Separating wheat from chaff: Diatom taxon selection using an artificial neural network pruning algorithm

This study addresses the question of what diatom taxa to include in a modern calibration set based on their relative contribution in a palaeolimnological calibration model. Using a pruning algorithm for Artificial Neural Networks (ANNs) which determines the functionality of individual taxa in terms of model performance, we pruned the Surface Water Acidification Project (SWAP) pH-diatom data-set until the predictive performance of the pruned set (as assessed by a jackknifing procedure) was statistically different from the initial full-set. Our results, based on the validation at each 5% data-set reduction, show that (i) 85% of the taxa can be removed without any effect on the pH model calibration performance, and (ii) that the complexity and the dimensionality reduction of the model by the removal of these non-essential or redundant taxa greatly improve the robustness of the calibration. A comparison between the commonly used ‘‘marginal’’ criteria for inclusion (species tolerance and Hill’s N2) and our functionality criterion shows that the importance of each taxon in an ANN palaeolimnological model calibration does not appear to depend on these marginal characteristics. Introduction grees, transparency as to the way information is extracted from the assemblage data and implemented Several types of algorithm have been proposed to in the predictive model. While it is clear that the develop quantitative inference models in palaeolimpredictive ability of these models can depend on the nology (Birks 1995): Weighted Averaging regression statistical characteristics of the calibration set (dis/calibration (WA) (ter Braak and van Dam 1989; tribution and range of the environmental variable, Birks et al. 1990), Weighted Averaging Partial Least number of samples, number of taxa, etc.), the modelSquare regression (WA-PLS) (ter Braak and Juggins ling approach is also important to the final success of 1993), Gaussian regression and maximum likelihood the model. Although some methods have been shown calibration (ter Braak and van Dam 1989; ter Braak et to outperform others in certain conditions (ter Braak al. 1993; Vasko et al. 2000), and back-propagation and Juggins 1993; ter Braak et al. 1993; ter Braak (BP) (Rumelhart et al. 1986) of Artificial Neural 1995; Racca et al. 2001), little is known about the Networks (ANNs) (Racca et al. 2001). All of these inclusion or exclusion of taxa based on their contribumethods have inherent but different abilities to model tion to the calibration model. Generally, calibration the complex relations between taxon assemblages and data-sets are large and sparse and the criterion for environmental variables and all yield successful pretaxon inclusion is typically ad hoc (e.g., all taxa with dictive models. However, they lack, to varying de1% relative abundance in at least one sample, present