Analysis: 1. level

Online testing

Preconditions for using online testing:

  • The stair functions produced as the result of online similarity analysis (be it standard or polynomial) draws an expert system (calculation table), based on which the Y-values of the input variants defined by any stair combination can be calculated.
  • The heart of calculation is the conversion of the test data, which was not used during learning, into ranking values.
  • This can be done manually (cf. online expert system- next menu point within the analysis menu)
  • Or by the algorithmic way (just as in this case): the corresponding algorithm does nothing else than searching for the closest known learning value in the attribute for every test value, then it makes its rank the stair level of the corresponding data.
  • Attention! The testing approach, which makes calculations from two neighboring ranks per attributes, so it makes 2^n calculation (where n is the number of attributes) in order to make the estimation interval-like, gives the basis for some kind of sensitivity check.
  • The initialization of the online test module is done in two parallel steps. The total data asset of the online similarity analysis is needed to be uploaded, which implies the existence of four matrixes of the same size, where the number of rows is equal with the number of objects (even so, when the number of stairs is fewer). The first matrix consists of the primary data, the second is for their learning ranks, the third is for the calculated stairs, and the fourth is for the estimations per objects. Besides that, the primary test data is needed to be uploaded too, where the number of columns in the OAM is equal with learning's, but the number of rows may be optional.
  • After initializing the two OAMs, they needed to be uploaded, and for that, two codes are needed to be given.
  • Note: The ranking values for real test-records can be simply derived in Excel through the COUNTIFS() function (also according to the direction of the attributes)...

Attached documents: (URL)