The Journal MIAU has been working for 25+ years as a kind of public service!
MATARKA-nézet
Statement for the Reader:
the thoughts presented on the virtual sites of the e-Journal of MIAU
should be interpreted like e.g. Ramanudjan's notices:
the value of the thoughts comes from the inner world of the publication -
the publication channel is not part of the value-building-processes...
Instead of social media: MYX-team-NEWS
(Last modified: 2015.VII.19.14:08 - MIAÚ-RSS)
Special rules for the value of attributes concerning parallel derivation of production functions – or staircase function with partial correlation instead of regression
Leading article: 2021. June (MIAU No. 274.)
(Previous article: MIAU No. 273.)
Keywords: regression, additivity, multiplicativity, impact, similarity analysis, partial correlation
Abstract: The naïve modelling experts think, models will become better if more attributes (Xi) with (as
far as possible higher) positive correlation values (compared to Y) can be involved into a model.
Therefore, it seems to be that the attributes (Xi) with the most negative correlations are the less useful
attributes to derive a production function for a Y-attribute. It is the same effect as in case of ships,
which could only be used for centuries if the wind comes from behind – and suddenly the appropriate
technique as such got identified. Unfortunately, the regression models (being since ever the
fundamental solution in modelling challenges) can really intagrate attributes (Xi) with a higher
probability when the correlation values (compared to Y) are positive and as higher as possible. All this
is but not valid for the similarity analyses (for the staircase functions). The models with double
attribute-sets are even capable of domesticating useless polynomial effects. In case of staircase
functions, it is to derive, that the primary correlation value of a production function between the facts
and estimations can be increased if (in case of a lot of Xi) the attributes of the first selection are hidden
and based on this step-by-step logic more parallel production functions can be built – while each single
attribute belongs to a production function AND a new production function will be created based on
the estimations (Xi) of the parallel models. Even, the primary layer (estimation) and therefore each of
its attributes can be excluded from the modelling process and the correlation value is still higher than
in case of the primary model with the first selection of (“full-blood”-) attributes. The number of the
full-blood-attributes can also be higher than the number of such (half-blood) attributes which are used
for the less robust production functions in a direct way – and these quasi useless attributes lead to the
quasi same high correlation values as the full-blood-attributes after the first selection phase – in form
of more alternative solutions. Therefore, the value of an attribute depends on the involved modelling
methods/techniques and not directly from the correlation between Xi and Y (c.f. from coal powder can
diamond be pressed).
All these are valid for additive models. The alternatives, the multiplicative models can not exclude
attributes in a way where the parameters will be set to zero because one single multiplication with null
leads to null independent from each further attribute. So, the multiplicative models always reflect the
basic rule: each attribute should always have a connection to each other attribute.
The invalidity of the naïve rule about correlation and usefulness can be proven based on one single
case study where a random-like input OAM can demonstrate the above-mentioned results. The paper
presents this successful case study – where the first analysis led to the needed evidence at once.
Based on this evidence, the attributes may not be interpreted from now on as before. If the new higher
correlation level is arbitrary near to 1.000, then a new aspect of the modelling should be focused (c.f.
STEP-IX-logic as an alternative approximation of the fuzzy-logic) – in order to be capable of talking
about robustness of the modelling.
Finally, the bridge between the naïve logic and the explored evidence is the term of the partial
correlation. This is only valid in case of staircase functions and not in case of regression models. The
partial correlation makes possible that an attribute with highly negative (full) correlation value may
deliver a highly positive (partial) correlation value if only parts of the known values will be interpreted.
More (DOC) ***
More (PDF)
Please, send Your comments per email!
((Back))
miau.my-x.hu
myxfree.tool
rss.services