Recommender systems for fossil community distribution modelling

Abstract

We propose to leverage recommender systems from machine learning to build large-scale community distribution models for the mammalian fossil record. Recommender systems are behind most online life today, from shopping to news personalisation, online dating, or the selection of study programmes or fastest routes. Many recommender systems work by predicting user preferences from items that occur together in user profiles. Technically, this setting closely resembles co-occurrence of species in natural environments. Here we frame community distribution modelling as a recommender systems task, tailor existing recommender techniques for this purpose and propose optimisation criteria for fitting the models in the ecological context. The predictive power comes from species co-occurrences. We demonstrate the potential of this approach for analysing past ecosystems on a case study of Miocene fossil sites in Europe, where we use the proposed community distribution modelling for reconstructing companionships and relative abundances of large mammals. The proposed approach to community distribution modelling, although not climatically explicit, can help to reconstruct past ecosystems and analyse their structure and dynamics over time and space. It also allows, even coarsely, to predict relative abundances of fossil species from presence–absence data. More generally, the proposed perspective is a means for analysis of fossil communities and the relationships between their ecological contexts.