17 November 2020 in Ecography
Species distribution models (SDMs) are widely used in ecology, biogeography and conservation biology to estimate relationships between environmental variables and species occurrence data and make predictions of how their distributions vary in space and time. During the past two decades, the field has increasingly made use of machine learning approaches for constructing and validating SDMs. Model accuracy has steadily increased as a result, but the interpretability of the fitted models, for example the relative importance of predictor variables or their causal effects on focal species, has not always kept pace. Here we draw attention to an emerging subdiscipline of artificial intelligence, explainable AI (xAI), as a toolbox for better interpreting SDMs. xAI aims at deciphering the behavior of complex statistical or machine learning models (e.g. neural networks, random forests, boosted regression trees), and can produce more transparent and understandable SDM predictions. We describe the rationale behind xAI and provide a list of tools that can be used to help ecological modelers better understand complex model behavior at different scales. As an example, we perform a reproducible SDM analysis in R on the African elephant and showcase some xAI tools such as local interpretable model‐agnostic explanation (LIME) to help interpret local‐scale behavior of the model. We conclude with what we see as the benefits and caveats of these techniques and advocate for their use to improve the interpretability of machine learning SDMs.
28 September 2018 in Journal of Open Source Software
5 November 2018 in bioRxiv
Word embeddings are omnipresent in Natural Language Processing (NLP) tasks. The same technology which defines words by their context can also define biological species. This study showcases this new method - species embedding ( species2vec ). By proximity sorting of 6761594 mammal observations from the whole world (2862 different species), we are able to create a training corpus for the skip-gram model. The resulting species embeddings are tested in an environmental classification task. The classifier performance confirms the utility of those embeddings in preserving the relationships between species, and also being representative of species consortia in an environment.
17 April 2020 in EcoEvoRxiv
29 October 2018 in bioRxiv
Species Distribution Models (SDMs) are used to generate maps of realised and potential ecological niches for a given species. As any other machine learning technique they can be seen as “black boxes”, due to a lack of interpretability. Advances in other areas of applied machine learning can be applied to remedy this problem. In this study we test a new tool relying on Local Interpretable Model-agnostic Explanations (LIME) by comparing its results of other known methods and ecological interpretations from domain experts. The findings confirm that LIME provides consistent and ecologically sound explanations of climate feature importance during the training of SDMs, and that the sdmexplain R package can be used with confidence.
25 October 2020 in OSF Preprints
The need for large amounts of data permeates almost all fields of research. New technologies related to machine learning (deep learning in particular), cloud computing, the Internet of Things (IoT) add to the increased complexity of data-related work and how researchers deal with that. Those trends are accelerating and result in a widespread need for new frameworks to increase the cost-benefit ratio of research work while driving innovation. Modern developments in the private sector, both in technologies and ways of working, can be readily adopted by research organizations. "Data strategy" is an umbrella term for those advancements. This article presents its research variant, Research Data Strategy (RDS), complete with its different elements, the sequence of execution, and supporting activities. This new methodology aims to build on top of existing research data management practices by providing a framework for ensuring innovative science in data-driven research organizations.