.show()

Meta-Learning Pipeline

Creating a meta-learning pipeline with selection strategy:

from metats.pipeline import MetaLearning

pipeline = MetaLearning(method='selection', loss='mse')

Adding AutoEncoder features:

from metats.features.unsupervised import DeepAutoEncoder
from metats.features.deep import AutoEncoder, MLPEncoder, MLPDecoder

enc = MLPEncoder(input_size=1, input_length=23, latent_size=8, hidden_layers=(16,))
dec = MLPDecoder(input_size=1, input_length=23, latent_size=8, hidden_layers=(16,))

ae = AutoEncoder(encoder=enc, decoder=dec)
ae_features = DeepAutoEncoder(auto_encoder=ae, epochs=200, verbose=True)

pipeline.add_feature(ae_features)

You can add as many features as you like:

from metats.features.statistical import TsFresh

stat_features = TsFresh()
pipeline.add_feature(stat_features)

Adding two sktime forecaster as base-forecasters

from sktime.forecasting.naive import NaiveForecaster
from sktime.forecasting.compose import make_reduction
from sklearn.neighbors import KNeighborsRegressor

regressor = KNeighborsRegressor(n_neighbors=1)
forecaster1 = make_reduction(regressor, window_length=15, strategy="recursive")

forecaster2 = NaiveForecaster() 

pipeline.add_forecaster(forecaster1)
pipeline.add_forecaster(forecaster2)

Specify some meta-learner

from sklearn.ensemble import RandomForestClassifier

pipeline.add_metalearner(RandomForestClassifier())

Training the pipeline

pipeline.fit(data, fh=7)

Prediction for another set of data

pipeline.predict(data, fh=7)

About the package

Contributors

• Sasan Barak
• Amirabbas Asadi

We wish to see your name in the list of contributors, So we are waiting for pull requests!

Papers using MetaTS

Below is a list of projects using MetaTS. If you have used MetaTS in a project, you're welcome to make a PR to add it to this list.

• Retail time series forecasting using an automated deep meta-learning framework