Boosting, Bagging and Ensembles in the Real World: An Overview, some
Textbox 3.1 On the suggested ‘best’ use of competing landcover, altitude and climate layers for better inference with Machine Learning: Being
3.6 Model Applications and Inference .1 Boosting Experiences and Applications
Stochastic gradient boosting is a rather strong concept per se and allows for a wide range of applications (see for instance De'ath 2007 for Ecology). I use it daily for rapid assessments, first error checks and pattern detections in data to gauge further analysis. It is specifically used for data mining, variance reduction and pattern detections in data across disciplines. Using it for predictions, and when optimized, becomes really strong for inference. I noticed a few atypical cases where boosting
can be erroneous for overfitting (=where a good model cannot be achieved, even when fine-tuned and with good predictors that achieve well in other algorithms).
In those cases -rare though - I tend to use ‘stumps’ (very simple split trees) and cut back on predictors and node depths.
Using the R implementations of boosting often asks for known distributions of data to be analyzed; I found this a question impossible to answer in complex data mining applications beforehand, and thus, I see it as a major drawback in the impli- cation (R). In my own work, primarily with SPM8, I use boosting for instance to explore data and to mine data for their initial patterns. Once those ‘signals’ in the data are found and become clear, I then follow up with other tools for a test, and whether those patterns can even improve further (often not the case though and thus I remain with the initial analysis!).
3.6.2 Bagging Experiences and Applications
Bagging is popular in several ways. By now, it is among the most widely used high- performance classifier out there (Ferandez-Delgado et al. 2014). It’s the method of choice for many remote sensing classification and modeling problems (e.g. Evans and Cushman 2009; Evans et al. 2010; Chunrong et al. 2017; see Cutler et al. 2007 for Ecology). Myself, I see it as the prime prediction machine for many species model applications (Drew et al. 2011; see Yen et al. 2004; Kandel et al. 2015; Chunrong et al. 2017 for a test and application with species of high conservation profile).
Using random forests with regression problems can prove less successful for the unexperienced user. Ensembles models are suggested in such cases.
Unfortunately, the vast implications of random forests in order to change, improve and progress science in itself have not been fully realized by the natural resource conservation management community. First steps have been done, but arguably, many large ones are to come still, e.g. to use RandomForest as the base approach for obtaining a robust ‘Learner.’ One way or another, this will move the conservation management of natural resources into computer-aided decision- making, for instance, beyond also having many other impacts, e.g. job descriptions, publications, ethics, business models and education and institutions. Textbooks may be re-written and are to be extended accordingly.
3.6.3 Ensembles
The global discussion on climate change - arguably the major scheme of our time and certainly for natural resource management as we know it - is driven by a set of ensemble models (IPPC). There is no need to stress the relevance of this approach,
and of its underlying algorithms, e.g. Baltensperger and Huettmann (2015). While bagging is on the rise, so are ensemble models even more so (see Elder 2003, for an application: Jiao et al. 2014).
3.6.4 Precautionary, Pro–Active, and Predictive Models for Better Resource Conservation Management
Applications that involve predictions for natural resource management are not only mature established, but for globally relevant ones they now become the norm (Table 3.3 for overview). In addition, the true relevance of those approaches is still not widely embraced, nor is climate change and how to manage natural resources in a predictive fashion (e.g. see in Silva 2012; but see O’Connor et al. 1996, Nielsen et al. 2008 and Chunrong et al. 2016 for examples). As long as man-made climate change is not accepted as a key problem, one cannot go ahead much with a mean- ingful approach to predictions used for policy. We are thus in true violation of the UN Agreement on Pre-cautionary, Pro-Active Management Principles. One can only be pro-active, in the best available fashion, when being predictive. And best predictions are only possible when using best-available approaches (See Tables 3.2
& 3.4) That is what Leo Breiman (2001a, b) and others offer, but that is widely missing from the textbooks on conservation and natural resource management (Verner et al. 1986; Romesburg 1989; Silva 2012). Table 3.4 shows some disciplines where Machine Learning still needs to be established further for progress.
Table 3.3 A small selection of application fields of CARTs, boosting, bagging and ensembles for wildlife conservation worldwide
Application Algorithm Detail of progress achieved Citation Climate
prediction
Ensembles Global progress of climate for sustainable decision-making
IPCC (https://www.ipcc.ch) Species
forecast Ensembles First-time forecast of the species
niche for an endemic subspecies Lawler et al. (2006, 2011), Hardy et al. (2011), Chunrong et al. (2016) Human
health
CART First-time connection on a national scale between avian diversity and human health
O’Connor et al. (1996)
Boosting &
Bagging
First-time global model for influenza
Herrick et al. (2013) Rapid
assessment
Boosting &
Bagging
First-time model for that species over large areas
Kandel et al. (2015), Regmi et al. (2018)