iSixSigma

How Do I Identify a Single Line in a Multiple-Line Graph?

Six Sigma – iSixSigma Forums General Forums General How Do I Identify a Single Line in a Multiple-Line Graph?

Viewing 4 posts - 1 through 4 (of 4 total)
  • Author
    Posts
  • #243679

    medivh
    Participant

    Hello everyone,
    I’ve recently made a few graphs of fitted curves from accumulated times for a process.
    I can’t post the graphs nor the data since they belong to my company, but I’ll try to be clear on the issue.
    When a line graph has to many curves on it, the labels (colors) will start to repeat themselves (adding symbols to the lines doesn’t help much).
    The entirety of the labels doesn’t even fit the screen. This is not a problem for the analysis per say, since we only try to identify the curves that follow different patterns (usually 1-5 in 100).
    Now, I can’t seem to find a way to identify the curves. I wish I could just hover the mouse over it and it would display the label of that curve, but the cross-hair only shows X,Y coordinates that are basically useless for me; and the select cursor highlights all curves at once.

    Is there a way to identify a single curve?
    For example, let’s say this graph attached from the support site had none or way too many labels on the screen… Is there a way we could get information on a specific curve to tell which label it belongs to?

    P.S: I don’t think adding labels when making the graphs helps either, because they are histograms with over 100 labels (but I just display the fitted curves accumulated), so it becomes a real mess when I add labels.

    Attachments:
    1. You must be signed in to download files.
    0
    #243684

    Robert Butler
    Participant

    My understanding of your post is that you have a large number of histograms which summarize elapsed time measurements.  You have taken these histograms and run a curve fitting routine to try to draw the contour of the histogram.  You have taken these contour plots and overlaid them and are now confronted with trying determine curves that tend to follow a certain shape.

    If the above is a fair description of what you are doing then a much easier approach to your problem would be a series of side-by-side boxplots.  They will give you the same information, you won’t have to worry about colors because all of the groups of data will be present.  Once you have these plots side-by-side it is an easy matter to spot similarities and differences in the boxplot patterns (you will want a boxplot routine that allows an overlay of the actual data points on the boxplot) and quickly assess such things as clustering, means, medians, standard deviations, etc.

    Attached is an example to give you some idea of what I’m describing.  The similarities and differences are readily apparent.  For example, the elapsed times for 2 and 5 years look very similar as do those for 3 and 6 whereas 4 stands alone.

    Attachments:
    1. You must be signed in to download files.
    0
    #243686

    medivh
    Participant

    Thank you Robert, but unfortunately the boxplot option does not quite solve the problem.
    I understand it is easier to see, but there’s a couple of problems with that approach as well.
    The number of labels (or ticks in the X axis) is still too high; so, for many graphs, the boxes are way too close to be distinguished, even in full screen (I also tried to zoom in to ~500x and Minitab crashed).
    The other problem is that, since the data follows a 3-parameter lognormal distribution, the boxplot shows a lot of outliers; thus it becomes a very polluted graph to show to costumers. I’m not sure if hiding the outliers, for the sake of looking nice, would be a good measure either.
    I wish there was a way to identify the odd-fitting curves just like it works for bloxpots, you hover the mouse and it shows the info on which label it is.

    I’ll attach 2 partial snapshots of both graphs. In this case, the boxplot can still be distinguished (especially since the outliers were hidden). So, I understand the odd-fitting curve corresponds to the box shown at higher values.  Is there a way to identify the curve without the need for the boxplot? We think the accumulated curves is much cleaner and easier to explain and to be understood by our workers and clients.

    I guess we could be doing both graphs and identifying the curves through the boxplot… but this is not ver practical since the boxplot graphs with more labels tend to crash when zoomed in.

    Attachments:
    1. You must be signed in to download files.
    2. You must be signed in to download files.
    0
    #243697

    Robert Butler
    Participant

    Since it sounds like the issue is that of identifying the “odd” line plot and given that the line plot you have shown is typical it would seem that a better approach would be to use the data you have to identify what looks like a “normal” envelope and then flag any curve that falls outside of that envelope.

    Just looking at the graph it would appear you have the ability to identify an acceptable range of starting points (minimum to maximum X)  and if you take a slice through the thickest part of the plot where the graphs roll over it looks like you can define a range of acceptability there as well (minimum vs maximum Y).

    A way to check this would be to take existing data – identify the unacceptable curves and see where they fall relative to what I assume is the “envelope” of acceptability.

    0
Viewing 4 posts - 1 through 4 (of 4 total)

You must be logged in to reply to this topic.