Experimental System Design And Analysis For Scatterplot Trend Perception

In big data era,data brings a lot of values.However,it is difficult for users to discover the value of data from a large amount of non-intuitive data.Visualization is an important tool for analyzing and understanding data.It can transform data into easy-to-understand images,helping users quickly and accurately discover the meanings behind the dataScatterplots are one of the most common ways for visualizing bivariate data.We rely on scatterplots for a variety of bivariate tasks including the perception of trends,correlation,and outliers.Usually,designers don't explicitly provide the trend lines or related statistical summaries.Although users can calculate the data trend through some regression models,these statistics are not always reliable.Common statistical measures of trend such as Ordinary Least Squares(OLS)is easily affected by data distribution,outliers,heteroscedasticity and other factors,so it may lead to inaccurate data modeling In fact,users obtain data trend information not only by numerical statistical methods,but also by human perception and their cognitive ability.Existing work suggests that viewers can accurately estimate trends in many standard visualizations of bivariate data.However,existing work still has some limitations.While using visual perception to estimate the trend of scatterplot,users may be influenced by data distribution,mark shapes and other factors.Hence,it is unknown whether this visual estimation is reliable in more complex scenarios,or if there are strategies by which the visualization designer can intervene in order to promote better estimationStudying how users estimate data trends in scatterplot has important theoretical significance and application value to help users quickly and accurately analyze and predict data.In order to explore the rules of human data perception,we cooperate with Tableau,a visualization design company,to investigate the effect of different data distribution and mark shapes on trend line estimation.In this paper,we generate a large amount of data with different data distribution,design and implement an experimental verification system for different types of mark shapes and data distributions.In addition,we also analyze the user results and summarize a series of perceptual rules.We find while in most cases viewers can accurately estimate trends,the use of marks with directionality that reinforces the trend direction can assist viewers in visually estimating the trend.