Avaliação de modelos de aprendizagem de máquina e técnicas de explicabilidade para descrição de microbioma de solo

Abstract

Machine learning has become increasingly popular and the need to better understand the inner workings of the models developed in the field makes itself clearer day by day. Among the existing resources to better understand said models its SHAP (Shapley Additive Explanations), an approach to explain machine learning models by computing the importance of the features that compose the data. One of the many uses of this tool its to better our understanding not only of the decision taken by the models, but also of the importance of features, and what they represent, in the context of the data’s domain. In this paper, we aim to use this tools to assess the viability of this kind of approach in the field of soil microbial ecology, investigating if machine learning models are capable of classifying metagenomic data and if the analysis of these models is helpful to our understanding of microbial diversity and capable of providing insights that traditional approaches are not. This type of data is characterized by high dimensionality and by the high cost in acquiring new samples, making this problem particularly challenging in the context of machine learning. Therefore, it was also evaluated the effects of projecting the data, for dimensionality reduction, to the obtained results. The results made clear that the dimensionality reduction gave no observable benefit. On the other hand, without projection, the results were promising, albeit limited by the reduced sample size. The explainability analysis showed a small relation with more traditional ones, but were also capable of highlighting features otherwise ignored. Overall, the results were positive, but the need for bigger datasets remained clear. Also, the need to pair these analysis with more in depth investigations became apparent, if we are to validate this approach with more confidence.