In more recent work, cows may have been motivated to access limited feed bunk space on commercial dairies or to obtain prime pasture . In this study, as all animals were locked following milking to facilitate feeding treatments and health checks, cows would have had ample access silage regardless of queue position. Alternatively, Rathore suggested greater intermammary pressure might motivate high yielding animals to be milked earlier. As this herd was milked three times daily, however, this biological driver may also have been attenuated. Indeed, among modern studies with herds milked thrice daily, Polikarpus et al. found no significant correlation and Grasso et al. also found high yielding cows frequented the rear of the queue. Ultimately, as yield is influenced by a wide range of health and management factors, any number of confounding variables might be implicated in explaining this somewhat unexpected result. In this study a significant linear association between ageand entry position was not found. Recent work by Berry and McCarthy , which identified a nonlinear trend across parity, and by Grasso et al. , which highlighting significant interactions of parity with other biological drivers of queue position, suggests that a linear effect may not adequately capture the underlying biological relationship. A larger and more structured sample may be necessary to bring more complex age dynamics into clearer resolution.Visual inspection of means plots produced from mixed model analysis of sensor records recovered only a handful of statistically significant differences between queue quartiles when hour and day effects were assessed individually, trimming trays but several global trends were still readily visible.
With respect to minutes recorded as active, the 1st-3rd queue quartiles were visually indistinguishable in their cyclical behavioral patterns, but cows in the fourth queue quartile were consistently more active, particularly during the night and morning lounging period. With respect to longitudinal trends across days, fourth queue quartile animals were again more active across the observation window, whereas cows in the first queue quartile were consistently the least active. These patterns were somewhat mirrored in the longitudinal and cyclical analysis of high activity minutes, but the pattern was both less distinct and less consistent. No clear qualitative insights could be drawn for cyclical or longitudinal patterns in nonactivity. Cyclical patterns in minutes spent eating were not seen overnight or in the afternoon, but first queue quantile cows may have spent slightly more time eating after the morning milking. Longitudinal analysis of eating patterns suggested cows in the fourth queue quartile spent relatively less time eating, whereas the cows in the first and second queue quartile consistently spent more time at the bunk. This contrasted with longitudinal results for minutes spent ruminating, where the cows in the second queue quartile were consistently low. No clear distinctions between groups were recovered in cyclical rumination patterns. Temperature patterns were, surprisingly, the most visually distinct of all the sensorparameters. Cows in the first queue quartile were consistently lower in body temperature in both the longitudinal and cyclical time dimensions as compared with the remainder of the herd.
While the preceding analyses revealed few statistically significant differences at individual time points, collective analysis of days and subsets of the 24-hour management cycle would uundoubtedlyreturn statistically significant differences for the broader qualitative trends visually identified via mean plots. Within a linear modeling framework, however, this constitutes no small task. Group-by-date interaction effects were also significant for activity, high activity, and temperature models . This suggests that these models should not be simplified to a single cyclical or longitudinal trend, which would allow overall differences between groups to be tested via a single group intercept term. Targeted hypotheses comparing comprehensive trends between groups would instead require formulation of linear contrasts ± a daunting task with so many fixed effects terms used to accommodate the high sampling frequency and extended observation period of this dataset. Further, as with the linear models with cow attributes, behavioral synchronization due to social cohesion or compensatory use of physical resources in the pen could again create non-independence between animals in such sensor records. Any such issues in estimation of model degrees of freedom, compounded with the inability to fit behaviorally and empirically compelling correlation and variance models, would only serve to further confound the estimation of appropriate p-values from these models. Fortunately, the qualitative trends identified via the preceding means plots largely aligned with the significant bivariate associations identified by mutual conditional entropy tests summarized in Table 2. Activity again proved to be the most distinctive behavioral axis. Significant associations were identified for all three lounging periods when analyzed both independently and in aggregate, with the afternoon lounging periods being the most distinct. High activity also showed a significant relation to queue records, but this association may have been driven predominantly by overnight lounging period. Whereas no clear qualitative patterns were identified for non-activity data via the means plots, a significant association with queue records was identified during the afternoon lounging period.
A highly significant relationship was identified for time spent eating for the full sensor record, but given that time budgets recorded by this platform were segmented somewhat arbitrarily at the start of each hour, this result may simply reflect a lag in the arrival of cows to the feed bunk after exiting the parlor. Significant associations were not found during the lounging periods at the standard 0.05 cutoff, though records from the afternoon lounging period approached significance. These results were mirrored in rumination patterns, where again no significant association was recovered, but the afternoon lounging period approached significance. Finally, as with the linear modeling results, temperature proved highly distinct between queue subgroups for all subperiods.Visual inspection of tube plots produced with median queue subgroup values again yielded insights comparable to the linear modeling results . Based on the results of the MCE tests, all behavioral axes cows were clustered into two subgroups based on queueing records, with Group 1 cows consisting of 80 animals at the front of the queue, and Group 2 cows constituting the 34 animals in the rear. Tube plots of minutes spent active revealed Group 2 cows to be more active across all three lounging periods. This pattern was the most consistent in the morning and overnight lounging periods, though this difference was ultimately quite subtle and seldom constituted more than a few minutes. In the afternoon subperiod there was evidence of several periods with anomalously high activity levels, most of which occurred post-pasture access. The significant association recovered for minutes spent highly active in the overnight subperiod appeared to be largely driven by increased activity immediately following the evening milking, trim tray which could reflect divergent home pen behaviors, but might also have been driven by delays in milking. To complement these results for active and high active minutes, the significant association for afternoon non-activity records appears to have been driven by increased non-activity among the Group 1 cows during the three hours immediately preceding the night milking. As anticipated, differences in time spent eating were largely restricted to the 2-3 hours immediately following milking. Cows only lingered at the feed bunk during the morning lounging period, where median eating times for Group 1 cows were perhaps slightly higher. Similarly, differences in rumination also appeared restricted to time periods immediately following milking, with no clear differences seen during the lounging period with this coarse stratification of animals.
Finally, as with the mean plots, body temperature values again produced surprisingly distinctive results. More finely segmented into five queue groups by the mutual conditional entropy test, the tube plots proved a slightly cumbersome means of comparing temperature records, but a clear visual distinction could still be made between the Group 2 animals and the remainder of the herd. For all three lounging periods, this relatively small cluster of 17 cows that constituted the very front of the milking queue demonstrated lower median body temperature values, a distinction seen most clearly at night.The strong agreement between the results of these two analytical pipelines suggests that UML and conventional linear modeling approaches could be used interchangeably or in concert to glean preliminary insights from exploratory analyses of large sensor-based datasets that may inform future hypothesis-driven studies. Perhaps the most surprising result of these analyses, that cows frequenting the back of the queue are consistently more active between milkings, may indeed warrant further exploration. In much of the prior literature, health challenges that impede movement have been identified as the main driver of delayed entry into the parlor . In fact, this mechanism is so well-established that it has even been proposed that milk order records might be incorporated into genetic evaluations to improve estimates of health traits . As these analyses were run on the subset of animals with no recorded health events, however, it is possible that this dataset has brought other behavioral mechanisms into focus. Previous studies have found that animals of low social status frequent the rear of the herd in voluntary movements , and social dominance is known to impact resource access in spatially constricted conditions such as those found at the entrance to the milking parlor. If low dominance animals are in turn also forced to wait longer or walk farther to access resources in the home pen, this could potentially explain the increased activity levels of animals found in the rear of the queue. While the early literature has found the relationship between dominance value and milking order to be tenuous at best , it is possible that such social mechanisms may have been confounded by health status, with linear analyses of limited sample size failing to disentangle these mechanisms in non-disaggregated data.On this farm, where resources are not severely restricted and animals are frequently remixed, energetic investments in a dominance hierarchy may offer few returns . Such a behavioral strategy might also explain why it is high-yielding multiparous cows and not heifers that occupy the end of the queue. Both these hypotheses are ultimately purely speculative interpretations of these exploratory results; however, if proposals to incorporate milk order records into genetic indices are progressed, any correlations between queueing position and consistent individual differences in home pen behaviors likely warrant closer inspection to mitigate the risk of unintended and potentially deleterious selection pressures.As with previous studies of milk order records, these analyses perhaps raise more questions than answers. As dairy record management systems grow to accommodate an ever wider range data streams, perhaps future work considering more herds from a wider range of management strategies will succeed in further untangling the complex web of explanatory variables at the individual, herd, and farm levels that drive variation in queueing patterns. This dataset has, none the less, demonstrated the utility of unsupervised machine learning tools in ethological studies using sensor platforms to study larger groups of animals over extended periods of time. While these analyses recovered no evidence of social cohesion amongst large or temporally consistent subgroups, information theoretic approaches succeeded in clarifying the underlying pattern of heterogeneity in error variance between animals and also demonstrated an advantage in recovering evidence of non-uniform patterns in temporal nonstationary over basic EDA tools. After incorporating these insights into the structure of subsequent linear models, these model-free tools then showed some capacity to confirm inferential results where probabilistic assumptions were not strictly met, as well as an aptitude for recovering significant associations not captured by a simple linear effect. This flexible clustering-based approach to identifying significant bivariate associations was then easily extended to accommodate two high dimensional behavioral axes, providing equivalent insights to more computationally taxing mixed effect models. While UML approaches are by no means infallible, as seen here with artifacts produced by the spectral embeddings, these analyses have demonstrated that such tools can add value at every stage of the standard hypothesis-driven linear analysis pipeline, and may even offer an advantages over model-based approaches in early-stage exploratory projects. While many new methodological developments are doubtless on the ethological horizon, we hope this algorithmic toolset will provide a meaningful step forwards to meet the challenges of a future defined by ever larger and more complex data.Precision Livestock Farming technologies produce prodigious amounts of data . While the behaviors encoded by such sensors are often much simpler than those that can be quantified by a human observer, the measurement granularity and perseverance provided by these technologies create new opportunities to study complex behavioral patterns across time and in a wider range of contexts.