csc2130:( empirical(research(methods(for(computer( sciensts(sme/csc2130/01-intro.pdf · this...
TRANSCRIPT
1
1
University of Toronto Department of Computer Science
© 2004-5 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license.
CSC2130: Empirical Research Methods for Computer
Scien<sts
Steve Easterbrook [email protected]
Barbara Neves [email protected]
www.cs.toronto.edu/~sme/CSC2130/
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 2
Course Goals ➜ Mo<vate the need for an empirical basis for research claims
➜ Prepare students for advanced research: Ä Learn how to plan, conduct and report on empirical inves<ga<ons. Ä Understand the key steps of a research project:
Ø formula<ng research ques<ons, Ø theory building, Ø data analysis (using both qualita<ve and quan<ta<ve methods), Ø building evidence, Ø assessing validity, Ø publishing.
➜ Cover the principal empirical methods applicable to human subjects studies in CS Ä controlled experiment, case studies, surveys, archival analysis, ac<on research,
ethnographies,…
➜ Relate these methods to relevant meta-‐theories in the philosophy and sociology of science.
2
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 3
Intended Audience
➜ This is an advanced course: Ä assumes a strong grasp of the key research ques<ons in your own research area, and that
you are already doing independent research
➜ Focus: Ä How do people use computer technology? Ä How does this technology (re-‐)shape human ac<vi<es? Ä How can we apply qualita<ve and quan<ta<ve techniques from the behavioural sciences
to help answer these ques<ons?
➜ The course is aimed at students who: Ä …plan to conduct research (in SE, HCI, etc) that demands some empirical valida<on Ä …wish to establish an empirical basis for an exis<ng research programme Ä …wish to apply these techniques in related fields (e.g. Cog Sci, )
➜ Note: we will *not* cover the kinds of experimental techniques used in CS systems areas, nor in medical/biological research
Ä Focus is on the rela<onship between human ac<vity and computer technology
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 4
Format
➜ Seminars: Ä 1 three-‐hour seminar per week Ä Mix of discussion, lecture, student presenta<ons
➜ Readings Ä Major component is discussion of weekly readings Ä Please read the set papers before the seminar
➜ Assessment: Ä 10% Class Par<cipa<on Ä 20% Oral Presenta<on – introduce the readings using no more than two slides
Ø Slide 1: Key takeaway messages from the paper Ø Slide 2: Discussion ques<ons
Ä 70% Wriden paper Ø A cri<cal literature review and study design for a specific research ques<on Ø preferably related to your own research
3
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 5
Course Outline (part 1) 1. Introduc<on & Orienta<on (today!)
Ä Intro to philosophy & sociology of science Ä Role of theory building
2. Research Design and Ethics Ä What counts as evidence? Ä Use of mixed methods Ä Ethics concerns for human subjects studies
3. Basics of Doing Research Ä Finding good research ques<ons Ä Theory building Ä Evidence and Measurement Ä Replica<on Ä Peer Review
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 6
Course Outline (cont) 4. Laboratory Experiments
Ä Controlled Experiments Ä Quasi-‐experiments Ä Sampling Ä Confounding Variables
5. Quan<ta<ve Analysis Ä Basic Stats Ä Interpre<ng significance measures Ä power analysis
6. Interviews and Observa<on Ä Conduc<ng Interviews Ä Par<cipant observa<on Ä Collec<ng field notes
7. Qualita<ve Analysis Ä Coding Strategies Ä Grounded Theory Ä Phenomenography
8. Case Studies Ä Single and Mul<-‐case Ä Longitudinal Case Studies
9. Survey Research Ä Ques<onnaire Design Ä Sample Size
10. Interven<on Methods Ä Ac<on Research Ä Pilot Studies Ä Benchmarking
11. Publishing and Reviewing Ä (mock PC mee<ng)
12. Replica<on and Beyond Ä Internal and External Replica<on Ä Biases and Influences Ä Threats to Validity Ä When to use empirical methods
4
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 7
Is this your research plan?
Step 1: Build a new tool
Step 2: ??
Step 3: Profit
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 8
Engineering vs. Science
➜ Tradi<onal View: Scien<sts… Engineers… create knowledge apply that knowledge study the world as it is seek to change the world are trained in scien<fic method are trained in engineering design use explicit knowledge use tacit knowledge are thinkers are doers
➜ More realis<c View Scien<sts… Engineers… create knowledge create knowledge are problem-‐driven are problem-‐driven seek to understand and explain seek to understand and explain design experiments to test theories design devices to test theories prefer abstract knowledge prefer con<ngent knowledge but rely on tacit knowledge but rely on tacit knowledge
Both involve a mix of design and discovery
5
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 9
Scien<fic Method
➜ No single “official” scien<fic method
➜ Somehow, scien<sts are supposed to do this:
World Theory
Observation
Validation
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 10
Observe!
6
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 11
Scien<fic Inquiry
Prior Knowledge (Initial Hypothesis)
Observe (what is wrong with the current theory?)
Theorize (refine/create a better theory)
Design (Design empirical tests
of the theory)
Experiment (manipulate the variables)
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 12
7
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 13
Some Characteris<cs of Science
➜ Science seeks to improve our understanding of the world.
➜ Explana<ons are based on observa<ons Ä Scien<fic truths must stand up to empirical scru<ny Ä Some<mes “scien<fic truth” must be thrown out in the face of new findings
➜ Theory and observa<on affect one another: Ä Our percep<ons of the world affect how we understand it Ä Our understanding of the world affects how we perceive it
➜ Crea<vity is important Ä Theories, hypotheses, experimental designs Ä Search for elegance, simplicity
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 14
All Methods are flawed
➜ E.g. Laboratory Experiments Ä Cannot study large scale sorware development in the lab! Ä Too many variables to control them all!
➜ E.g. Case Studies Ä How do we know what’s true in one project generalizes to others? Ä Researcher chose what ques<ons to ask, hence biased the study
➜ E.g. Surveys Ä Self-‐selec<on of respondents biases the study Ä Respondents tell you what they think they ought to do, not what they actually do
➜ …etc...
8
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 15
Strategies to overcome weaknesses
➜ Theory-‐building Ä Tes<ng a hypothesis is pointless (single flawed study!)… Ä …unless it builds evidence for a clearly stated theory
➜ Empirical Induc<on Ä Series of studies over <me… Ä Each designed to probe more aspects of the theory Ä …together build evidence for a clearly stated theory
➜ Mixed Methods Research Ä Use mul<ple methods to inves<gate the same research ques<on Ä Each method compensates for the flaws of the others Ä …together build evidence for a clearly stated theory
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 16 Source: http://xkcd.com/882/"
9
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 17
What is a research contribu<on?
➜ A beder understanding of how people use sorware technology?
➜ Iden<fica<on of problems with the current state-‐of-‐the-‐art?
➜ A characteriza<on of the proper<es of new tools/techniques?
➜ Evidence that approach A is beder than approach B?
How will you validate your claims?"
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 18
Meet Stuart Dent
➜ Name: Ä Stuart Dent (a.k.a. “Stu”)
➜ Advisor: Ä Prof. Helen Back
➜ Topic: Ä Merging Stakeholder views in Model Driven
Development
➜ Status: Ä 2 years into his PhD Ä Has built a tool Ä Needs an evalua<on plan
10
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 19
Stu’s Evalua<on Plan ➜ Formal Experiment
Ä Independent Variable: Stu-‐Merge vs. Ra<onal Architect Ä Dependent Variables: Correctness, Speed, Subjec<ve Assessment Ä Task: Merging Class Diagrams from two different stakeholders’ models Ä Subjects: Grad Students in SE Ä H1: “Stu-‐Merge produces correct merges more oren than RA” Ä H2: “Subjects produce merges faster with Stu-‐Merge than with RA” Ä H3: “Subjects prefer using Stu-‐Merge to RA”
➜ Results Ä H1 accepted (strong evidence) Ä H2 & H3 rejected Ä Subjects found the tool unintui<ve
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 20
Threats to Validity
➜ Construct Validity Ä What do we mean by a merge? What is correctness? Ä 5-‐point scale for subjec<ve assessment -‐ insufficient discriminatory power
Ø (both tools scored very low)
➜ Internal Validity Ä Confounding variables: Time taken to learn the tool; familiarity
Ø Subjects were all familiar with RA, not with Stu-‐merge
➜ External Validity Ä Task representa<veness
Ø class models were of a toy problem
Ä Subject representa<veness Ø Grad students as sample of what popula<on?
➜ Theore<cal Reliability Ä Researcher bias
Ø subjects knew Stu-‐merge was Stu’s own tool
11
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 21
What went wrong?
➜ What was the research ques<on? Ä “Is tool A beder than tool B?”
➜ What would count as an answer?
➜ What use would the answer be? Ä How is it a “contribu<on to knowledge”?
➜ How does this evalua<on relate to the exis<ng literature?
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 22
Experiments as Clinical Trials
Is drug A better than drug B?
Why would we expect it to be better?
Better at doing what?
Better in what way?
Better in what situations?
Why do we need to know?
What will we do with the
answer?
12
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 23
You goda have a theory!
Why would we expect it to be better?
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 24
Some Defini<ons
➜ A model is an abstract representa<on of a phenomenon or set of related phenomena Ä Some details included, others excluded
➜ A theory is a set of statements that explain a set of phenomena Ä Serves to explain and predict Ä Precisely defined terminology Ä Concepts, rela<onships, causal inferences Ä (opera<onal defini<ons for theore<cal terms)
➜ A hypothesis is a testable statement derived from a theory Ä A hypothesis is not a theory!
➜ In CS, we have mostly folk theories
13
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 25
A simpler defini<on
A (good) Theory is the best explanation of all the available evidence"
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 26
The Role of Theory Building
➜ Theories lie at the heart of what it means to do science. Ä Produc<on of generalizable knowledge
➜ Theory provides orienta<on for data collec<on Ä Cannot observe the world without a theore<cal perspec<ve
➜ Theories allow us to compare similar work Ä Theories include precise defini<on for the key terms Ä Theories provide a ra<onale for which phenomena to measure
➜ Theories support analy<cal generaliza<on Ä Provide a deeper understanding of our empirical results Ä …and hence how they apply more generally Ä Much more powerful than sta<s<cal generaliza<on
14
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 27
Stu’s Theory ➜ Background Assump<ons
Ä Large team projects, models contributed by many actors Ä Models are fragmentary, capture par<al views Ä Par<al views are inconsistent and incomplete most of the <me
➜ Basic Theory Ä (Brief summary:) Ä Model merging is an exploratory process, in which the aim is to discover intended
rela<onships between views. ‘Goodness’ of a merge is a subjec<ve judgment. If an adempted merge doesn’t seem ‘good’, many need to change either the models, or the way in which they were mapped together.
Ä [S<ll needs some work]
➜ Derived Hypotheses Ä Useful merge tools need to represent rela<onships explicitly Ä Useful merge tools need to be complete (work for any models, even if inconsistent)
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 28
What type of ques<on are you asking? ➜ Existence:
Ä Does X exist?
➜ Descrip<on & Classifica<on Ä What is X like? Ä What are its proper<es? Ä How can it be categorized? Ä How can we measure it? Ä What are its components?
➜ Descrip<ve-‐Compara<ve Ä How does X differ from Y?
➜ Frequency and Distribu<on Ä How oren does X occur? Ä What is an average amount of X?
➜ Descrip<ve-‐Process Ä How does X normally work? Ä By what process does X happen? Ä What are the steps as X evolves?
➜ Rela<onship Ä Are X and Y related? Ä Do occurrences of X correlate with occurrences
of Y?
➜ Causality Ä Does X cause Y? Ä Does X prevent Y? Ä What causes X? Ä What effect does X have on Y?
➜ Causality-‐Compara<ve Ä Does X cause more Y than does Z? Ä Is X beder at preven<ng Y than is Z? Ä Does X cause more Y than does Z under one
condi<on but not others?
➜ Design Ä What is an effec<ve way to achieve X? Ä How can we improve X?
15
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 29
What type of ques<on are you asking? ➜ Existence:
Ä Does X exist?
➜ Descrip<on & Classifica<on Ä What is X like? Ä What are its proper<es? Ä How can it be categorized? Ä How can we measure it? Ä What are its components?
➜ Descrip<ve-‐Compara<ve Ä How does X differ from Y?
➜ Frequency and Distribu<on Ä How oren does X occur? Ä What is an average amount of X?
➜ Descrip<ve-‐Process Ä How does X normally work? Ä By what process does X happen? Ä What are the steps as X evolves?
➜ Rela<onship Ä Are X and Y related? Ä Do occurrences of X correlate with occurrences
of Y?
➜ Causality Ä Does X cause Y? Ä Does X prevent Y? Ä What causes X? Ä What effect does X have on Y?
➜ Causality-‐Compara<ve Ä Does X cause more Y than does Z? Ä Is X beder at preven<ng Y than is Z? Ä Does X cause more Y than does Z under one
condi<on but not others?
➜ Design Ä What is an effec<ve way to achieve X? Ä How can we improve X?
Exploratory
Baserate
Correlation
Design
Causal Relationship
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 30
Stu’s Research Ques<on(s) ➜ Existence
Ä Does model merging ever happen in prac<ce?
➜ Descrip<on/Classifica<on Ä What are the different types of model merging that occur in prac<ce on large scale
systems?
➜ Descrip<ve-‐Compara<ve Ä How does model merging with explicit representa<on of rela<onships differ from model
merging without such representa<on?
➜ Causality Ä Does an explicit representa<on of the rela<onship between models cause developers to
explore different ways of merging models?
➜ Causality-‐Compara<ve Ä Does the algebraic representa<on of rela<onships in Stu’s tool lead developers to explore
more than do pointcuts in AOM?
Pick just one for now…"
16
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 31
Puyng the Ques<on in Context
The Research Question
How does this relate to the established literature?
What new perspectives are you bringing to this field?
What methods are appropriate for answering this question?
Existing Theories
New Concepts
Methodological Choices Empirical Method
Data Collection Techniques
Data Analysis Techniques
Philosophical Context Positivist Constructivist
Critical theory Pragmatist
What will you accept as valid truth?
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 32
Many available methods… Common “in the lab” Methods
● Controlled Experiments
➜ Ra<onal Reconstruc<ons
➜ Exemplars
➜ Benchmarks
➜ Simula<ons
Common “in the wild”
Methods
● Quasi-Experiments ● Case Studies ● Survey Research ● Ethnographies ● Action Research
❍ Artifact/Archive Analysis (“mining”!)
17
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 33
Stu’s Method(s) Selec<on… ➜ Existence
Ä Does model merging ever happen in prac<ce?
➜ Descrip<on/Classifica<on Ä What are the different types of model merging that
occur in prac<ce on large scale systems?
➜ Descrip<ve-‐Compara<ve Ä How does model merging with explicit representa<on
of rela<onships differ from model merging without such representa<on?
➜ Causality Ä Does an explicit representa<on of the rela<onship
between models cause developers to explore different ways of merging models?
➜ Causality-‐Compara<ve Ä Does the algebraic representa<on of rela<onships in
Stu’s tool lead developers to explore more than do pointcuts in AOM?
Case study
Survey Research
Controlled Experiment
Action Research
Ethnography
?
? ?
?
?
?
?
?
?
?
?
?
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 34
Warning
No method is perfect
Don’t get hung up on methodological purity
Pick something and get on with it
Some knowledge is better than none
18
35
University of Toronto Department of Computer Science
© 2004-5 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license.
Okay, but…
University of Toronto Department of Computer Science
© 2012 Steve Easterbrook. This presentation is available free for non-commercial use with attribution under a creative commons license. 36
Why Build a Tool?
➜ Build a Tool to Test a Theory Ä Tool is part of the experimental materials needed to conduct your study
➜ Build a Tool to Develop a Theory Ä Theory emerges as you explore the tool
➜ Build a Tool to Explain your Theory Ä Theory as a concrete instan<a<on of (some aspect of) the theory
?