Bayesian networks to explain the effect of label information on product perception

Interdisciplinary approaches in food research require new methods in data analysis that are able to deal with complexity and facilitate the communication among model users. Four parallel full factorial within-subject designs were performed to examine the relative contribution to consumer product evaluation of intrinsic product properties and information given on packaging. Detailed experimental designs and results obtained from analyses of variance were published [1]. The data was analyzed again with the machine learning modelling technique Bayesian networks. The objective of the current paper is to explain basic features of this technique and its advantages over the standard statistical approach regarding handling of complexity and communication of results. With analysis of variance, visualization and interpretation of main effects and interactions effects becomes difficult in complex systems. The Bayesian network model offers the possibility to formally incorporate (domain) experts knowledge. By combining empirical data with the pre-defined network structure, new relationships can be learned, thus generating an update of current knowledge. Probabilistic inference in Bayesian networks allows instant and global use of the model; its graphical representation makes it easy to visualize and communicate the results. Making use of the most of data from one single experiment, as well as combining data of independent experiments makes Bayesian networks for analysing these and similarly complex and rich data sets