Journal article 1054 views 94 downloads
Constructing compact Takagi-Sugeno rule systems: Identification of complex interactions in epidemiological data
PLoS ONE, Volume: 7, Issue: 12, Start page: e51468
PDF | Version of Record
Distributed under the terms of a Creative Commons Attribution (CC-BY-4.0)Download (857.8KB)
In the identification of non-linear interactions between variables, the Takagi-Sugeno (TS) fuzzy rule system as a widely used data mining technique suffers from the limitations that the number of rules increases dramatically when applied to high dimensional data sets (the curse of dimensionality). H...
|Published in:||PLoS ONE|
Check full text
No Tags, Be the first to tag this record!
In the identification of non-linear interactions between variables, the Takagi-Sugeno (TS) fuzzy rule system as a widely used data mining technique suffers from the limitations that the number of rules increases dramatically when applied to high dimensional data sets (the curse of dimensionality). However, few robust methods are available to tackle this issue, and this results in limited applicability in fields such as epidemiology or bioinformatics where the interaction of many variables must be considered. In this study, we develop a new parsimonious TS rule system. We propose three statistics: R, L, and ω-values, to rank the importance of each TS rule, and a forward selection procedure to construct a final model. We use our method to predict how key components of childhood deprivation combine to influence educational achievement outcome. We show that a parsimonious TS model can be constructed, based on a small subset of rules, that provides an accurate description of the relationship between deprivation indices and educational outcomes. The selected rules shed light on the synergistic relationships between the variables, and reveal that the effect of targeting specific domains of deprivation is crucially dependent on the state of the other domains. Policy decisions need to incorporate these interactions, and deprivation indices should not be considered in isolation. The TS rule system provides a basis for such decision making, and has wide applicability for the identification of non-linear interactions in complex biomedical data.
Health informatics, data mining, interactions, epidemiology, rule modelling, deprivation
Faculty of Medicine, Health and Life Sciences