Designing Test Collections for Comparing Many Systems
Tetsuya SakaiWaseda University, Japan
@tetsuyasakai
November 4, 2014@CIKM 2014, Shanghai
Acknowledgement
This research is a part of Waseda University’s project “Taxonomising and Evaluating Web Search Engine User Behaviours,” supported by Microsoft Research.
THANK YOU!
Takeaways (1)
• Using one‐way ANOVA‐based power analysis, researchers can determine the topic set size n by specifying:
α: Type I error probability β: Type II error probabilityminD: minimum detectable range (performance diff between best and worst systems) for ensuring a statistical power of 1‐βm: number of systems to be compared
: estimated variance of each system• Different measures have different s, so researchers should decide on the evaluation measure at the test collection design phase
Takeaways (2)
• Our method can provide different test collection designs (n, pd) that satisfy the same statistical requirement.
n: topic set size pd: pool depth• The assessment cost of a pd=100 test collection can be reduced to 18% or less while keeping it statistically equally reliable.
• Our method can be used to compare evaluation measures in terms of practical significance = judgment cost.
• Our tools and data are available athttp://www.f.waseda.jp/tetsuya/tools.htmlhttp://www.f.waseda.jp/tetsuya/data.html
TALK OUTLINE
1. How Information Retrieval (IR) test collections are constructed2. Statistical reform3. How test collections SHOULD be constructed4. Experimental results5. Conclusions and future work
Test collections = standard data sets for evaluation
Test collection A Test collection BEvaluationmeasurevalues
Evaluationmeasurevalues
An Information Retrieval (IR) test collection
Topic Relevance assessments(relevant/nonrelevant documents)
Document collection
Topic Relevance assessments(relevant/nonrelevant documents)
Topic Relevance assessments(relevant/nonrelevant documents)
: :
Topic set “Qrels = query relevance sets”
CIKM 2014 home page cikm2014.fudan.edu.cn/: highly relevant
cikmconference.org/: partially relevantwww.cikm2013.org: nonrelevant
How IR people build test collections (1)
Okay, let’s build a test collection…
Organiser
How IR people build test collections (2)
…with maybe n=50topics (search requests)…
Well n>25 sounds good for statistical significance testing, but why 50? Why not 100? Why not 30?
TopicTopicTopicTopicTopic 1
How IR people build test collections (3)
TopicTopicTopicTopicTopic 1
50 topicsOkay folks, give me your runs (search results)!
run run run
Participants
How IR people build test collections (4)
TopicTopicTopicTopicTopic 1
50 topicsPool depth pd=100 looks
affordable…
run run run
Top pd=100 documentsfrom each run
Pool for
Topic 1Document collection too large to doexhaustive relevance assessments sojudge pooled documents only
How IR people build test collections (5)
TopicTopicTopicTopicTopic 1
50 topics
Top pd=100 documentsfrom each run
Pool for
Topic 1Relevance assessments
Highly relevant
Partially relevant
Nonrelevant
An Information Retrieval (IR) test collection
Topic Relevance assessments(relevant/nonrelevant documents)
Document collection
Topic Relevance assessments(relevant/nonrelevant documents)
Topic Relevance assessments(relevant/nonrelevant documents)
: :
Topic set “Qrels = query relevance sets”
CIKM 2014 home page cikm2014.fudan.edu.cn/: highly relevant
cikmconference.org/: partially relevantwww.cikm2013.org: nonrelevant
n=50topics…why?
Pool depth pd=100(not exhaustive)
TALK OUTLINE
1. How Information Retrieval (IR) test collections are constructed2. Statistical reform3. How test collections SHOULD be constructed4. Experimental results5. Conclusions and future work
NHST = null hypothesis significance testing (1)
EXAMPLE: paired t‐test for comparing systems X and Y with n topics
Assumptions:
Null hypothesis:
Test statistic:
Population means are the same
NHST = null hypothesis significance testing (2)
EXAMPLE: paired t‐test for comparing systems X and Y with n topics
Null hypothesis:
Test statistic:
Under H0, t0 obeys a t distribution with n‐1 degrees of freedom.
NHST = null hypothesis significance testing (3)
EXAMPLE: paired t‐test for comparing systems X and Y with n topicsNull hypothesis:Under H0, t0 obeys a t distribution with n‐1 degrees of freedom.
Given a significance criterion α(=0.05), reject H0 if |t0| >= t(n‐1; α).
0
0.1
0.2
0.3
0.4
‐t(n‐1; α)
n=50
t(n‐1; α)
“H0 is probably not true because the chance of observing t0 under H0
is very small”
NHST = null hypothesis significance testing (4)
EXAMPLE: paired t‐test for comparing systems X and Y with n topicsNull hypothesis:Given a significance criterion α(=0.05), reject H0 if |t0| >= t(n‐1; α).
0
0.1
0.2
0.3
0.4
‐t(n‐1; α)
n=50
t(n‐1; α)
0
0.1
0.2
0.3
0.4
‐t(n‐1; α)
n=50
t(n‐1; α)
Conclusion:X ≠ Y!
t0 t0Conclusion:
H0 not rejected, so don’t know
NHST is not good enough [Cumming12]
• Dichotomous thinking ( “different or not different?” )A more important question is “what is the magnitude of the difference?” Another is “How accurate is my estimate?”• p‐values a little more informative than “significant at α=0.05” but…
0
0.1
0.2
0.3
0.4
‐t(n‐1; α)
n=50
t(n‐1; α)
t0
Probability of observing t0 or something more extreme under H0
The p‐value is not good enough either [Ellis10,Nagata03]
Reject H0 if |t0| >= t(n‐1; α) where
But a large |t0| could mean two things:(1) Sample effect size (ES) is large;(2) Topic set size n is large.
If you increase the sample size n, you can always achieve statistical significance!
Difference between X and Y measured in standard deviation
units
Statistical reform – effect sizes [Cumming12,Okubo12]
• ES: “how much difference is there?”• ES for paired t test measures difference in standard deviation unitsPopulation ES =
Sample ES as an estimate of the above =
In several research disciplines such as psychology and medicine, it is required to report ESs! In this study, we determine the topic set size n by ensuring high power 1‐β whenever ES is large.
Statistical reform – confidence intervals
• CIs are much more informative than NHST(point estimate + uncertainty/accuracy)• Estimation thinking, not dichotomous thinking[Cumming12]
In several research disciplines such as psychology and medicine, it is required to report CIs! See [Sakai14FIT] (Designing Test Collections that Provide Tight Confidence Intervals)
[Sakai14SIGIRforum]
TALK OUTLINE
1. How Information Retrieval (IR) test collections are constructed2. Statistical reform3. How test collections SHOULD be constructed4. Experimental results5. Conclusions and future work
Overview of our approaches
• Using the paired t‐test to determine n (for m=2 systems)INPUT: α, β, minDt (minimum detectable difference for ensuring power=1‐β), (estimated variance of performance difference)• Using one‐way ANOVA to determine n (for m>=2 systems)INPUT: α, β, m, minD (minimum detectable range for ensuring power=1‐β), (estimated variance of system performance)• Methods for estimating and INPUT: test collections with runs and an evaluation measure(topic‐by‐run matrices)
See Appendix
The ANOVA approach (1)
Assume
i=1,…,m (systems)j=1,…,n (topics)
Homoscedasticity(equal variance)
Let
where
Hypotheses
: ai ≠ 0 for some i
System effect
No system effect
The ANOVA approach (2)i=1,…,m (systems)j=1,…,n (topics)
Total variation
can be decomposed intoST = SA + SE where
Between‐systemvariation
Within‐systemvariation
Sample grand mean
Sample system mean
The ANOVA approach (3)i=1,…,m (systems)j=1,…,n (topics)
Test statistic F0 = VA/VE where VA = SA/φA, VE = SE/φE,φA = m‐1, φE = m(n‐1)
Under H0,F0~ F distribution with (φA, φE) degrees of freedom.One‐way ANOVA rejects H0if F0 >= F(φA; φE; α).
F(φA; φE; α)F0
α
1‐α
How large is the between‐system variance compared to the within‐system variance?
The ANOVA approach (4)i=1,…,m (systems)j=1,…,n (topics)
The probability of rejecting H0
Under H0, this is exactly α (rejecting H0 that is true)
F(φA; φE; α)F0
α
1‐α
The ANOVA approach (5)i=1,…,m (systems)j=1,…,n (topics)
The probability of rejecting H0
Under H1, this is exactly the power 1‐β (rejecting H0 that is false)
Under H1, F0~ noncentral F distribution with (φA, φE) degrees of freedomand a noncentrality parameter λ = nΔ, where
Measures total system effects in variance units
The ANOVA approach (6)i=1,…,m (systems)j=1,…,n (topics)
The power (probability of rejecting H0 that is false)
1‐βF0~ noncentralF distribution
For a random variable F’ that obeys a noncenral F distribution,Pr{ F’ <= w } can be approximated using a normal distribution[Nagata03] (Eqs. 14 and 15 in my paper).
Given α, n, Δ, m, the power 1‐β can be computed.But what we want is: Given α, β, Δ, m, compute n!
Under H0, we know Δ=0. But under H1, we only know that Δ ≠ 0.Δ needs to be specified to guarantee power=1‐β.Let’s guarantee power=1‐β (i.e. correctly reject H0 with 100(1‐β)% confidence)whenever Δ >= minΔ (minimum detectable delta).
How shall we set minΔ?
The ANOVA approach (7)i=1,…,m (systems)j=1,…,n (topics)
Let
where minD is the minimum detectable range that you specify.Whenever the difference between the best system and the worst system is minD or more, we guarantee power=1‐β.
Given α, n, Δ, m, the power 1‐βcan be computed.Using α, n, minD, , m, the worst‐case power can be computed (See Eq.17 in my paper).
The ANOVA approach (8)i=1,…,m (systems)j=1,…,n (topics)
Estimate variance from past data
μ
best
worst
μi
minDai = μi ‐ μ
The ANOVA approach (9)i=1,…,m (systems)j=1,…,n (topics)
Given (α, β, minD, , m),
Here, λ can be obtained if we use the following approximation [Nagata03]:
λ = nΔ: Noncentrality parameter
1‐β ≒
Let φE = m(n‐1) ≒∞ ~ noncentral chi‐square distribution with φA degrees of
freedom and the same noncentrality parameter λ
The ANOVA approach (10)i=1,…,m (systems)j=1,…,n (topics)
Given (α, β, minD, , m), λ = nΔ: Noncentrality parameter
φA = m‐1
For noncentral chi‐square distributions, use the following λ values [Nagata03]:
The ANOVA approach (11)i=1,…,m (systems)j=1,…,n (topics)
Given (α, β, minD, , m),
Obtain n using λ from the table, and check if (α, n, minΔ, m) satisfies the power requirement:
λ = nΔ: Noncentrality parameter
1‐β
Normal approximation available (Eqs. 15, 16 in my
paper)
If not, n++ and try the above again.
Demo: determine n from (α, β, minDt, , m)
http://www.f.waseda.jp/tetsuya/CIKM2014/samplesizeANOVA.xlsx
Obtaining
A time‐honoured method using one‐way ANOVA statistics [Okubo12]:
Estimate of the population between‐system variance
Estimate of the population within‐system variance
Estimate of the populationvariance
Given multiple ANOVA statistics (test collections + runs),the estimated variances can be pooled to enhance reliability:
Overview of our approaches
• Using the paired t‐test to determine n (for m=2 systems)INPUT: α, β, minDt (minimum detectable difference for ensuring power=1‐β), (estimated variance of performance difference)• Using one‐way ANOVA to determine n (for m>=2 systems)INPUT: α, β, m, minD (minimum detectable range for ensuring power=1‐β), (estimated variance of system performance)• Methods for estimating and INPUT: test collections with runs and an evaluation measure(topic‐by‐run matrices)
See Appendix
TALK OUTLINE
1. How Information Retrieval (IR) test collections are constructed2. Statistical reform3. How test collections SHOULD be constructed4. Experimental results5. Conclusions and future work
Data for estimating
Data #topics runs pd #docsTREC03new 50 78 125 528,155 news articlesTREC04new 49 78 100 dittoTREC11w 50 37 25 One billion web pagesTREC12w 50 28 20/30 dittoTREC11wD 50 25 25 dittoTREC12wD 50 20 20/30 ditto
Adhoc news IR
Adhoc web IR
Diversified web IR
We have a topic‐by‐run matrix for each data set and evaluation measure
Evaluation measures
News(l=10,1000)Web (l=10)
Web (l=10)
l: measurement depth
t‐test vs ANOVAWhen m=2,• minD for ANOVA (range) reducesto minDt for t‐test (difference).• Results are similar,with ANOVA giving slightly larger estimates of n.Henceforth we discussANOVA as it can also considerm>3 and we prefer to“err on the side of oversampling”[Ellis10] .
0
1000
2000
3000
4000
5000
6000
0 20 40 60 80 100 120 140 160 180 200AP Q nDCG nERR
(a2) adhoc/news (l=10)(α, β, minD)=(0.05, 0.20, 0.05)
m
n
For comparing m=100 systems,Q/nDCG/AP/nERR require
2198/2382/2863/4063 topics
0
500
1000
1500
2000
2500
3000
3500
4000
0 20 40 60 80 100 120 140 160 180 200AP Q nDCG nERR
(b) adhoc/web (l=10)(α, β, minD)=(0.05, 0.20, 0.05)
m
n
For comparing m=100 systems,Q/nDCG/AP/nERR require
1240/1291/2801/2921 topics
0
500
1000
1500
2000
2500
3000
3500
4000
0 20 40 60 80 100 120 140 160 180 200α‐nDCG nERR‐IA D‐nDCG D#‐nDCG
(c) diversity/web (l=10)(α, β, minD)=(0.05, 0.20, 0.05)
m
n
For comparing m=100 systems,D/D#/α/nERR‐IA require
1201/1749/2662/2869 topics
What if we reduce the pool depth pd?
TopicTopicTopicTopicTopic 1
n=50 topics
Top pd=100 documentsfrom each run
Pool for
Topic 1Relevance assessments
Highly relevant
Partially relevant
Nonrelevant
For adhoc/news l=1000 (pd=100) only
when pd is reduced
As pd gets smaller,• Average #judged/topic decreases (naturally)• Variance increases (fewer data points hurts stability)Re‐estimate n for (α, β, minD, new )
0
200
400
600
800
1000
1200
1400
1600
0 100 200 300 400 500 600 700 800AP Q nDCG nERR
pd=50
n
(α, β, minD)=(0.05, 0.20, 0.05)m=10
pd=100pd=70
pd=30pd=10
#Average judged/topic
Total cost for AP:96 docs/topic *
879 topics = 84,384 docs
Total cost for AP:731 docs/topic *
652 topics = 476,612 docs
TREC ad hocpool depth
Alternative design with costreduced to 18%
TALK OUTLINE
1. How Information Retrieval (IR) test collections are constructed2. Statistical reform3. How test collections SHOULD be constructed4. Experimental results5. Conclusions and future work
Takeaways (1)
• Using one‐way ANOVA‐based power analysis, researchers can determine the topic set size n by specifying:
α: Type I error probability β: Type II error probabilityminD: minimum detectable range (performance diff between best and worst systems) for ensuring a statistical power of 1‐βm: number of systems to be compared
: estimated variance of each system• Different measures have different s, so researchers should decide on the evaluation measure at the test collection design phase
Takeaways (2)
• Our method can provide different test collection designs (n, pd) that satisfy the same statistical requirement.
n: topic set size pd: pool depth• The assessment cost of a pd=100 test collection can be reduced to 18% or less while keeping it statistically equally reliable.
• Our method can be used to compare evaluation measures in terms of practical significance = judgment cost.
• Our tools and data are available athttp://www.f.waseda.jp/tetsuya/tools.htmlhttp://www.f.waseda.jp/tetsuya/data.html
Future work
• Investigating the relationship between our power‐based approach and a CI (confidence interval)‐based approach: DONE [Sakai14EVIA]
• Estimating n for various tasks (not just IR) – our methods are applicable to any paired‐data evaluation tasks
• Given a set of statistically equally reliable designs (n,pd), choose the best one based on reusability and assessment cost
Can we evaluate new systems fairly?
References[Cumming12] Cumming, G.: Understanding The New Statistics: Effect Sizes, Confidence Intervals, and Meta‐Analysis. Routledge, 2012.[Ellis10] Ellis, P.D.: The Essential Guide to Effect Sizes, Cambridge, 2010.[Nagata03] Nagata, Y.: How to Design the Sample Size. Asakura Shoten, 2003.[Okubo12] Okubo, M. and Okada, K. Psychological Statistics to Tell Your Story: Effect Size, CondenceInterval (in Japanese). Keiso Shobo, 2012.[Sakai14PROMISE] Sakai, T.: Metrics, Statistics, Tests. PROMISE Winter School 2013: Bridging between Information Retrieval and Databases (LNCS 8173), pp.116‐163, Springer, 2014. [Sakai14SIGIRforum] Sakai, T.: Statistical Reform in Information Retrieval?, SIGIR Forum, 48(1), 2014. [Sakai14FIT] Sakai, T.: Designing Test Collections that Provide Tight Confidence Intervals, FIT 2014, RD‐003, 2014.[Webber08] Webber, W., Moffat, A. and Zobel, J.: Statistical power in Retrieval Experimentation. ACM CIKM 2008, pp.571–580, 2008.
Appendices
• The t‐test approach• Delta over [Webber08]
The t‐test approach (1)
vs.
Hypotheses
Assume
⇒
where
Systems X and Y are equally effective
The t‐test approach (2)
Test statistic
Under H0, t0~ t distribution with φ=n‐1 degrees of freedom.The paired t test rejects H0 if |t0| >= t(φ; α).
α/2 α/2
t(φ; α)‐t(φ; α)
1‐αt0
Two‐sided critical t value
The t‐test approach (3)
The probability of rejecting H0
α/2 α/2
t(φ; α)‐t(φ; α)
1‐αt0
Under H0, this is exactly α (rejecting H0 that is true)
The t‐test approach (4)The probability of rejecting H0
Under H1, this is exactly the power 1‐β (rejecting H0 that is false)
Under H1, t0~ noncentral t distribution with φ=n‐1 degrees of freedom anda noncentrality parameter λt =where
Effect size
The t‐test approach (5)The power (probability of rejecting H0 that is false)
1‐β =
t0~ noncentralt distribution
For a random variable t’ that obeys a noncenral t distribution,Pr{ t’ <= w } can be approximated using a normal distribution[Nagata03] (Eqs. 4 and 5 in my paper).
Given α, n, effect size Δt, the power 1‐β can be computed.But what we want is: Given α, β, Δt, compute n!
The t‐test approach (6)Given α, n, effect size Δt, the power 1‐β can be computed.But what we want is: Given α, β, Δt, compute n!
Under H0, we know Δt = 0. But under H1, we need to specify Δt to discuss power.So let’s correctly reject H0 with 100(1‐β)% confidencewhenever |Δt| >= minΔt (minimum detectable effect).
Don’t miss a real difference if it’s minΔt or larger!
The t‐test approach (7)
Given (α, β, minΔt), the required n can be approximated [Nagata03]:
zp: one‐sided critical z value
1‐β <=
Check if the above n actually satisfies the power requirement:
t0~ noncentralt distribution
If not , n++ and try the above again.
The t‐test approach (8)
Effect size
In practice, instead of setting minΔt (in terms of effect size),set minDt (minimum detectable difference):
|μX – μY | > = minDtthen convert it to
Need a variance estimatefrom past data!
Demo: determine n from (α, β, minDt, )
http://www.f.waseda.jp/tetsuya/CIKM2014/samplesizeTTEST.xlsx
Delta over [Webber08]
• They addressed the problem of building a test collection incrementally (add a topic, judge, re‐estimate variance…).
• They considered the t‐test only.
• They used heuristics to estimate the variance.
• They considered AP and adhoc IR only.
We ask the direct question: “How many topics do we need to create?”
We use both the t‐test and ANOVA to handle m(>=2) systems
We use estimates from ANOVA
We consider a variety of graded‐relevance measures and three different IR tasks.