Title: Multi-class modelling for muscle level prediction of beef eating quality
Authors: Garth Tarr - University of Sydney (Australia) [presenting]
Ines Wilms - Maastricht University (Netherlands)
Abstract: The Meat Standards Australia (MSA) beef grading system was developed to improve the eating quality consistency of Australian beef for a range of muscles and cooking methods. The MSA system aims to predict the eating quality of beef given various inputs. Data has been collected and merged from experiments conducted over the past 20 years, resulting in a highly unbalanced data set where the majority of observations have been recorded on only a small subset of muscles with a far fewer observations on the majority of other muscles. The current approach to building a MSA grading system is ad hoc and labour intensive. We present a new method of multi-class modelling using lasso-type penalties to encourage similar coefficient estimates that bridge a spectrum of muscle level models. At one extreme, each muscle is modelled independently and at the other extreme a single pooled regression model for all muscles. An ideal model is somewhere between the two extremes, where muscles with limited data are encouraged to borrow information from other similar muscles. We illustrate our method on real data and through simulation studies. The choice of tuning parameter is also discussed.