The purpose of this paper is to propose a noise identification method for data without temporal structure, in which application of typical mathematical white or colored noise models is very limited due to observation order requirements. The method is used to identify the destructive elements and to eliminate them what finally brings prediction improvement.
The paper concerns noise detection problem presented in the framework of ensemble methods via blind signals separation. The authors utilize the Extended Generalized Lambda Distribution (EGLD) model to compare the signals with the target.
The authors proposed novel signals similarity measure which is based on the EGLD system. The authors showed that it can be applied for data with or without time structure, as well as for data which are mutually uncorrelated. It turned out that method is effective for noise identification and can be an alternative, in many cases, to correlation approach, particularly for noise identification problems.
In this method the improvement of prediction results is associated with elimination of the real physical factors rather than mathematical averaging in terms of arbitrary assumed distributions. In this approach, it does not matter what is the structure of aggregated models, what significantly distinct this approach from such techniques as boosting or bagging, in which the aggregation process applies to the models of similar structure. For this reason the methodology is focussed on physical noises elimination from predictions and it is complementary to the other ensemble approaches.
The work was funded by the National Science Center in Poland based on decision number DEC-2011/03/B/HS4/05092.
Szupiluk, R. and Ząbkowski, T. (2014), "EGLD system for noise identification in ensemble predictors", COMPEL - The international journal for computation and mathematics in electrical and electronic engineering, Vol. 33 No. 6, pp. 2006-2015. https://doi.org/10.1108/COMPEL-11-2013-0354Download as .RIS
Emerald Group Publishing Limited
Copyright © 2014, Emerald Group Publishing Limited