One of the common tasks faced by the U.S. Department of Agriculture (USDA) food safety analysts is to estimate the risk of observing positive outcomes of microbial tests of food samples collected at the slaughter and food processing establishments. Resulting risk estimates can be used, among other criteria, to drive allocation of FSIS investigative resources. The Activity From Demographics and Links (AFDL) algorithm is a computationally efficient method for estimating activity of unlabeled entities in a graph from patterns of connectivity of known active entities, and from their demographic profiles. It has been successfully used in social network analysis and intelligence applications. In order to test its utility in the food safety context, we treat a co-occurrence of the same strain of bacteria (in particular a specific serotype of Salmonella) in samples taken at different establishments at roughly the same time, as a link in the graph spanning all of the USDA controlled establishments. Now, given the historical patterns of linkage and the information about the distribution of the currently observed microbial positives (which make the corresponding establishments “active” in the AFDL terminology), we aim at predicting which of the remaining establishments are likely to also report positive results of tests. Even though such definition of a link produces uncertain data given that the co-occurrences of specific test results at different establishments may be purely coincidental and our analysis does not attempt to distinguish them from truly correlated instances, we expect that using this inherently noisy data in combination with demographic features of establishments, would lead to useful predictability of microbial events.
Objective
The objective of the research summarized in this paper is to evaluate utility of the AFDL in predicting likelihood of positive isolates obtained from microbial testing of food samples collected at the USDA controlled establishments.