It makes me think about what I value and what's actually valuable to society at large and to specific groups. I don't generally value things like graphs or lines of best fit too highly. They feel a bit superficial. I understand that they represent data to enable understanding, but I don't worry about the details too much. The value of those lines are also a bit subjective.
The research in part considers the impact of adjusting a parametre α used by the techniques. What is an ideal parametre. Parametres like that bother me a little. I don't doubt their importance and their necessity. I understand why people often want to find a "best" value for it that can be generally used, but I suppose I feel it will never be completely true. There can't be a magic value for parametres like that because they're often context specific. Ultimately, you just want a value that results in an answer that feels best. What intuitively is most useful. The parametre can be "wrong" almost objectively: in this case, choose one of the extremes of 0 or 1, and you end up with either useless smoothing (that is, no smoothing, so no trend is illustrated) or total smoothing (a straight line, which rashly treats almost everything as noise). So you want good correlation (r²) without just mimicking the data.
To me, I'd be more interested in analysing and understanding the algorithms and equations behind the two goodness-of-fit measures and articulating the contribution of α from there, using that as a way to determine useful considerations in selecting α. Rather than recommending a default α or range for α, and letting users of the techniques pick one that feels "good", it would be nice to simply distill the meaning of α and what considerations a user should be making in picking one. Give α real meaning that can be justified in the resulting paper, not just "we used this value for α because it worked when we tried it". An acquaintance used a variety of software for their research and they're disturbed by how superficial their knowledge of the software and its parametres are, and the lack of guidance available for using them. (It appears documentation for scientific applications can be ... sketchy.)
But it's true, society benefits from standardised visuals, when they're even constructed meaningfully. My Experimental Design course last year emphasised the importance of properly selected and labelled graphing. The course did wonders for my appreciation for graphs. They weren't just pretty objects anymore taking up precious space or displacing deeper understanding with a shallow paraphrase. When done correctly, they were the deeper understanding succinctly demonstrated. However, so many visuals and graphs are junk. You look at them and they give you an idea, but many of them lack the most fundamental details to give that idea anything but superficial meaning. And tragically it's often simple things like scale, units, names of axes, legends, titles. It's ridiculous that you can be shown graphs even in professional material that lack these. It's ridiculous that missing trivial but fundamental parts those is something I get to question right now. I would rather question the absence of comparative data, I'd rather question the absence of sufficient context in which to understand the data in the graph more generally.
When useful data is usefully represented in a graph, in an appropriate graph with appropriate form, graphs change from a communication art to an informational science for me. (Disclaimer: I don't support the dichotomy of art versus science, though my words just encouraged it. :D) Lots of magazines and infographics display meaningless pie charts or line graphs that just sky rocket into nowhere, and when I see them, I just assume "These are pretty and are supposed to convey the gist of an idea but aren't intended on actually educating anyone in a convincing way". They have a limited usefulness in that regard. They achieve the publication's minimum goal in using them. The only thing worse would be to randomly generate data and then publish a crayon portrait of a horse by the editor's child. (I think the portrait of the horse would actually prove more useful than the decontextualised data.) I'd rather publications strive to only present information in a useful way. Sure, even most of the readers don't need the detail I want, but I'm not asking for a giant appendix of tables of values, I just want a useful graph. It's like error messages in software: sure, to most users "An error has occurred" and "Couldn't find /boot at dm-3. Please check dm-1:/etc/fstab." would be comparably useful, but the second doesn't really leave the user any more confused than the former, it just provides them with an opportunity to understand, which is denied in the first situation (they could ask a techy friend to help diagnose their problem). Similarly, a neutered graph denies interested parties the opportunity to properly understand the origin of graph.
So regarding the research, it seems strange to me because it seems concerned with arbitrary and subjective representation of data. It's definitely useful because it will help people make better use of the subject smoothing techniques, but still feels like the focus is on something subjective and aesthetic rather than helping identify something
No comments:
Post a Comment