1
00:00:08,986 --> 00:00:13,470
Another heuristic which Edward Tufte
introduces us to is called chartjunk.

2
00:00:13,470 --> 00:00:16,980
Now Tufte is much more damning of
chartjunk than he is of other forms

3
00:00:16,980 --> 00:00:18,530
of non-data ink.

4
00:00:18,530 --> 00:00:22,920
Indeed, he suggests that artistic
decorations on statistical graphs are like

5
00:00:22,920 --> 00:00:23,970
weeds in our data graphics.

6
00:00:25,200 --> 00:00:28,120
He suggested there's really
three kinds of chartjunk.

7
00:00:28,120 --> 00:00:30,500
The first is unintended optical art.

8
00:00:30,500 --> 00:00:33,770
For instance, excessive shading or
patterning of chart features,

9
00:00:33,770 --> 00:00:37,650
such as shown in this economics graphic
which Tufte shares in his book,

10
00:00:37,650 --> 00:00:40,790
The Visual Display of
Quantitative Information.

11
00:00:40,790 --> 00:00:44,745
Here, the patterns make the human
eye jump and cause visual fatigue.

12
00:00:44,745 --> 00:00:47,270
These phenomenon
are called moiré patterns.

13
00:00:47,270 --> 00:00:51,435
And this is the same reason you don't
usually see people wearing stripe shirts

14
00:00:51,435 --> 00:00:53,769
when filming video content like this MOOC,

15
00:00:53,769 --> 00:00:57,383
since the low resolution of the video
accentuates the moiré issue.

16
00:00:58,625 --> 00:01:01,790
Tufte suggests that instead
of patterning this content,

17
00:01:01,790 --> 00:01:05,410
one is better off labeling
the chart graphics directly.

18
00:01:05,410 --> 00:01:09,338
We saw that being used in the data ink
ratio example by Darkhorse Analytics.

19
00:01:10,710 --> 00:01:13,912
The second form of chartjunk is the grid.

20
00:01:13,912 --> 00:01:17,100
Tufte suggested the grid is both
unnecessary as data ink, but

21
00:01:17,100 --> 00:01:20,000
also causes competition with
the actual data being shared.

22
00:01:21,020 --> 00:01:26,140
Thinning, removing, or desaturating grid
lines makes it easier to see the data,

23
00:01:26,140 --> 00:01:28,920
instead of being overwhelmed by
the number of lines on the page.

24
00:01:30,080 --> 00:01:34,255
Direct labeling of data is another great
way to reduce this form of chartjunk.

25
00:01:35,870 --> 00:01:39,690
But it's the third form of chartjunk
that I want to focus on here.

26
00:01:39,690 --> 00:01:43,855
It's the one which typically comes to
mind when talking about chartjunk.

27
00:01:43,855 --> 00:01:46,410
Tufte calls this the duck.

28
00:01:46,410 --> 00:01:50,009
Broadly, he's referring to
non-data creative graphics,

29
00:01:50,009 --> 00:01:54,520
whether they be line art or photographs
and they're inclusion in the chart.

30
00:01:55,630 --> 00:02:00,498
Newspapers and news magazines are a place
where this kind of imagery is often used.

31
00:02:00,498 --> 00:02:05,210
One well-known graphic artist,
Nigel Holmes, has used the duck to

32
00:02:05,210 --> 00:02:10,300
display data in a way which is memorable
yet aesthetically interesting.

33
00:02:10,300 --> 00:02:14,284
One of the most memorable images
he's created, to me, is entitled,

34
00:02:14,284 --> 00:02:16,454
Diamonds Were a Girl's Best Friend.

35
00:02:16,454 --> 00:02:19,760
And this showed up in
a Times Magazine in 1982.

36
00:02:19,760 --> 00:02:23,935
This graphic shows the trend of the price
of diamonds from 1978 to 1982, and

37
00:02:23,935 --> 00:02:29,855
while the dollar amounts
are easily forgotten, it's easy

38
00:02:29,855 --> 00:02:33,845
to remember that the trend have a spike
because of the shape of the woman's leg.

39
00:02:35,335 --> 00:02:38,095
So is the doc really a useful heuristic,
or

40
00:02:38,095 --> 00:02:41,150
is it something about
docs which are memorable?

41
00:02:41,150 --> 00:02:43,780
I was part of a team led by Scott Bateman

42
00:02:43,780 --> 00:02:46,220
who wanted to understand
this issue in more detail.

43
00:02:46,220 --> 00:02:48,720
So we set up some user testing.

44
00:02:48,720 --> 00:02:51,610
I've linked the full academic
paper in this week's readings.

45
00:02:51,610 --> 00:02:54,950
In short, we provided participants
a variety of homes as image,

46
00:02:54,950 --> 00:02:59,810
including the one which you were shown,
which could be considered chartjunk.

47
00:02:59,810 --> 00:03:03,910
We also included a variety of plain
graphs with high data-ink ratios,

48
00:03:03,910 --> 00:03:06,610
with no decorative embellishments.

49
00:03:06,610 --> 00:03:09,160
We brought 24 subjects
into the laboratory and

50
00:03:09,160 --> 00:03:14,120
assigned them to either view the Holmes or
high data-ink ratio conditions.

51
00:03:14,120 --> 00:03:15,980
We asked the subjects to describe and

52
00:03:15,980 --> 00:03:19,980
summarize the charts with
a series of guided questions.

53
00:03:19,980 --> 00:03:21,800
This was followed with a recall test,

54
00:03:21,800 --> 00:03:26,920
both conducted immediately, as well as
a second test two to three weeks later.

55
00:03:26,920 --> 00:03:30,510
Finally, we used eye tracking to determine
where subjects were paying attention

56
00:03:30,510 --> 00:03:32,650
when they looked at these charts.

57
00:03:32,650 --> 00:03:34,630
Our findings were interesting.

58
00:03:34,630 --> 00:03:37,730
While there was no recall
difference when tested immediately,

59
00:03:37,730 --> 00:03:40,900
recall after two to three weeks
was significantly better for

60
00:03:40,900 --> 00:03:44,620
the Holmes Charts,
which showed the duck chartjunk.

61
00:03:44,620 --> 00:03:48,824
Further subjects indicated subjectively
that the Holmes Charts were more

62
00:03:48,824 --> 00:03:52,025
enjoyable, more attractive,
easier to remember, and

63
00:03:52,025 --> 00:03:53,838
easier to remember details of.

64
00:03:53,838 --> 00:03:58,580
And they felt it was faster to describe
and remember the Holmes Charts.

65
00:03:58,580 --> 00:04:02,480
The non-embellished designs didn't perform
better than any of the Holmes designs did.

66
00:04:03,940 --> 00:04:07,550
Does this mean that you should use
embellishments in your charts?

67
00:04:07,550 --> 00:04:08,600
Well, maybe.

68
00:04:08,600 --> 00:04:11,480
I don't know of anyone who has
replicated our findings on a larger and

69
00:04:11,480 --> 00:04:15,520
more diverse population, and there are
certainly other kinds of chartjunk, like

70
00:04:15,520 --> 00:04:20,080
the unintended optical art in the grid
which seem like really good heuristics.

71
00:04:20,080 --> 00:04:23,610
But I think there's more to be told in the
story and more nuance to be worked out.

72
00:04:24,930 --> 00:04:28,440
At this point, you should have, at least,
some thoughts from designers as well as

73
00:04:28,440 --> 00:04:32,110
from a user study on
specific kinds of charts.

74
00:04:32,110 --> 00:04:35,640
And as you go about creating your
data science graphics, it's worth

75
00:04:35,640 --> 00:04:39,760
not only reflecting on the principles you
use and the results you are sharing, but

76
00:04:39,760 --> 00:04:42,664
also the process by which you
came up to create the graphics.

77
00:04:43,900 --> 00:04:47,973
Setting up a user's study is much simpler
today with crowd sourcing services,

78
00:04:47,973 --> 00:04:51,300
such as CrowdFlower and
Amazon Mechanical Turk.

79
00:04:51,300 --> 00:04:55,180
It's now very reasonable for the average
data scientist to test out whether

80
00:04:55,180 --> 00:04:58,310
more embellished visuals
might be more effective.

81
00:04:58,310 --> 00:04:59,930
Whether it's in terms of time,

82
00:04:59,930 --> 00:05:03,835
memorability, accuracy
than those with data-ink.

83
00:05:05,300 --> 00:05:09,270
In the next lecture, we're going to go on
to touch on two more items from Tufte,

84
00:05:09,270 --> 00:05:10,900
the Lie Factor and sparklines.