STABLE ISOTOPE MIXING MODEL EVALUATION: QUANTIFYING THE QUALITY OF PREDICTIONS

Stable isotope technique is of importance to study the ecology of food webs. Based on stable isotope mixing model of mass balance, stable isotope technology can be used for consumer nutrition traceability, that is, to determine the contribution of multiple sources of nutrition to the consumer. Stable isotope mixing model of mass balance has been one of the necessary methods for the traceability analysis of consumer nutrition sources. Bayesian mixing models are often used to estimate the contribution of different sources of nutrition. Such models provide probabilistic distribution characteristics of each nutrient source’s contribution to the consumer. However, the result of mixing model fitting and its matching level with the actual ecology theory are important evaluation contents of model performance. In order to ensure the accuracy of the modeling analysis, the modeling data must be corrected and verified first. Second, before data modeling, important prior information must be considered. Furthermore, the process of model selection and model evaluation for complete reproduction is a necessary condition for modeling, training, verification and evaluation. Model selection is to select the best model based on a set of model representations with different complexities and model evaluation is to evaluate the predicted error after selecting the model. According to the specific research, there are various evaluation indexes in practice, and the information loss of the relative “real model” is described respectively. Due to the unknown nature of the real model, these evaluations only reflect the relatively good performance of existing models in the construction process, so specific problems still need to be analyzed. Based on the measured isotope data set (isotope data set for Culter mongolicus mongolicus), this paper constructed a series of Bayesian models by identifying the characteristics of consumer nutrition functional groups and changing the a priori information characteristics of nutritional sources; and described the methods and processes of model performance evaluation by comparing the overall performance of the model, the difference between measured and predicted values, and the difference between prior and post-test information, so as to provide a model performance evaluation system for the application of stable isotope technology to carry out consumer nutrition traceability research. The fit quality of the model can be judged by the evaluation method, which focuses on the model’s ability to predict consumer isotope values. In addition, in view of the characteristics of the Bayesian mixing model, that is, if a priori information error is low, the mixed posterior distribution information of the model will converge to a priori information, and further evaluation will be based on information theory and probability distance statistical method to provide complementary assessment method for the quality of the output for the isotopic mixture model. The integrated use of these methods further improves the consumer nutrition source accuracy, and provides a scientific support for a more profound understanding of the food web laws. This paper reviews the best practices for fitting and evaluating Bayesian mixing model, and how to directly avoid many technical issues involved in isotope construction in consumer nutrition traceability analysis.