Heterogeneity Of Expert Evaluation

Tobacco industry is an important support for the economic and social development of Yunnan Province and the main source of fiscal and tax revenue.It is an important pillar industry to promote the sustainable and healthy development of the province's economy for a long time at present and in the future.In the tobacco industry,sensory evaluation and smoking technology has become an important technical means in many aspects,such as the cultivation of new tobacco varieties,the research and development of new cigarette products,the replacement of formula ingredients,quality management,market prediction,etc.,which can solve complex physiological sensation problems that cannot be solved by general physical and chemical analysis,with strong practicability,high sensitivity and reliable results.As the final consumption form of cigarette products is smoking,the style characteristics and flavor quality of cigarettes can provide consumers with direct feelings after burning,so the sensory evaluation results of cigarettes can best represent the will of consumers.However,there is also a problem that can not be ignored in the process of sensory quality evaluation of cigarettes: because there are differences in evaluation scale,capture degree of various sensory quality indicators,smoke tolerance,sensory threshold and sensitivity among evaluation experts,it is impossible for different experts to compare the evaluation results of different samples of cigarettes horizontally.The direct comparison of the evaluation results will not truly reflect the real differences between the samples,and the direct comparison of the evaluation results will cause a large error,which is largely due to the heterogeneity between the evaluation experts.This paper analyzes the heterogeneity of evaluation behavior from three aspects: experts' evaluation preference,experts' sensitive indicators and evaluation results' horizontal comparability.According to the historical scoring data,the research on the evaluation behavior of experts has certain application value and objective practical significance for tobacco manufacturers and enterprises in the selection of evaluation experts and the measurement of evaluation results.In this paper,the sensory quality evaluation data of several cigarette manufacturers were analyzed.The basic statistics is used to simulate the filling of a small number of missing scores,so that there is an intersection of evaluation products among experts.Then,the sparse data is filled based on collaborative filtering algorithm,and the best filling method of sparse score matrix is found by combining cross validation and average absolute error,and the data is filled.Then,the association rules are used to mine the experts' evaluation under different specifications of cigarettes.According to the strong association rules of preference,the experts with high sensitivity in each sensory quality index are mined.The core idea is whether the individual decisionmaking of experts and the overall decision-making of experts tend to be the same to measure the sensitivity of experts in each index under the condition of different smoke sensory evaluation.The high sensitivity expert groups on each index are mined and in each sensory quality index experts' sensitivity of quality index are graded and ranked.Finally,from the perspectives of individual difference and the connection of experts,the support degree of frequent item set is transformed into individual weight of experts,and the average score of experts is adjusted according to the weight,and the score value is compared to judge the quality of products;the logical association between variables is obtained by using mountain climbing algorithm to build the bayesian network which relates expert scores and tested products.By using the probability density distribution of the comprehensive scores of different products to evaluate the differences of sensory quality,two methods effectively reduce the impact of heterogeneity to a certain extent,and improve the quality of comparability.