Fedot.Industrial 0.4.2

In this version:

• added Riemann geometry feature extractor
• we migrated from PIP to Poetry dependancy management
• refactored and updated examples directory
• introduced ssa+ar and glm models for time-series forecasting
• implemented channel filtration method and n-components selection after RSVD
• minor fixes and improvements

Fedot. Industrial ver 0.3

Refactoring feature extraction models. Quantile/topological/recurrence feature extraction models was refactored

• Models have been redesigned and brought to a "single" type of class, in which there are 2 key methods: transform() (for calling inside the Pipeline constructor Pipeline) and the generate_features_from_ts() method for calling these models from the outside. The __transform()_ method inherits from the abstract BaseExtractor class.
• For these models, 2 hyperparameters were derived and obtained: window width and stride. For the quantile model: window width is the width of the window in which the features are calculated, stride is the step size of the sliding window. For recurrent and topological models: window width and stride are the parameters used in the time series Gankelization (converting a series into a trajectory matrix).
• Usage examples are in examples/pipeline_examples/time_series/ts_classification

Refactoring basis models (fedot_ind/core/operation/transformation/basis). Eigen/fourier/wavelet basis models was refactored

• Models have been redesigned and brought to a "single" type of class, in which there are 2 key methods: transform() (for calling inside the Pipeline constructor Pipeline) and the _get_1d_basis() and get_multidim_basis() method for calling these models from the outside. The __transform()_ method inherits from the abstract AbstractBasis class.
• For EigenBasis a new algorithm for low-rank approximation based on Krylov subspaces was developed. This algorithm is located in fedot_ind/core/operation/decomposition/matrix_decomposition/power_iteration_decomposition.py . Due to this algorithm, it was possible to significantly reduce the calculation time of the trajectory matrix eigenvectors basis of the time series, achieve better separability of time series components, and also it became possible to work with "long" time series (>100,000 elements)
• Usage examples are in examples/pipeline_examples/time_series/ts_classification

New type of time series machine learning task. Time Series Forecasting (TSF).

• New forecasting model: Singular Spectrum Analysis (SSA) based on EigenBasis model.
• This model uses EigenBasis to decompose a time series, represented as a trajectory matrix, into components (trend, seasonality, noise). The resulting components are predicted using standard models (e.g. ARIMA).
• Usage examples are in examples/pipeline_examples/time_series/ts_forecasting

New type of time series machine learning task: Time Series Extinct Regression (TSER).

• TSER involves a predictive model built on time series to predict a real-valued variable distinct from the training input series.
• Usage examples are in examples/api_examples/time_series/ts_regression

Unit and Integration tests was refactored.