5.6. GTIVE to GTSDA Migration Guide

Prior to version 6.0 the pSeven Core package included the Generic Tool for Important Variable Extraction (GTIVE) module which provided global sensitivity analysis methods. This module had been deprecated since pSeven Core 4.0 due to the appearance of a stable release of the GTSDA module (the beta version of GTSDA was introduced earlier in pSeven Core 3.0). Finally, in pSeven Core 6.0 the GTIVE module was removed and active development of sensitivity analysis methods now continues in GTSDA.

Feature score estimation methods that were available from GTIVE are mostly covered by the rank() method in GTSDA (see the details below). Note that GTIVE provided score variance and standard deviation in addition to score values; GTSDA rank() calculates scores only.

GTIVE provided the following ranking techniques:

  • RidgeFS — Ridge regression technique, sample-based.
  • Mutual Information — mutual information technique, sample-based.
  • Elementary Effects — elementary effects technique, blackbox-based.
  • FAST — Fourier Amplitude Sensitivity Testing technique, blackbox-based.
  • SMBFAST — Surrogate Model Based FAST, the sample-based version of the FAST technique.

GTSDA modified the choice of techniques as follows:

  1. RidgeFS is no longer available as a dedicated technique, but similar results are obtained when GTSDA/Ranker/Technique is "Sobol" and the approx_options argument to rank() sets GTApprox/Technique to "RSM".

  2. Mutual information is no longer used in ranking tasks. In GTSDA an updated and improved implementation of the Mutual Information technique is used to estimate correlations (see Correlation Analysis, Mutual Information). To calculate mutual information with GTSDA, use check() and set GTSDA/Checker/Technique to "MutualInformation".

    Technique behavior is now more automated, and fine-tuning options are no longer available. Also note that intended usage of the Mutual Information technique is to test if dependency exists (not to rank features), so it is recommended to use this technique for correlation analysis only.

  3. To use the Elementary Effects technique, set GTSDA/Ranker/Technique to "Screening" (the other name for this technique is Morris Screening — see section Screening Indices for details).

  4. Both the FAST and SMBFAST techniques have been unified into a single FAST technique. To use FAST, set GTSDA/Ranker/Technique to "Sobol" and GTSDA/Ranker/Sobol/Method to "FAST".

    GTSDA now computes Sobol indices in both the sample- and blackbox-based modes — see the argument combinations for rank(). Results from the sample-based rank() are equivalent to SMBFAST results in GTIVE, and results from the blackbox-based rank() are equivalent to those of the GTIVE FAST method. Note also that GTSDA allows to configure the approximator used by the sample-based mode rank() — see the approx_options argument.