tspymfe 0.0.4

Univariate time-series expansion for Pymfe package

ts-pymfe: meta-feature extractor for one-dimensional time-series

A backup for the pymfe expansion for time-series data. Currently, this repository contains the methods for meta-feature extraction and an modified pymfe core to run extract the meta-features.

Please note that tspymfe is not intended to be a stand-alone package, and will be oficially merged (hopefully soon) to the original Pymfe package. Until then, this package is available as a beta version.

There is 149 distinct metafeature extraction methods in this version, distributed in the following groups:

1. General
2. Local statistics
3. Global statistics
4. Statistical tests
5. Autocorrelation
6. Frequency domain
7. Information theory
8. Randomize
9. Landmarking
10. Model based

Install

From pip:

pip install -U tspymfe

or:

python3 -m pip install -U tspymfe

Usage

To extract the meta-features, the API behaves pretty much like the original Pymfe API:

import pymfe.tsmfe
import numpy as np

# random time-series
ts = 0.3 * np.arange(100) + np.random.randn(100)

extractor = pymfe.tsmfe.TSMFE()
extractor.fit(ts)
res = extractor.extract()

print(res)

Dev-install

If you downloaded directly from github, install the required packages using:

pip install -Ur requirements.txt

You can run some test scripts:

python test_a.py <data_id> <random_seed> <precomp 0/1>
python test_b.py <data_id> <random_seed> <precomp 0/1>

Where the first argument is the test time-series id (check data/comp-engine-export-sample.20200503.csv file.) and must be between 0 (inclusive) and 19 (also inclusive), the random seed must be an integer, and precomp is a boolean argument ('0' or '1') to activate the precomputation methods, used to calculate common values between various methods and, therefore, speed the main computations.

Example:

python test_a.py 0 16 1
python test_b.py 0 16 1

The code format style is checked using flake8, pylint and mypy. You can use the Makefile to run all verifications by yourself:

pip install -Ur requirements-dev.txt
make code-check

Available meta-features by group

Below I present the full list of available meta-features in this package separated by meta-feature group. Also note that you can use the following methods to recover the available meta-feature, groups, and summary functions:

import pymfe.tsmfe

groups = pymfe.tsmfe.TSMFE.valid_groups()
print(groups)

metafeatures = pymfe.tsmfe.TSMFE.valid_metafeatures()
print(metafeatures)

summaries = pymfe.tsmfe.TSMFE.valid_summary()
print(summaries)

landmarking:

1. model_arima_010_c
2. model_arima_011_c
3. model_arima_011_nc
4. model_arima_021_c
5. model_arima_100_c
6. model_arima_110_c
7. model_arima_112_nc
8. model_exp
9. model_gaussian
10. model_hwes_ada
11. model_hwes_adm
12. model_linear
13. model_linear_acf_first_nonpos
14. model_linear_embed
15. model_linear_seasonal
16. model_loc_mean
17. model_loc_median
18. model_mean
19. model_mean_acf_first_nonpos
20. model_naive
21. model_naive_drift
22. model_naive_seasonal
23. model_ses
24. model_sine

general:

1. bin_mean
2. cao_e1
3. cao_e2
4. diff
5. emb_dim_cao
6. emb_lag
7. embed_in_shell
8. fnn_prop
9. force_potential
10. frac_cp
11. fs_len
12. length
13. moving_threshold
14. peak_frac
15. period
16. pred
17. step_changes
18. step_changes_trend
19. stick_angles
20. trough_frac
21. turning_points
22. turning_points_trend
23. walker_cross_frac
24. walker_path

global-stat:

1. corr_dim
2. dfa
3. exp_hurst
4. exp_max_lyap
5. ioe_tdelta_mean
6. kurtosis_diff
7. kurtosis_residuals
8. kurtosis_sdiff
9. opt_boxcox_coef
10. sd_diff
11. sd_residuals
12. sd_sdiff
13. season_strenght
14. skewness_diff
15. skewness_residuals
16. skewness_sdiff
17. spikiness
18. t_mean
19. trend_strenght

local-stat:

1. local_extrema
2. local_range
3. lumpiness
4. moving_acf
5. moving_acf_shift
6. moving_approx_ent
7. moving_avg
8. moving_avg_shift
9. moving_gmean
10. moving_gmean_shift
11. moving_kldiv
12. moving_kldiv_shift
13. moving_kurtosis
14. moving_kurtosis_shift
15. moving_lilliefors
16. moving_sd
17. moving_sd_shift
18. moving_skewness
19. moving_skewness_shift
20. moving_var
21. moving_var_shift
22. stability

model-based:

1. avg_cycle_period
2. curvature
3. des_level
4. des_slope
5. ets_level
6. ets_season
7. ets_slope
8. gaussian_r_sqr
9. ioe_std_adj_r_sqr
10. ioe_std_slope
11. linearity

info-theory:

1. low_freq_power
2. ps_entropy
3. ps_freqs
4. ps_peaks
5. ps_residuals

stat-tests:

1. test_adf
2. test_adf_gls
3. test_dw
4. test_earch
5. test_kpss
6. test_lb
7. test_lilliefors
8. test_pp
9. test_za

autocorr:

1. acf
2. acf_detrended
3. acf_diff
4. acf_first_nonpos
5. acf_first_nonsig
6. autocorr_crit_pt
7. autocorr_out_dist
8. first_acf_locmin
9. gen_autocorr
10. gresid_autocorr
11. gresid_lbtest
12. pacf
13. pacf_detrended
14. pacf_diff
15. tc3
16. trev

randomize:

1. itrand_acf
2. itrand_mean
3. itrand_sd
4. resample_first_acf_locmin
5. resample_first_acf_nonpos
6. resample_std
7. surr_tc3
8. surr_trev

freq-domain:

1. ami
2. ami_curvature
3. ami_detrended
4. ami_first_critpt
5. approx_entropy
6. control_entropy
7. hist_ent_out_diff
8. hist_entropy
9. lz_complexity
10. sample_entropy
11. surprise

Main references

Papers

1. T.S. Talagala, R.J. Hyndman and G. Athanasopoulos. Meta-learning how to forecast time series (2018)..
2. Kang, Yanfei., Hyndman, R.J., & Smith-Miles, Kate. (2016). Visualising Forecasting Algorithm Performance using Time Series Instance Spaces (Department of Econometrics and Business Statistics Working Paper Series 10/16).
3. C. Lemke, and B. Gabrys. Meta-learning for time series forecasting and forecast combination (Neurocomputing Volume 73, Issues 10–12, June 2010, Pages 2006-2016)
4. B.D. Fulcher and N.S. Jones. hctsa: A computational framework for automated time-series phenotyping using massive feature extraction. Cell Systems 5, 527 (2017).
5. B.D. Fulcher, M.A. Little, N.S. Jones. Highly comparative time-series analysis: the empirical structure of time series and their methods. J. Roy. Soc. Interface 10, 83 (2013).

Books

1. Hyndman, R.J., & Athanasopoulos, G. (2018) Forecasting: principles and practice, 2nd edition, OTexts: Melbourne, Australia. OTexts.com/fpp2. Accessed on April 29 2020.

Packages

1. tsfeatures (R language)
2. hctsa (Matlab language)

Data

Data sampled from: https://comp-engine.org/