This paper is only available as a PDF. To read, Please Download here.
The objective of this article is to review fun- damental differences between model-dependent and mod- el-free approaches to data analysis, and to explore the potential advantages of more open-ended machine learn- ing approaches in recovering complex behavioral patterns from precision livestock farming data streams.
Case studies using simulated data were de- signed to mimic a real-world scenario. Data from a feeding trial in an organic dairy were reanalyzed using the Live- stock Informatics Toolkit.
Case studies using simulated data are used to demonstrate how incomplete information about the management system can prohibit the development of an appropriate model for information compression, allow- ing aggregation bias to mask important behavioral indi- cators of compromised welfare. These hidden behavioral patterns are then recovered using unsupervised machine learning approaches that are able to leverage the intrinsic behavioral codependencies of group-housed animals. This simulated case study is then extended to demonstrate how model-based approaches can also overlook causes of com- promised welfare when the link between environmental factors and behavioral responses is strong but nonlinear, whereas model-free information-theoretic tools can easily recover and characterize such complex dynamics. Finally, in an empirical case study with data from a commercial organic dairy, the Livestock Informatics Toolkit is used to recover from milk parlor metadata complex associations between herd age structure, levels of milk production, and order of milking.
Conclusions and Applications
Model-free machine learning algorithms provide a more open-ended approach to knowledge discovery that require fewer up-front as- sumptions about the management system. This can yield more comprehensive insights into large precision livestock farming data sets now commonly encountered in on-farm research trials and in applied data auditing scenarios.
☆The authors have not declared any conflicts of interest.
© 2023 Published by Elsevier Inc.