data analysis04 morethantwovariables
DESCRIPTION
Revised version - spell checked.TRANSCRIPT
http://publicationslist.org/junio
Data AnalysisMore Than Two Variables: Graphical Multivariate Analysis
Prof. Dr. Jose Fernando Rodrigues JuniorICMC-USP
http://publicationslist.org/junio
What is it about?More than two variables determine a tough analytical problem
Although, there are multivariate graphical techniques, they cannot deal too many variables (less than 15 – 25)
http://publicationslist.org/junio
What is it about?More than two variables determine a tough analytical
problem
Although, there are multivariate graphical techniques, theycan not deal too many variables (less than 15 – 25)
http://publicationslist.org/junio
Three variablesFor example, consider the data defined by function:
푦 = 푓 푥, 푎 = 푥2 + 푎푥 −
푥2 +
푎4
that corresponds to the three-variable setting y, x, and a
One way to analyze this is by means of a surface plot:
http://publicationslist.org/junio
Three variablesHowever, the surface plot does not permit an intuitive
reading of quantitative informationAnother way is to use a two-dimensional xy plot with
multiple curves, one for each value of interest of one of thevariables; in the example, variable a is considered for values 2,1, 0, -1, -2
http://publicationslist.org/junio
Three variablesSurface plots and multiple-curve xy plots can be in
combination, one providing an aesthetically appealingoverview, the other providing fine details for values ofinterest
It is interesting to note that surface plots go against thecommonsense that 3D plots should be moreinformative than 2D plots
Yet another possibility is to project the function into thebase plane bellow the surface, using either:contour plots false-color plots
http://publicationslist.org/junio
Three variablesThe false-color plot is achieved by mapping all values of the
dependent variable y following a palette of colorsThe false-color plot for function f(x, a)
http://publicationslist.org/junio
Three variablesThe false-color plot is achieved by mapping all values of the
dependent variable y following a palette of colorsThe false-color plot for function f(x, a)
False-color plots are very effective for presenting quantitative information
It is important to note, however, that its efficiency depends heavily on the color mapping, which must be
intuitive according to the task at hand
For an overview of color-mapping guidelines, see the textbook on page 104
http://publicationslist.org/junio
MoreThere are basically two ways to get more information on a
plotPut similar graphs next to each other and vary the
variables in a systematic fashion from one subgraph to the nextmultiplots
Make the graph elements themselves richer with color, shape,and interaction
MultiplotsThe most common forms of multiplots are the scatter-plot
matrix, and the co-plot,
http://publicationslist.org/junio
Scatter-plot matrixThe scatter-plot matrix is constructed considering all the
possible two-variable combinations achieved from theset of variables
For each combination, a sub-region of the space isreserved and all the combinations are put together accordingto a straight layout
The more variables, the bigger must be the screen,limits start to manifest around 10 variables, the same for thenumber of data points, limited around 100
For example, consider a 250 wines data set consisting ofseven different properties: acidity, sugar, chlorides, sulfurdioxide, density, alcohol, and quality
The data can be found in the “Wine Quality” data set, available at the UCI MachineLearning repository - http://archive.ics.uci.edu/ml/.
http://publicationslist.org/junio
Scatter-plot matrix
http://publicationslist.org/junio
Scatter-plot matrix
The scatter plot shows:
• sugar content and density are positively correlated• as the alcohol content goes up, density goes down, inverse
correlation• wine quality seems to increase with increasing alcohol content:
apparently, more potent wines are considered to be better
http://publicationslist.org/junio
Co-plotsCo-plots work by partitioning the data according to one of
the variables and plotting each partition in a different plot
In this example, the upperfigure shows how one ofthe variables was used topartition the data
It is possible to notice thatthe intervals overlap and havedifferent sizes so that eachplot has the same numberof points
http://publicationslist.org/junio
Co-plotsCo-plots work by partitioning the data according to one of
the variables and plotting each partition in a different plot
In this example, the upperfigure shows how one ofthe variables was used topartition the data
It is possible to notice thatthe intervals overlap and havedifferent sizes so that eachplot has the same numberof points
The partitioning can be determined in many differentways according to the goals of the analysis
http://publicationslist.org/junio
CompositionAnother way to visualize more than two variables is to
compose multiple plots according to some of the variables
For example: imagine a company that makes five productslabeled A, B, C, D, and E, and two questions:how many items of each kind are produced overallhow the item mix is changing over time
A simple solution, but not quiteeffective is to plot (days x quantity)the different curves all together
http://publicationslist.org/junio
CompositionAnother solution is to use the same plot but stacking the
information, so as to have a notion of total In absolute numbers (left), or In relative contributions (percentage)
http://publicationslist.org/junio
CompositionAnother solution is to use the same plot but stacking the
information, so as to have a notion of total In absolute numbers (left), or In relative contributions (percentage)
Surprisingly, the total number ofitems does not change appreciablyover the long run
Here, one can see the contribution ofeach product, an information thatwas lost in the left plot
http://publicationslist.org/junio
CompositionAnother solution is to use the same plot but stacking the
information, so as to have a notion of total In absolute numbers (left), or In relative contributions (percentage)
Surprisingly, the total number ofitems does not change appreciablyover the long run
Here, one can see the contribution ofeach product, an information thatwas lost in the left plot
It is desirable to use multiple kinds of plots combinedin order to verify the many aspects of data
http://publicationslist.org/junio
Parallel CoordinatesIn a parallel coordinate plot, the coordinate axes are parallel
to each otherFor every data point, its value for each of the variables is
marked on the corresponding axis, and then all these pointsare connected with lines
http://publicationslist.org/junio
Information VisualizationMany other techniques are presented according the findings of the field
known as InformationVisualization:GlyphsChernoff FacesTree-maps Star coordinatesTable LensMultidimensional Projection
And many others, all improved by means of interaction techniques:Querying and zooming Linking and BrushingCombined projections, and so forth
http://publicationslist.org/junio
References Philipp K. Janert, Data Analysis with Open Source Tools,
O’Reilly, 2010. Wikipedia, http://en.wikipedia.org Wolfram MathWorld, http://mathworld.wolfram.com/