5. Steffen Staab
staab@uni-koblenz.de
5WeST
Why do we loathe experiments?
Experiments
require effort
require time
require money
may be complicated
may involve human subjects (BAD!)
may even involve real people, i.e. not students (WORSE!!!)
have unpredictable outcomes (THE WORST!!!!!)
May have to be redone (THE NIGHTMARE!!!!!)
7. Steffen Staab
staab@uni-koblenz.de
7WeST
1 What do you talk about?
Define your key terms clearly!
Don‘t redefine existing terms!
Don‘t use figurative speech.
Negative example: Beer is a kind
of bread made from barley.
Define them independently of
how you compute them:
Negative example: We are doing this experiment to show that
Lakhsa is enjoyed by more people than Surströmming
Use the terms exactly as you defined them:
Negative example: The strömming must be enjoyed cold.
8. Steffen Staab
staab@uni-koblenz.de
8WeST
2 Formulate a hypothesis
What is your experiment about?
Only possible to define after
key concepts are clear!
As precise as possible:
...enjoy more...?
vs
...would do immediately again, if given the choice....
Briefly say what you abstract from and why.
A clear hypothesis saves time!
9. Steffen Staab
staab@uni-koblenz.de
9WeST
3 Fail-safe hypothesis
Hypothesis that is always interesting,
no matter whether the experiment succeeds or fails
comparing different methods is often interesting
Be impartial / not biased in favor of your method
AND
write in this way
not: „we want to show that our method is better“
BUT
„we want to find out whether fermentation makes the fish
more enjoyable for Swedish and non-swedish“
Tackling unpredictability!
10. Steffen Staab
staab@uni-koblenz.de
10WeST
4 Write before performing experiment
Write
The experiment
The description of the outcome
Receive input/criticism on this writing
BEFORE
you perform the experiment
Avoid having to redo experiment!
11. Steffen Staab
staab@uni-koblenz.de
11WeST
5 Reverse experiment planning
Lead question: What can make your experiment go wrong?
May not be understood / handled by test subjects
Do assume that someone can misunderstand something
Keep your original data, video recordings if possible
...
Real people are somewhat
unpredictable!
12. Steffen Staab
staab@uni-koblenz.de
12WeST
6 More is better...
Not just one data set
Do not vary arbitrarily, but vary
on what may help to explore
your hypotheses
Not just one baseline
Especially not one that is very weak,
only because it is available
• Tackling the „hard baseline“ is not impossible
• Tackling the „hard baseline“ right away may be easier
than doing 5 rounds until the paper is accepted
More....is less time!
13. Steffen Staab
staab@uni-koblenz.de
13WeST
7 When more is worse
Grounding in theory
Counterexample:
„....we investigated all the 100680 parameter configurations....“
Experiments produce outliers!
Statistical effects
Normalization exists, but is very strict
Save effort!
14. Steffen Staab
staab@uni-koblenz.de
14WeST
8 When reporting: Focus the reader‘s eye
Issue: Experiments may be complex
Hard to do
Hard to understand for the reader
Measures:
Do not overwhelm the reader
Focus the reader‘s eyes
Do not just show graphs/charts/etc.
Each graph/chart should give an answer to a question
Reader should know why he sees the chart
Manage complexity!
15. Steffen Staab
staab@uni-koblenz.de
15WeST
9 Sound experiment structure
Evaluation procedure must be independent from the method
you used-7should also be applicable for a large set of
competitors
1. Set up; i.e. data, conditions
2. Evaluation measure, baselines/competitors
3. Results
4. Discussion
Keep this order and do not discuss results before you have
described setup (maybe briefly in intro/abstract)