Subjects and facility
We observed two groups of Atlantic bottlenose dolphins (six different individuals in total) housed at the marine zoo “Marineland Mallorca”. One of the groups was composed of four individuals (G1) and the other was constituted by five individuals (G2). The two adult males and one of the females were the same in both groups (Table 1). Group composition changed due to the transfer of individuals to another pool of the zoo and due to the arrival of new individuals from another aquatic park.
The dolphins were kept in three outdoor interconnecting pools: the main performance pool (1.6 million liters of water), a medical pool (37.8 thousand liters of water) and a small pool (636.8 thousand liters of water). During the observational periods, the dolphins had free access to all the pools. Underwater viewing at the main and the small pool was available through the transparent walls around the rim of the pools.
Ethics statement
This study was approved by the UIB Committee of Research Ethics and Marineland Mallorca. This research was conducted in compliance with the standards of the European Association of Zoos and Aquaria (EAZA). All subjects tested in this study were housed in Marineland Mallorca following the Directive 1999/22/EC on the keeping of animals in zoos. This study was strictly non-invasive and did not affect the welfare of dolphins.
Behavioral observations and data collection
Behavioral data were collected in situ by APM from May to November 2016 for G1 and from November 2017 to February 2018 for G2. All observational periods were also recorded using two waterproof cameras SJCAM SJ4000. Observations were conducted at the main pool between 8:00 a.m. and 11:00 a.m. Due to the schedules and dynamics of the zoo, we were unable to collect data outside this period. Dolphin social behavior was registered and videotaped for 30 min–2 h each day. Only data from sessions that lasted at least 30 min were included in the analysis. We did not collect any data during training or medical procedures and resumed the observational session a few minutes after the end of these events.
We recorded all occurrences of affiliative and aggressive interactions, the identities of the involved individuals and the identity of the dolphin initiating the contact. Aggressive contacts were defined by the occurrence of chasing, biting, and hitting, as established in previous studies37,38,39,40,41. Affiliative contacts were defined as contact swimming, synchronous breathing and swimming (at least 30″ of continuous swimming) or flipper-rubbing, as established in previous studies37,39,40,41,43.
To assess the strength of the affiliative bonds in both groups, we calculated the index of affiliative relationships (IA) between dolphins following the procedure described in Yamamoto et al. For calculating the IA we recorded the relative frequencies of synchronous swimming since it is a well-defined affiliative behavior in dolphins. Data of synchronous swimming were recorded using group 0–1 sampling44 at 3-min intervals. This method consists of the observation of individuals during short periods and the recording of the occurrence (assigning to that period a 1) or non-occurrence (assigning to that period a 0) of a well-defined behavior44. For calculating the IA for each couple, the number of sampling periods in which synchronous swimming between individuals A and B occurred (XAB) was divided by the number of sampling periods in which individuals A and B were observed (YAB): (IA=frac{{X}_{AB}}{{Y}_{AB}})39,45. Therefore, the IA reflects the level of affiliation for each dolphin dyad based on the pattern of synchronous swimming. This index served to construct the general affiliative social networks of both groups of dolphins.
Temporal network construction
Temporal networks can provide insight into social events such as conflicts and post-conflict interactions in which the order of interactions and the timing is crucial. Furthermore, they allow us to calculate the probabilities of the different affiliative and aggressive interactions occurring in the group.
We used behavioral observations to construct temporal networks for each group. Each dolphin was treated as a node (N) with their aggressive and affiliative interactions supplying the network links. We divided the daily observations into periods of 3 min. In each period, we assigned a positive (+ 1), negative (− 1) or neutral (0) interaction to each pair of dolphins. That is, if during the period a pair of dolphins displayed affiliative interactions, we assigned a + 1 to the link between that pair of nodes, if they were involved in a conflict, we assigned a − 1, and if the pair did not engage in any interaction, we assigned to that link a 0. If during the same period, the pair displayed both aggressive and affiliative interactions we considered the last observed interaction. Therefore, we obtained an adjacency matrix (an N × N matrix describing the links in the network) for each group of dolphins. Thus, for each day we had a series of different signed networks of the group, each network representing a 3-min period.
Social network analysis: time-aggregated networks and network motifs
We collapsed the temporal networks of each day in time-aggregated networks. This procedure consists in aggregating the data collected over time within specific intervals to create weighted networks. The sign and the weight of the links characterize these networks, indicating the valence and duration of the interaction respectively. Thus, they are static representations of the social structure of the group of dolphins. To obtain these time-aggregated networks we proceeded as follows:
First, for each day we aggregated the values of each interaction of the temporal networks until one link qualitatively changed. We considered a qualitative change if one interaction passed from being negative (− 1) to positive (+ 1) meaning that the pair of dolphins reconciled after the conflict or vice versa, or if a new affiliation (+ 1) or aggression (− 1) took place, that is the link changed from being neutral (0) to positive or negative. If a link changed from being negative or positive to being neutral, we did not consider that this interaction has changed qualitatively. For example, if dolphins interacted positively during two periods of time, then they ceased to interact (neutral) and finally they engaged in an aggressive interaction, the total weight of the interaction in the resulting time-aggregated network would be of + 2. Therefore, a conflict or an affiliation may extend over multiple periods containing several contacts, and is considered finished when the interaction changes its valence. In this way, we obtained a series of time-aggregated networks for each day, which retain the information on the duration, timing, and ordering of the affiliative and aggressive events in the group.
We examined the local-scale structure of the affiliative-aggressive social networks using motif analysis. Thus, for each group, we analyzed the network motif representation of the temporal and time-aggregated networks, identifying and recording the number of occurrences of each motif.
Model of affiliative and aggressive interactions
We built two models (a simple and a complex one) that aim to simulate the dynamics of aggressive and affiliative interactions of a group of four dolphins. These models were created using the observed probabilities of each affiliative or aggressive interaction between individuals in group G1. We only used the data of G1 since we had more hours of video recordings and, thus, more statistics of the pattern of dolphins’ interactions. Both models return affiliative/aggressive temporal networks constituted by four nodes and different aggressive, affiliative, or neutral interactions between the six possible pairs of individuals in the network. We simulated data for 20 periods of 3 min per day for a total of 80 days to mimic the empirical data time structure. We obtained one temporal network for each period (1600 temporal networks in total) and ran 100 realizations of each model.
Our models work as follows: At the beginning of the simulations, all the interactions between the four nodes are neutral (0). In each period, we select a pair of nodes randomly and assign to that link a positive (+ 1) or a negative (− 1) interaction with probability p (calculated previously for each type of interaction). These interactions correspond to spontaneous aggressions and affiliations. In the complex model, if in the previous period a conflict took place, before assessing spontaneous interactions we first evaluated the different possible post-conflict contacts that could occur (reconciliation, new aggressions, and affiliations). Therefore, for reconciliations, we change the valence of the interaction from negative to positive with a certain probability. Then, we also randomly choose a pair of nodes including one of the former opponents and assign to that link a positive or negative interaction with the observed probabilities to simulate the occurrence of new affiliations (third party-affiliation) or redirected aggressions arising from the previous conflict. We keep on doing this procedure period by period. Lastly, we obtained the time-aggregated networks for the two models.
The simpler model only includes the probability of aggression and affiliation between group members, whereas the complex one also includes the patterns of conflict resolution previously observed. In this way, the complex model serves to assess the influence of post-conflict management mechanisms on the observed pattern of aggressive/affiliative networks. That is, the complex model also keeps track of past actions. Thus, depending on the interaction of the previous step, the probability of the following interaction changes based on the observed pattern of conflict resolution strategies.
Calculation of the observed probabilities of affiliative and aggressive interactions
For the simple model, we calculated the probability of general aggression and affiliation per day without distinguishing between types of positive and negative interactions. Thus, we obtained the number of periods in which an aggressive or affiliative contact took place per day and divided it by the total number of periods of that day (probability of general aggression or affiliation per 3-min period). With these probabilities, we calculated the mean probability of general aggression and affiliation per period.
For the complex model, we calculated the probabilities of reconciliation, new affiliations/aggressions, and spontaneous affiliations/aggressions per day. That is, the probability that former opponents exchange affiliative contacts after an aggressive encounter (reconciliation), the probabilities that a conflict may promote new affiliations (third-party affiliation) or new conflicts (redirected aggression) between one of the opponents and a bystander in the same day, and the probability of affiliative or aggressive interactions not derived from a previous conflict (spontaneous interactions). To classify affiliations and aggressions in these categories we used the temporal networks, examining the interactions that took place after a conflict between opponents and between them and bystanders. If the opponents reconciled or affiliated with a bystander after a fight, we assumed that the following affiliative or aggressive interactions were spontaneous and were not a consequence of that conflict. Thus, to calculate the number of spontaneous affiliations, we subtracted the number of reconciliations and new affiliations from the total number of affiliations per day. For spontaneous aggressions, we subtracted the number of new aggressions to the total number of aggressions per day. Then, we obtained the probability of spontaneous affiliation and aggression per period.
Using the previous probabilities, we obtained the rate (r) of reconciliation, new aggression and new affiliation per minute with the following formula:({p=1-e}^{-rDelta t}). Using the same formula, we finally calculated the probability of reconciliation, new aggression and affiliation per 3-min period used in the complex model (Supplementary Table 1 for details of probabilities calculation).
Network-motif analysis
We also carried out a network-motif analysis. As we did not consider the identities or sex of the nodes in these models, we grouped the obtained motifs into equivalent categories considering the pattern of interactions between nodes. We also classified the motifs obtained from the real data of G1 into those equivalent categories. Finally, we compared the pattern of equivalent network motifs of the observed social network of dolphins and the ones of the two models. To do so we calculated the Spearman’s rank correlation coefficient (rs), defined as a nonparametric measure of the statistical dependence between the rankings of two variables: ({r}_{s}=frac{covleft({rg}_{X}{rg}_{Y}right)}{{sigma }_{{rg}_{X}}}{sigma }_{{rg}_{Y}}); rgX and rgY are the rank variables; cov (rgX rgY) is the covariance of the rank variables, and σrgX and σrgY are the standard deviations of the rank variables. Therefore, this coefficient allows us to assess the statistical dependence between the motif ranking of the real data and the one of each model.
Computational implementations
All the models, network construction, visualization and motif analysis were generated and implemented using MATLAB R2018b.
Source: Ecology - nature.com
