Résumé: Groups or aggregations of animals can result from individuals being attracted to a common resource or because of synchronised patterns of daily or seasonal activity. Although mostly solitary throughout its distribution, white sharks (Carcharodon carcharias) seasonally aggregate at a number of sites worldwide to feed on calorie-rich pinnipeds. At the Neptune Islands, South Australia, large numbers of white sharks can be sighted throughout the year, including during periods of low seal abundance. We use a combination of photo-identification and network analysis based on co-occurrence of individuals visiting the site on the same day to elucidate the population structure and aggregatory behaviour of Australia's largest aggregation of sub-adult and adult white sharks. We photo-identified 282 sharks (183 males, 97 females, 2 unknown) over a 4.5-year period (June 2010-November 2014) and found that white sharks did not randomly co-occur with their conspecifics, but formed four distinct communities. Tendency to co-occur varied across months with males co-occurring with more individuals than females. Sex-dependent patterns of visitation at the Neptune Islands and resulting intraspecific competition likely drive the observed community structure and temporal variability in co-occurrences. This study provides new insights into the aggregatory behaviour of white sharks at a seal colony and shows for the first time that white shark co-occurrence can be non-random.

Résumé: 1. The behaviour of individuals affect their distributions and is therefore fundamental in determining ecological patterns. While, the direct observation of behaviour is often limited due to logistical constraints, collection of movement data has been greatly facilitated through the development of bio-logging. Movement data obtained through tracking instrumentation may potentially constitute a relevant proxy to infer behaviour. 2. To infer behaviour from movement data is a key focus within the “movement ecology” discipline. Statistical learning constitutes a number of methods that can be used to assess the link between given variables from a fully informed training dataset and then predict the values on a non-informed variable. We chose the random forest algorithm for its high prediction accuracy and its ease of implementation. The strength of random forest partly lies in its ability to handle a very large number of variables. Our methodology is accordingly based on the derivation of multiple predictor variables from movement data over various temporal scales, to capture as much information as possible from changes and variations in movement. 3. The methodology is described in four steps, using examples on foraging seabirds and fishing vessels for illustration. The models showed very high prediction accuracy (92%-97%), thereby confirming the influence of behaviour on movement decisions and demonstrating the ability to derive multiple variables from movement data to predict behaviour with random forests. 4. The codes developed for this methodology are published in the “M2B” (Movement to Behaviour) R package, available at https://CRAN.R-project.org/package=m2b. They can be used and adapted to datasets where movement was sampled from a wide range of taxa, sampling schemes or tracking devices. Observations are needed for a subset of the data, but once the model is trained, it can be used on any dataset with similar movement data.