evaluation in (music) information retrieval through the audio music similarity task
DESCRIPTION
Test-collection based evaluation in (Music) Information Retrieval has been used for half a century now as the means to evaluate and compare retrieval techniques and advance the state of the art. However, this paradigm makes certain assumptions that remain a research problem and that may invalidate our experimental results. In this talk I will approach this paradigm as an estimator of certain probability distributions that describe the final user experience. These distributions are estimated with a test collection, computing system-related distributions assumed to reliably correlate with the target user-related distributions. Using the Audio Music Similarity task as an example, I will talk about issues with our current evaluation methods, the degree to which they are problematic, how to analyze them and improve the situation. In terms of validity, we will see how the measured system distributions correspond to the target user distributions, and how this correspondence affects the conclusions we draw from an experiment. In terms of reliability, we will discuss optimal characteristics of test collections and statistical procedures. In terms of efficiency, we discuss models and methods to greatly reduce the annotation cost of an evaluation experiment.TRANSCRIPT
Evaluation in (Music) Information Retrieval through the Audio Music Similarity task
Julián Urbano
Barcelona, Spain · January 16th 2014
Spam
• @julian_urbano
• Postdoctoral researcher
– Music Technology Group, Universitat Pompeu Fabra
• Recently: PhD, Computer Science
– (Evaluation in) (Music) Information Retrieval
2
Information Retrieval
• Automatic representation, storage and search of unstructured information
– Traditionally textual information
– Lately multimedia too: images, video, music
• A user has an information need and uses an IR system that retrieves the relevant or significant information from a collection of documents
3
Information Retrieval Evaluation
• IR systems are based on models to estimate relevance, implementing different techniques
• How good is my system? What system is better?
– Answered with IR Evaluation experiments
– “if you can’t measure it, you can’t improve it”
– But we need to be able to trust our measurements
• Research on IR Evaluation
– Improve our methods to evaluate systems
– Critical for the correct development of the field
4
Disclaimer
• If you see…
A system is evaluated with a test collection containing queries, documents and judgments
telling how relevant a document is to a query
• …you can think of
An algorithm is evaluated with a dataset containing queries, songs and annotations
telling how similar a song is to a query
5
Talk outline
• Why we want to Evaluate…
• …and what we do with Cranfield
• Validity: users versus systems
• Reliability: estimating from samples
• Efficiency: reducing annotations
6
Introduction: Why we want to Evaluate…
The two questions
• How good is my system?
– What does good mean?
– What is good enough?
• Is system A better than system B?
– What does better mean?
– How much better?
• Efficiency? Effectiveness? Ease? 8
Measure user experience
• We are interested in user-measures
– Time to complete task, idle time, success rate, failure rate, frustration, ease to learn, ease to use …
• Their distributions describe user experience, fully
– For an arbitrary user, query and document collection, what is the distribution of…
9
0 time to complete task
none frustration
much some
The big(ger) picture
• Different user-measures attempting to assess the same thing: user satisfaction
– How likely is it that an arbitrary user, with an arbitrary query (and with an arbitrary document collection) will be satisfied by the system?
• This is the ultimate goal: the good, the better
10
The big(ger) question
• User satisfaction…as Bernoulli trial
• Probability of satisfaction P(Sat = yes)?
• Probability that k in n users are satisfied?
• Probability of >80% users satisfied?
11
satisfaction yes no
Introduction: …what we do with Cranfield
Sources of variability
user-measure = f(documents, query, user, system)
• Our goal is the distribution of the user-measure for our system, which is impossible to calculate
– (Possibly?) infinite population
• The best we can do is estimate it
– Sample documents, queries and users
– Measure user experience, implicitly or explicitly
– Representativeness, cost, ethics, privacy…
13
Fix samples
• Hard to replicate experiment and repeat results
• Just plain impossible to reproduce results
• Get a (hopefully) good sample and fix it
– Documents and queries
• But we can’t fix the users!
14
Simulate users…and fix them
• Cranfield paradigm: remove users, but include a user-abstraction, fixed across experiments
– Static user component: judgments in the ground truth
– Dynamic user component: effectiveness measures
• Remove all sources of variability, except systems
user-measure = f(documents, query, user, system)
15
Simulate users…and fix them
• Cranfield paradigm: remove users, but include a user-abstraction, fixed across experiments
– Static user component: judgments in the ground truth
– Dynamic user component: effectiveness measures
• Remove all sources of variability, except systems
user-measure = f(documents, query, user, system)
user-measure = f(system)
15
Test collections
• Controlled set of documents, queries and judgments, shared across researchers
• (Most?) important resource for IR research
– Experiments are inexpensive (collections are not!)
– Research becomes systematic
– Reproducibility becomes possible and easy
16
Wait a minute
• Are we estimating distributions about users or distributions about systems?
system-effectiveness = f(system, scale, measure)
• We come up with different distributions of system-effectiveness, depending on how we abstract users from the experiment – Different scales to assess relevance
– Different measures to model user behavior
17
Assumption
• System-measures correspond to user-measures
18
Users Systems
Time to complete task Idle time
Success rate Failure rate Frustration
Ease to learn Ease to use Satisfaction
…
P AP RR DCG nDCG ERR GAP Q …
Assumption
• System-measures correspond to user-measures
18
Users Systems
Time to complete task Idle time
Success rate Failure rate Frustration
Ease to learn Ease to use Satisfaction
…
P AP RR DCG nDCG ERR GAP Q …
Assumption
• System-measures correspond to user-measures
18
Users Systems
Time to complete task Idle time
Success rate Failure rate Frustration
Ease to learn Ease to use Satisfaction
…
P AP RR DCG nDCG ERR GAP Q …
Assumption
• System-measures correspond to user-measures
18
Users Systems
Time to complete task Idle time
Success rate Failure rate Frustration
Ease to learn Ease to use Satisfaction
…
P AP RR DCG nDCG ERR GAP Q …
Assumption
• System-measures correspond to user-measures
19
Users Systems
Time to complete task Idle time
Success rate Failure rate Frustration
Ease to learn Ease to use Satisfaction
…
P AP RR DCG nDCG ERR GAP Q …
Experiments with Test Collections
• Our goal is the users
user-measure = f(system)
• but Cranfield tells us about systems
system-effectiveness = f(system, scale, measure)
20
Experiments with Test Collections
• Our goal is the users
user-measure = f(system)
• but Cranfield tells us about systems
system-effectiveness = f(system, scale, measure)
20
Experiments with Test Collections
• Our goal is the users
user-measure = f(system)
• but Cranfield tells us about systems
system-effectiveness = f(system, scale, measure)
• This poses several problems
– That we have been dealing with for over 50 years
– But hey, they’re extremely interesting!
20
Validity, Reliability and Efficiency
• Validity: are we measuring what we want to? – Internal: are observed effects due to hidden factors?
– External: are queries, documents and users representative?
– Construct: do system-measures match user-measures?
– Conclusion: how good is good and how better is better?
• Reliability: how repeatable are the results? – How large do collections need to be?
– What statistical methods should be used?
• Efficiency: how inexpensive is it to get valid and reliable results? (i.e. to build a test collection) – Can we estimate results with fewer judgments?
21
In this talk
How to study and improve the validity, reliability and efficiency
of the methods used to evaluate IR systems
• Audio Music Similarity task as example – Song as query input to system, audio signal
– Retrieve songs musically similar to it, by content
– Resembles traditional Ad Hoc retrieval in Text IR
– Important task in Music IR • Music recommendation
• Playlist generation
• Plagiarism detection
22
Validity: Effectiveness and Satisfaction
Assumption of Cranfield
• Systems with better effectiveness are perceived by users as more useful, more satisfactory
• Tricky: different effectiveness measures and relevance scales produce different distributions
– Which one is better to predict satisfaction?
• Map system effectiveness onto user satisfaction, experimentally
– If P@10 = 0.2, how likely is it that an arbitrary user will find the results satisfactory?
– What is P(Sat | P@10 = 0.2)? 24
User-oriented System-measures
• Effectiveness measures are generally not formulated to correlate with user-satisfaction
– If effectiveness is λ = 0, we expect P(Sat) = 0
– If effectiveness is λ = 1, we expect P(Sat) = 1
– In general, we expect P(Sat | λ) = λ
• But this is not what we have
– Effectiveness measures need to be reformulated
– Upper bounds, recall components, ideal rankings
– Many mathematical details omitted in this talk 25
User Components: Measures and Scales
• How is relevance measured in the judgments?
– Nominal, ordinal, interval, ratio
• How are results consumed?
– Set, list
• What determines document utility?
– Positional, cascade
– Linear, exponential
• What determines user persistence?
– Navigational, informational
26
Measures and Scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=3 nℒ=4 nℒ=5 ℓmin=20 ℓmin=40 ℓmin=60 ℓmin=80
P@5 X X X X
AP@5 X X X X
RR@5 X X X X
CGl@5 X X X X X P@5 P@5 P@5 P@5
CGe@5 X X X X P@5 P@5 P@5 P@5
DCGl@5 X X X X X X X X X
DCGe@5 X X X X DCGl@5 DCGl@5 DCGl@5 DCGl@5
EDCGl@5 X X X X X X X X X
EDCGe@5 X X X X EDCGl@5 EDCGl@5 EDCGl@5 EDCGl@5
Ql@5 X X X X X AP@5 AP@5 AP@5 AP@5
Qe@5 X X X X AP@5 AP@5 AP@5 AP@5
RBPl@5 X X X X X X X X X
RBPe@5 X X X X RBPl@5 RBPl@5 RBPl@5 RBPl@5
ERRl@5 X X X X X X X X X
ERRe@5 X X X X ERRl@5 ERRl@5 ERRl@5 ERRl@5
GAP@5 X X X X X AP@5 AP@5 AP@5 AP@5
ADR@5 X X X X X X X X
27
Measures and Scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=3 nℒ=4 nℒ=5 ℓmin=20 ℓmin=40 ℓmin=60 ℓmin=80
P@5 X X X X
AP@5 X X X X
RR@5 X X X X
CGl@5 X X X X X P@5 P@5 P@5 P@5
CGe@5 X X X X P@5 P@5 P@5 P@5
DCGl@5 X X X X X X X X X
DCGe@5 X X X X DCGl@5 DCGl@5 DCGl@5 DCGl@5
EDCGl@5 X X X X X X X X X
EDCGe@5 X X X X EDCGl@5 EDCGl@5 EDCGl@5 EDCGl@5
Ql@5 X X X X X AP@5 AP@5 AP@5 AP@5
Qe@5 X X X X AP@5 AP@5 AP@5 AP@5
RBPl@5 X X X X X X X X X
RBPe@5 X X X X RBPl@5 RBPl@5 RBPl@5 RBPl@5
ERRl@5 X X X X X X X X X
ERRe@5 X X X X ERRl@5 ERRl@5 ERRl@5 ERRl@5
GAP@5 X X X X X AP@5 AP@5 AP@5 AP@5
ADR@5 X X X X X X X X
27
Measures and Scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=3 nℒ=4 nℒ=5 ℓmin=20 ℓmin=40 ℓmin=60 ℓmin=80
P@5 X X X X
AP@5 X X X X
RR@5 X X X X
CGl@5 X X X X X P@5 P@5 P@5 P@5
CGe@5 X X X X P@5 P@5 P@5 P@5
DCGl@5 X X X X X X X X X
DCGe@5 X X X X DCGl@5 DCGl@5 DCGl@5 DCGl@5
EDCGl@5 X X X X X X X X X
EDCGe@5 X X X X EDCGl@5 EDCGl@5 EDCGl@5 EDCGl@5
Ql@5 X X X X X AP@5 AP@5 AP@5 AP@5
Qe@5 X X X X AP@5 AP@5 AP@5 AP@5
RBPl@5 X X X X X X X X X
RBPe@5 X X X X RBPl@5 RBPl@5 RBPl@5 RBPl@5
ERRl@5 X X X X X X X X X
ERRe@5 X X X X ERRl@5 ERRl@5 ERRl@5 ERRl@5
GAP@5 X X X X X AP@5 AP@5 AP@5 AP@5
ADR@5 X X X X X X X X
27
Measures and Scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=3 nℒ=4 nℒ=5 ℓmin=20 ℓmin=40 ℓmin=60 ℓmin=80
P@5 X X X X
AP@5 X X X X
RR@5 X X X X
CGl@5 X X X X X P@5 P@5 P@5 P@5
CGe@5 X X X X P@5 P@5 P@5 P@5
DCGl@5 X X X X X X X X X
DCGe@5 X X X X DCGl@5 DCGl@5 DCGl@5 DCGl@5
EDCGl@5 X X X X X X X X X
EDCGe@5 X X X X EDCGl@5 EDCGl@5 EDCGl@5 EDCGl@5
Ql@5 X X X X X AP@5 AP@5 AP@5 AP@5
Qe@5 X X X X AP@5 AP@5 AP@5 AP@5
RBPl@5 X X X X X X X X X
RBPe@5 X X X X RBPl@5 RBPl@5 RBPl@5 RBPl@5
ERRl@5 X X X X X X X X X
ERRe@5 X X X X ERRl@5 ERRl@5 ERRl@5 ERRl@5
GAP@5 X X X X X AP@5 AP@5 AP@5 AP@5
ADR@5 X X X X X X X X
27
Measures and Scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=3 nℒ=4 nℒ=5 ℓmin=20 ℓmin=40 ℓmin=60 ℓmin=80
P@5 X X X X
AP@5 X X X X
RR@5 X X X X
CGl@5 X X X X X P@5 P@5 P@5 P@5
CGe@5 X X X X P@5 P@5 P@5 P@5
DCGl@5 X X X X X X X X X
DCGe@5 X X X X DCGl@5 DCGl@5 DCGl@5 DCGl@5
EDCGl@5 X X X X X X X X X
EDCGe@5 X X X X EDCGl@5 EDCGl@5 EDCGl@5 EDCGl@5
Ql@5 X X X X X AP@5 AP@5 AP@5 AP@5
Qe@5 X X X X AP@5 AP@5 AP@5 AP@5
RBPl@5 X X X X X X X X X
RBPe@5 X X X X RBPl@5 RBPl@5 RBPl@5 RBPl@5
ERRl@5 X X X X X X X X X
ERRe@5 X X X X ERRl@5 ERRl@5 ERRl@5 ERRl@5
GAP@5 X X X X X AP@5 AP@5 AP@5 AP@5
ADR@5 X X X X X X X X
27
Measures and Scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=3 nℒ=4 nℒ=5 ℓmin=20 ℓmin=40 ℓmin=60 ℓmin=80
P@5 X X X X
AP@5 X X X X
RR@5 X X X X
CGl@5 X X X X X P@5 P@5 P@5 P@5
CGe@5 X X X X P@5 P@5 P@5 P@5
DCGl@5 X X X X X X X X X
DCGe@5 X X X X DCGl@5 DCGl@5 DCGl@5 DCGl@5
EDCGl@5 X X X X X X X X X
EDCGe@5 X X X X EDCGl@5 EDCGl@5 EDCGl@5 EDCGl@5
Ql@5 X X X X X AP@5 AP@5 AP@5 AP@5
Qe@5 X X X X AP@5 AP@5 AP@5 AP@5
RBPl@5 X X X X X X X X X
RBPe@5 X X X X RBPl@5 RBPl@5 RBPl@5 RBPl@5
ERRl@5 X X X X X X X X X
ERRe@5 X X X X ERRl@5 ERRl@5 ERRl@5 ERRl@5
GAP@5 X X X X X AP@5 AP@5 AP@5 AP@5
ADR@5 X X X X X X X X
27
Measures and Scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=3 nℒ=4 nℒ=5 ℓmin=20 ℓmin=40 ℓmin=60 ℓmin=80
P@5 X X X X
AP@5 X X X X
RR@5 X X X X
CGl@5 X X X X X P@5 P@5 P@5 P@5
CGe@5 X X X X P@5 P@5 P@5 P@5
DCGl@5 X X X X X X X X X
DCGe@5 X X X X DCGl@5 DCGl@5 DCGl@5 DCGl@5
EDCGl@5 X X X X X X X X X
EDCGe@5 X X X X EDCGl@5 EDCGl@5 EDCGl@5 EDCGl@5
Ql@5 X X X X X AP@5 AP@5 AP@5 AP@5
Qe@5 X X X X AP@5 AP@5 AP@5 AP@5
RBPl@5 X X X X X X X X X
RBPe@5 X X X X RBPl@5 RBPl@5 RBPl@5 RBPl@5
ERRl@5 X X X X X X X X X
ERRe@5 X X X X ERRl@5 ERRl@5 ERRl@5 ERRl@5
GAP@5 X X X X X AP@5 AP@5 AP@5 AP@5
ADR@5 X X X X X X X X
27
Measures and Scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=3 nℒ=4 nℒ=5 ℓmin=20 ℓmin=40 ℓmin=60 ℓmin=80
P@5 X X X X
AP@5 X X X X
RR@5 X X X X
CGl@5 X X X X X P@5 P@5 P@5 P@5
CGe@5 X X X X P@5 P@5 P@5 P@5
DCGl@5 X X X X X X X X X
DCGe@5 X X X X DCGl@5 DCGl@5 DCGl@5 DCGl@5
EDCGl@5 X X X X X X X X X
EDCGe@5 X X X X EDCGl@5 EDCGl@5 EDCGl@5 EDCGl@5
Ql@5 X X X X X AP@5 AP@5 AP@5 AP@5
Qe@5 X X X X AP@5 AP@5 AP@5 AP@5
RBPl@5 X X X X X X X X X
RBPe@5 X X X X RBPl@5 RBPl@5 RBPl@5 RBPl@5
ERRl@5 X X X X X X X X X
ERRe@5 X X X X ERRl@5 ERRl@5 ERRl@5 ERRl@5
GAP@5 X X X X X AP@5 AP@5 AP@5 AP@5
ADR@5 X X X X X X X X
27
Measures and Scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=3 nℒ=4 nℒ=5 ℓmin=20 ℓmin=40 ℓmin=60 ℓmin=80
P@5 X X X X
AP@5 X X X X
RR@5 X X X X
CGl@5 X X X X X P@5 P@5 P@5 P@5
CGe@5 X X X X P@5 P@5 P@5 P@5
DCGl@5 X X X X X X X X X
DCGe@5 X X X X DCGl@5 DCGl@5 DCGl@5 DCGl@5
EDCGl@5 X X X X X X X X X
EDCGe@5 X X X X EDCGl@5 EDCGl@5 EDCGl@5 EDCGl@5
Ql@5 X X X X X AP@5 AP@5 AP@5 AP@5
Qe@5 X X X X AP@5 AP@5 AP@5 AP@5
RBPl@5 X X X X X X X X X
RBPe@5 X X X X RBPl@5 RBPl@5 RBPl@5 RBPl@5
ERRl@5 X X X X X X X X X
ERRe@5 X X X X ERRl@5 ERRl@5 ERRl@5 ERRl@5
GAP@5 X X X X X AP@5 AP@5 AP@5 AP@5
ADR@5 X X X X X X X X
27
Measures and Scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=3 nℒ=4 nℒ=5 ℓmin=20 ℓmin=40 ℓmin=60 ℓmin=80
P@5 X X X X
AP@5 X X X X
RR@5 X X X X
CGl@5 X X X X X P@5 P@5 P@5 P@5
CGe@5 X X X X P@5 P@5 P@5 P@5
DCGl@5 X X X X X X X X X
DCGe@5 X X X X DCGl@5 DCGl@5 DCGl@5 DCGl@5
EDCGl@5 X X X X X X X X X
EDCGe@5 X X X X EDCGl@5 EDCGl@5 EDCGl@5 EDCGl@5
Ql@5 X X X X X AP@5 AP@5 AP@5 AP@5
Qe@5 X X X X AP@5 AP@5 AP@5 AP@5
RBPl@5 X X X X X X X X X
RBPe@5 X X X X RBPl@5 RBPl@5 RBPl@5 RBPl@5
ERRl@5 X X X X X X X X X
ERRe@5 X X X X ERRl@5 ERRl@5 ERRl@5 ERRl@5
GAP@5 X X X X X AP@5 AP@5 AP@5 AP@5
ADR@5 X X X X X X X X
27
Measures and Scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=3 nℒ=4 nℒ=5 ℓmin=20 ℓmin=40 ℓmin=60 ℓmin=80
P@5 X X X X
AP@5 X X X X
RR@5 X X X X
CGl@5 X X X X X P@5 P@5 P@5 P@5
CGe@5 X X X X P@5 P@5 P@5 P@5
DCGl@5 X X X X X X X X X
DCGe@5 X X X X DCGl@5 DCGl@5 DCGl@5 DCGl@5
EDCGl@5 X X X X X X X X X
EDCGe@5 X X X X EDCGl@5 EDCGl@5 EDCGl@5 EDCGl@5
Ql@5 X X X X X AP@5 AP@5 AP@5 AP@5
Qe@5 X X X X AP@5 AP@5 AP@5 AP@5
RBPl@5 X X X X X X X X X
RBPe@5 X X X X RBPl@5 RBPl@5 RBPl@5 RBPl@5
ERRl@5 X X X X X X X X X
ERRe@5 X X X X ERRl@5 ERRl@5 ERRl@5 ERRl@5
GAP@5 X X X X X AP@5 AP@5 AP@5 AP@5
ADR@5 X X X X X X X X
27
Measures and Scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=3 nℒ=4 nℒ=5 ℓmin=20 ℓmin=40 ℓmin=60 ℓmin=80
P@5 X X X X
AP@5 X X X X
RR@5 X X X X
CGl@5 X X X X X P@5 P@5 P@5 P@5
CGe@5 X X X X P@5 P@5 P@5 P@5
DCGl@5 X X X X X X X X X
DCGe@5 X X X X DCGl@5 DCGl@5 DCGl@5 DCGl@5
EDCGl@5 X X X X X X X X X
EDCGe@5 X X X X EDCGl@5 EDCGl@5 EDCGl@5 EDCGl@5
Ql@5 X X X X X AP@5 AP@5 AP@5 AP@5
Qe@5 X X X X AP@5 AP@5 AP@5 AP@5
RBPl@5 X X X X X X X X X
RBPe@5 X X X X RBPl@5 RBPl@5 RBPl@5 RBPl@5
ERRl@5 X X X X X X X X X
ERRe@5 X X X X ERRl@5 ERRl@5 ERRl@5 ERRl@5
GAP@5 X X X X X AP@5 AP@5 AP@5 AP@5
ADR@5 X X X X X X X X
27
MIREX
Experimental design
28
What can we infer?
• Preference: difference noticed by user
– Positive: user agrees with evaluation
– Negative: user disagrees with evaluation
• Non-preference: difference not noticed by user
– Good: both systems are satisfactory
– Bad: both systems are not satisfactory
29
Data
• Queries, documents and judgments from MIREX
• 4115 unique and artificial examples
– At least 200 examples per (measure-scale-λ)
• 432 unique queries, 5636 unique documents
• Answers collected via Crowdsourcing
– Quality control with trap questions
• 113 unique subjects 30
Single system: how good is it?
• For 2045 examples (49%) users could not decide which system was better
What do we expect?
31
Single system: how good is it?
• For 2045 examples (49%) users could not decide which system was better
31
Single system: how good is it?
• Large ℓmin thresholds underestimate satisfaction
32
Single system: how good is it?
• Users don’t pay attention to ranking?
33
Single system: how good is it?
• Exponential gain underestimates satisfaction
34
Single system: how good is it?
• Document utility independent of others
35
Two systems: which one is better?
• For 2090 examples (51%) users did prefer one system over the other one
What do we expect?
36
Two systems: which one is better?
• For 2090 examples (51%) users did prefer one system over the other one
36
Two systems: which one is better?
• Large differences needed for users to note them
37
Two systems: which one is better?
• More relevance levels are better to discriminate
38
Two systems: which one is better?
• Cascade and navigational user models are not appropriate
39
Two systems: which one is better?
• Users do prefer the (supposedly) worse system
40
Summary
• Effectiveness and satisfaction are clearly correlated – There is a 20% bias: P(Sat | 0) > 0 and P(Sat | 1) < 1 – Room to improve: personalization, better user abstraction
• Magnitude of differences does matter – Just looking at rankings is very naive
• Be careful with statistical significance
– Need Δλ≈0.4 for users to agree with effectiveness • Historically, only 20% of times in MIREX
• Differences among measures and scales – Linear gain slightly better than exponential gain – Informational and positional user models better than
navigational and cascade – The more relevance levels, the better
41
Measures and scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=3 nℒ=4 nℒ=5 ℓmin=20 ℓmin=40 ℓmin=60 ℓmin=80
P@5 X X X X
AP@5 X X X X
RR@5 X X X X
CGl@5 X X X X X P@5 P@5 P@5 P@5
CGe@5 X X X X P@5 P@5 P@5 P@5
DCGl@5 X X X X X X X X X
DCGe@5 X X X X DCGl@5 DCGl@5 DCGl@5 DCGl@5
EDCGl@5 X X X X X X X X X
EDCGe@5 X X X X EDCGl@5 EDCGl@5 EDCGl@5 EDCGl@5
Ql@5 X X X X X AP@5 AP@5 AP@5 AP@5
Qe@5 X X X X AP@5 AP@5 AP@5 AP@5
RBPl@5 X X X X X X X X X
RBPe@5 X X X X RBPl@5 RBPl@5 RBPl@5 RBPl@5
ERRl@5 X X X X X X X X X
ERRe@5 X X X X ERRl@5 ERRl@5 ERRl@5 ERRl@5
GAP@5 X X X X X AP@5 AP@5 AP@5 AP@5
ADR@5 X X X X X X X X
42
Measures and scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=3 nℒ=4 nℒ=5 ℓmin=20 ℓmin=40 ℓmin=60 ℓmin=80
P@5 X X X X
AP@5 X X X X
RR@5 X X X X
CGl@5 X X X X X P@5 P@5 P@5 P@5
CGe@5 X X X X P@5 P@5 P@5 P@5
DCGl@5 X X X X X X X X X
DCGe@5 X X X X DCGl@5 DCGl@5 DCGl@5 DCGl@5
EDCGl@5 X X X X X X X X X
EDCGe@5 X X X X EDCGl@5 EDCGl@5 EDCGl@5 EDCGl@5
Ql@5 X X X X X AP@5 AP@5 AP@5 AP@5
Qe@5 X X X X AP@5 AP@5 AP@5 AP@5
RBPl@5 X X X X X X X X X
RBPe@5 X X X X RBPl@5 RBPl@5 RBPl@5 RBPl@5
ERRl@5 X X X X X X X X X
ERRe@5 X X X X ERRl@5 ERRl@5 ERRl@5 ERRl@5
GAP@5 X X X X X AP@5 AP@5 AP@5 AP@5
ADR@5 X X X X X X X X
43
Validity: Satisfaction over Samples
Evaluate in terms of user satisfaction
• So far, arbitrary users for a single query
– P Sat Ql@5 = 0.61 = 0.7
• Easily for n users and a single query
– P Sat15 = 10 Ql@5 = 0.61 = 0.21
• What about a sample of queries 𝒬?
– Map queries separately for the distribution of P(Sat)
– For easier mappings, P(Sat | λ) functions are interpolated with simple polynomials
45
Expected probability of satisfaction
• Now we can compute point and interval estimates of the expected probability of satisfaction
• Intuition fails when interpreting effectiveness
46
System success
• If P(Sat) ≥ threshold the system is successful
– Setting the threshold was rather arbitrary before
– Now it is meaningful, in terms of user satisfaction
• Intuitively, we want the majority of users to find the system satisfactory
– P Succ = P P Sat > 0.5 = 1 − FP Sat (0.5)
• Improving queries for which we are bad is worthier than further improving those for which we are already good
47
Distribution of P(Sat)
• But we (will) only have a handful queries, estimates will probably be bad – Need to estimate the cumulative distribution function of
user satisfaction: FP(Sat)
– Not described by any typical distribution family
• More than ≈25 queries in the collection – ecdf approximates better
• Less than ≈25 queries in the collection – Normal for graded scales, ecdf for binary scales
• Beta is always the best with the Fine scale – Which turns out to be the best scale, overall
48
Intuition fails, again
Intuitive conclusions based on effectiveness alone contradict those based on user satisfaction
– E Δλ = −0.002
– E ΔP Sat = 0.001
– E ΔP Succ = 0.07
49
Intuition fails, again
Intuitive conclusions based on effectiveness alone contradict those based on user satisfaction
– E Δλ = −0.002
– E ΔP Sat = 0.001
– E ΔP Succ = 0.07
49
Intuition fails, again
Intuitive conclusions based on effectiveness alone contradict those based on user satisfaction
– E Δλ = −0.002
– E ΔP Sat = 0.001
– E ΔP Succ = 0.07
49
Historically, in MIREX
• Systems are not as satisfactory as we thought
• But they are more successful
– Good (or bad) for some kinds of queries
50
Measures and scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=4 nℒ=5 ℓmin=20 ℓmin=40
P@5 X X
AP@5 X X
CGl@5 X X X X P@5 P@5
CGe@5 X X X P@5 P@5
DCGl@5 X X X X X X
DCGe@5 X X X DCGl@5 DCGl@5
Ql@5 X X X X AP@5 AP@5
Qe@5 X X X AP@5 AP@5
RBPl@5 X X X X X X
RBPe@5 X X X RBPl@5 RBPl@5
GAP@5 X X X X AP@5 AP@5
51
Measures and scales
Measure Original Artificial Graded Artificial Binary
Broad Fine nℒ=4 nℒ=5 ℓmin=20 ℓmin=40
P@5 X X
AP@5 X X
CGl@5 X X X X P@5 P@5
CGe@5 X X X P@5 P@5
DCGl@5 X X X X X X
DCGe@5 X X X DCGl@5 DCGl@5
Ql@5 X X X X AP@5 AP@5
Qe@5 X X X AP@5 AP@5
RBPl@5 X X X X X X
RBPe@5 X X X RBPl@5 RBPl@5
GAP@5 X X X X AP@5 AP@5
52
Reliability: Optimal Statistical Significance Tests
Random error
• Test collections are just samples from larger, possibly infinite, populations
• If we conclude system A is better than B, how confident can we be?
– Δλ𝒬 is just an estimate of the population mean μΔλ
• Usually employ some statistical significance test for differences in location
• If it is statistically significant, we have confidence that the true difference is at least that large
54
Statistical hypothesis testing
• Set two mutually exclusive hypotheses
– H0: μΔλ = 0
– H1: μΔλ ≠ 0
• Run test, obtain p-value= P μΔλ ≥ Δλ𝒬 H0
– p ≤ α: statistically significant, high confidence
– p > α: statistically non-significant, low confidence
• Possible errors in the binary decision
– Type I: incorrectly reject H0
– Type II: incorrectly accept H0
55
Statistical significance tests
• (Non-)parametric tests
– t-test, Wilcoxon test, Sign test
• Based on resampling
– Bootstrap test, permutation/randomization test
• They make certain assumptions about distributions and sampling methods
– Often violated in IR evaluation experiments
– Which test behaves better, in practice, knowing that assumptions are violated?
56
Optimality criteria
• Power
– Achieve significance as often as possible (low Type II)
– Usually increases Type I error rates
• Safety
– Minimize Type I error rates
– Usually decreases power
• Exactness
– Maintain Type I error rate at α level
– Permutation test is theoretically exact
57
Experimental design
• Randomly split query set in two
• Evaluate all systems with both subsets
– Simulating two different test collections
• Compare p-values with both subsets
– How well do statistical tests agree with themselves?
– At different α levels
• All systems and queries from MIREX 2007-2011
– >15M p-values 58
Power and success
• Bootstrap test is the most powerful
• Wilcoxon, bootstrap and permutation are the most successful, depending on α level
59
Conflicts
• Wilcoxon and t-test are the safest at low α levels
• Wilcoxon is the most exact at low α levels, but bootstrap is for usual levels
60
Summary
• Bootstrap test is the most powerful, and still it has smaller Type I error rates, so we are safe
• Power and success:
– CGl@5 > GAP@5 > DCGl@5 > RBPl@5
– Fine > Broad > binary
• Conflicts:
– Very similar across measures and scales
– Corrections for multiple comparisons (e.g. Tukey) do not seem necessary
61
Reliability: Test Collection Size
Acceptable sample size
• Reliability is higher with larger sample sizes
– But it is also more expensive
– What is an acceptable test collection size?
• Answer with Generalizability Theory
– G-Study: estimate variance components
– D-Study: estimate reliability of different sample sizes and experimental designs
63
G-study: variance components
Fully crossed experimental design: s × q
λq,A = λ + λA + λq + εqA
σ2 = σs
2 + σq2 + σsq
2
• Estimated with Analysis of Variance
64
G-study: variance components
Fully crossed experimental design: s × q
λq,A = λ + λA + λq + εqA
σ2 = σs
2 + σq2 + σsq
2
• Estimated with Analysis of Variance
64
G-study: variance components
Fully crossed experimental design: s × q
λq,A = λ + λA + λq + εqA
σ2 = σs
2 + σq2 + σsq
2
• Estimated with Analysis of Variance
64
G-study: variance components
Fully crossed experimental design: s × q
λq,A = λ + λA + λq + εqA
σ2 = σs
2 + σq2 + σsq
2
• Estimated with Analysis of Variance
64
G-study: variance components
Fully crossed experimental design: s × q
λq,A = λ + λA + λq + εqA
σ2 = σs
2 + σq2 + σsq
2
• Estimated with Analysis of Variance
64
G-study: variance components
Fully crossed experimental design: s × q
λq,A = λ + λA + λq + εqA
σ2 = σs
2 + σq2 + σsq
2
• Estimated with Analysis of Variance
64
G-study: variance components
Fully crossed experimental design: s × q
λq,A = λ + λA + λq + εqA
σ2 = σs
2 + σq2 + σsq
2
• Estimated with Analysis of Variance
64
G-study: variance components
Fully crossed experimental design: s × q
λq,A = λ + λA + λq + εqA
σ2 = σs
2 + σq2 + σsq
2
• Estimated with Analysis of Variance
64
G-study: variance components
Fully crossed experimental design: s × q
λq,A = λ + λA + λq + εqA
σ2 = σs
2 + σq2 + σsq
2
• Estimated with Analysis of Variance
64
G-study: variance components
Fully crossed experimental design: s × q
λq,A = λ + λA + λq + εqA
σ2 = σs
2 + σq2 + σsq
2
• Estimated with Analysis of Variance
64
Intuition
• If σs2 is small or σq
2 is large, we need more queries
65
𝜆𝐴 𝜆𝐶 𝜆𝐷 𝜆𝐸 𝜆𝐹 𝜆𝐵
Intuition
• If σs2 is small or σq
2 is large, we need more queries
65
𝜆𝐴 𝜆𝐶 𝜆𝐷 𝜆𝐸 𝜆𝐹 𝜆𝐵
𝜆𝐴 𝜆𝐶 𝜆𝐷 𝜆𝐸 𝜆𝐹 𝜆𝐵
Larger σs2
Intuition
• If σs2 is small or σq
2 is large, we need more queries
65
𝜆𝐴 𝜆𝐶 𝜆𝐷 𝜆𝐸 𝜆𝐹 𝜆𝐵
𝜆𝐴 𝜆𝐶 𝜆𝐷 𝜆𝐸 𝜆𝐹 𝜆𝐵 𝜆𝐴 𝜆𝐶 𝜆𝐷 𝜆𝐸 𝜆𝐹 𝜆𝐵
Larger σs2 Smaller σq
2 or
more queries
D-study: variance ratios
• Stability of absolute scores
Φ nq =σs2
σs2 +
σq2 + σe
2
nq
• Stability of relative scores
Eρ2 nq =σs2
σs2 +
σe2
nq
• We can easily estimate how many queries are needed to reach some level of stability (reliability)
66
D-study: variance ratios
• Stability of absolute scores
Φ nq =σs2
σs2 +
σq2 + σe
2
nq
• Stability of relative scores
Eρ2 nq =σs2
σs2 +
σe2
nq
• We can easily estimate how many queries are needed to reach some level of stability (reliability)
66
Effect of query set size • Average absolute stability Φ = 0.97 • ≈65 queries needed for Φ2 = 0.95, ≈100 in worst cases • Fine scale slightly better than Broad and binary scales • RBPl@5 and nDCGl@5 are the most stable
67
Effect of query set size • Average relative stability Eρ 2 = 0.98
• ≈35 queries needed for Eρ2 = 0.95, ≈60 in worst cases
• Fine scale better than Broad and binary scales
• CGl@5 and RBPl@5 are the most stable
68
Effect of cutoff k
• What if we use a deeper cutoff, k=10?
– From 100 queries and k=5 to 50 queries and k=10
– Should still have stable scores
– Judging effort should decrease
– Rank-based measures should become more stable
• Tested in MIREX 2012
– In 2013 too, but not analyzed here
69
Effect of cutoff k
• Judging effort reduced to 72% of the usual
• Generally stable – From Φ = 0.81 to Φ = 0.83
– From Eρ 2 = 0.93 to Eρ 2 = 0.95
70
Effect of cutoff k
• Reliability given a fixed budged for judging?
– k=10 allows us to use fewer queries, about 70%
– Slightly reduced relative stability
71
Effect of assessor set size
• More assessors or simply more queries?
– Judging effort is multiplied
• Can be studied with MIREX 2006 data
– 3 different assessors per query
– Nested experimental design: s × h: q
72
Effect of assessor set size
• Broad scale: σ s2 ≈ σ h:q
2
• Fine scale: σ s2 ≫ σ h:q
2
• Always better to spend resources on queries
73
Summary
• MIREX collections generally larger than necessary
• For fixed budget
– More queries better than more assessors
– More queries slightly better than deeper cutoff
• Worth studying alternative user model?
• Employ G-Theory while building the collection
• Fine better than Broad, better than binary
• CGl@5 and DCGl@5 best for relative stability
• RBPl@5 and nDCGl@5 best for absolute stability 74
Implications
• Fixing the number of queries across years is unrealistic – Especially because they are not intended for reuse
• Fixing the number of queries across task is simply nonsense
• Need to analyze on a case-by-case basis, while building the collections – GT4IReval, R package online – https://github.com/julian-urbano/GT4IREval
75
Efficiency: Learning Relevance Distributions
Probabilistic evaluation
• The MIREX setting is still expensive – Need to judge all top k documents from all systems
– Takes days, even weeks sometimes
• Model relevance probabilistically – Relevance judgments are random variables over the space
of possible assignments of relevance
E Rd = P 𝑅𝑑 = ℓ · ℓ
ℓ∈ℒ
Var 𝑅𝑑 = P 𝑅𝑑 = ℓ · ℓ2
ℓ∈ℒ
− E 𝑅𝑑2
• Effectiveness measures are also probabilistic 77
Probabilistic evaluation
• Accuracy increases as we make judgments
– E Rd ← rd
• Reliability increases too (confidence)
– Var Rd ← 0
• Iteratively estimate relevance and effectiveness
– If confidence is low, make judgments
– If confidence is high, stop
• Judge as few documents as possible 78
Learning distributions of relevance
• Uniform distribution is very uninformative
• Historical distribution in MIREX has high variance
• Estimate from a set of features: P Rd = ℓ θd
– For each document separately
– Ordinal Logistic Regression
• Three sets of features
– Output-based, can always be used
– Judgment-based, to exploit known judgments
– Audio-based, to exploit musical similarity 79
Learned models
• Mout : can be used even without judgments
– Similarity between systems’ outputs
– Genre and artist metadata
• Genre is highly correlated to similarity
– Decent fit, R2 ≈ 0.35
• Mjud : can be used when there are judgments
– Similarity between systems’ outputs
– Known relevance of same system and same artist
• Artist is extremely correlated to similarity
– Excellent fit, R2 ≈ 0.91 80
Estimation errors
• Actual vs. predicted by Mout
– 0.36 with Broad and 0.34 with Fine
• Actual vs. predicted by Mjud
– 0.14 with Broad and 0.09 with Fine
• Among assessors in MIREX 2006
– 0.39 with Broad and 0.31 with Fine
• Negligible under the current MIREX setting
81
Efficiency: Probabilistic Evaluation
Probabilistic effectiveness measures
• Effectiveness scores become random variables too
• Example: DCGl@k
– (Usual) deterministic formulation:
𝐷𝐶𝐺𝑙@𝑘 = 𝑟𝑖/ log2 𝑖 + 1𝑘𝑖=1
𝑛ℒ − 1 / log2 𝑖 + 1𝑘𝑖=1
– (New) probabilistic formulation:
E 𝐷𝐶𝐺𝑙@𝑘 =1
𝜂𝐷𝐶𝐺𝑙
E 𝑅𝑖log2 𝑖 + 1
𝑘
𝑖=1
Var 𝐷𝐶𝐺𝑙@𝑘 =1
𝜂𝐷𝐶𝐺𝑙2
Var 𝑅𝑖log2 𝑖 + 1 2
𝑘
𝑖=1
83
Probabilistic effectiveness measures
• Effectiveness scores become random variables too
• Example: DCGl@k
– (Usual) deterministic formulation:
𝐷𝐶𝐺𝑙@𝑘 = 𝑟𝑖/ log2 𝑖 + 1𝑘𝑖=1
𝑛ℒ − 1 / log2 𝑖 + 1𝑘𝑖=1
– (New) probabilistic formulation:
E 𝐷𝐶𝐺𝑙@𝑘 =1
𝜂𝐷𝐶𝐺𝑙
E 𝑅𝑖log2 𝑖 + 1
𝑘
𝑖=1
Var 𝐷𝐶𝐺𝑙@𝑘 =1
𝜂𝐷𝐶𝐺𝑙2
Var 𝑅𝑖log2 𝑖 + 1 2
𝑘
𝑖=1
83
𝜂𝐷𝐶𝐺𝑙
Probabilistic effectiveness measures
• From there we can compute Δ𝐷𝐶𝐺𝑙@𝑘AB
• And averages over a sample of queries 𝒬
• Different approaches to compute estimates
– Deal with dependence of random variables
– Different definitions of confidence
• For measures based on ideal ranking (nDCGl@k and RBPl@k) we do not have a closed form
– Approximated with Delta method and Taylor series
84
Ranking without judgments
1. Estimate relevance with Mout
2. Estimate relative differences and rank systems
• Average confidence in the rankings is 94%
• Average accuracy of the ranking is 92%
85
Ranking without judgments
• Can we trust individual estimates?
– Ideally, we want X% accuracy when X% confidence
– Confidence slightly overestimated in [0.9, 0.99)
86
DCGl@5
Confidence Broad Fine
In bin Accuracy In bin Accuracy
[0.5, 0.6) 23 (6.5%) 0.826 22 (6.2%) 0.636
[0.6, 0.7) 14 (4%) 0.786 16 (4.5%) 0.812
[0.7, 0.8) 14 (4%) 0.571 11 (3.1%) 0.364
[0.8, 0.9) 22 (6.2%) 0.864 21 (6%) 0.762
[0.9, 0.95) 23 (6.5%) 0.87 19 (5.4%) 0.895
[0.95, 0.99) 24 (6.8%) 0.917 27 (7.7%) 0.926
[0.99, 1) 232 (65.9%) 0.996 236 (67%) 0.996
E[Accuracy] 0.938 0.921
Relative estimates with judgments
1. Estimate relevance with Mout
2. Estimate relative differences and rank systems
3. While confidence is low (<95%) 1. Select a document and judge it
2. Update relevance estimates with Mjud when possible
3. Update estimates of differences and rank systems
• What documents should we judge? – Those that are the most informative
– Measure-dependent 87
Relative estimates with judgments
• Judging effort dramatically reduced – 1.3% with CGl@5, 9.7% with RBPl@5
• Average accuracy still 92%, but improved individually – 74% of estimates with >99% confidence, 99.9% accurate
– Expected accuracy improves slightly from 0.927 to 0.931
88
Absolute estimates with judgments
1. Estimate relevance with Mout
2. Estimate absolute effectiveness scores
3. While confidence is low (expected error >±0.05) 1. Select a document and judge it
2. Update relevance estimates with Mjud when possible
3. Update estimates of absolute effectiveness scores
• What documents should we judge? – Those that reduce variance the most
– Measure-dependent 89
Absolute estimates with judgments • The stopping condition is overly confident – Virtually no judgments are even needed (supposedly)
• But effectiveness is highly overestimated – Especially with nDCGl@5 and RBPl@5 – Mjud, and especially Mout, tend to overestimate relevance
90
Absolute estimates with judgments
• Practical fix: correct variance
• Estimates are better, but at the cost of judging
– Need between 15% and 35% of judgments
91
Summary
• Estimate ranking of systems with no judgments
– 92% accuracy on average, trustworthy individually
– Statistically significant differences are always correct
• If we want more confidence, judge documents
– As few as 2% needed to reach 95% confidence
– 74% of estimates have >99% confidence and accuracy
• Estimate absolute scores, judging as necessary
– Around 25% needed to ensure error <0.05
92
Implications
• We do not need dozens of volunteers to make thousands of judgments over several days
• Just one person spending a couple hours is fine
• The spare manpower can be put to better use – Redundant judgments to have better estimates
– Make annotations for other tasks
• It naturally promotes collaborative creation of test collections by iteratively adding the judgments needed in each experiment (if any)
93
Future Work
Validity
• User studies to understand user behavior
• What information to include in test collections
• Other forms of relevance judgment to better capture document utility
• Explicitly define judging guidelines
• Similar mapping for Text IR
– Different user models within the same task
95
Reliability
• Corrections for Multiple Comparisons
• Methods to reliably estimate reliability while building test collections
96
Efficiency
• Better models to estimate document relevance
• Correct variance when having just a few relevance judgments available
• Estimate relevance beyond k=5
• Other stopping conditions and document weights
97
Conduct similar studies
for the wealth of tasks in
Music Information Retrieval