Modeling The Trend Of Product Rating Based On The Symbolic Regression

With the rapid development of the Chinese economy and the rise of the Web2.0 era,ecommerce has sprung up like mushrooms after a rain,and online shopping has gradually become a domain way of shopping.Related research shows that more and more consumers and merchants rely on online reviews for purchasing or marketing decisions.Considering the problem of time efficiency,it is unrealistic to directly obtain and use all the review information.Obtaining information from the rating is relatively more feasible.Therefore,it is particularly important to grasp the law of the rating.Existing studies on online rating have verified the dynamic changing pattern of average rating;nevertheless,different research concluded various results.Some hold that the previous rating will have a negative effect on subsequent rating,while others consider the previous rating will positive feedback following rating.Similarly,some research has found two trends of the average rating while others found more trends.There is currently a lack of analysis and discussion on the dynamic changes of average rating of different types of products.Most of the existing studies are empirical studies,and the results obtained are mostly linear models,which can only describe the correlation between the average rating and other factors or variables,but cannot describe its true laws.In order to discover the laws of average rating for different products,we do our research based on about 8.5 million Amazon product reviews from 1996 to 2018.First of all,inspired by divide and conquer,divide the product into four levels from coarse to fine on the whole,and use a data-driven based method—symbolic regression for each independent product layer to construct a mathematical model of its average rating;Secondly,under the calculation framework of the divide and conquer,based on the Estimate Distribution Algorithm,and we propose a general model learning method for different products,which is to extract metaknowledge from a large number of mathematical models and learn a general mathematical model.Finally,based on this method,we obtained the general model of average rating for different types of products.Its robustness and superiority are verified through a large number of test data sets.The main research conclusions of this paper include the following aspects:(1)In search and experiential products at different prices,we found that average rating shows four trends in time series: increasing,deceasing,undershooting and inverted U trends,in which the inverted U trend is the new trend discovered in this paper;(2)In terms of the distribution of different trends,this article finds that the decreasing trend takes up most percent(each over 65%);(3)The symbolic regression is used to construct a mathematical model of average rating,which can more accurately describe the true law of average rating;(4)The general model of average rating for different products learned based on the divide and conquer idea and the distribution estimation algorithm show higher the r-square and the significance rate than models in existing researches.