SlideShare a Scribd company logo
Utility Theory, Minimum Effort,
and Predictive Coding
Fabrizio Sebastiani
Istituto di Scienza e Tecnologie dell’Informazione
Consiglio Nazionale delle Ricerche
56124 Pisa, Italy
DESI V – Roma, IT, 14 June 2013
What I’ll be talking about
A talk about text classification (“predictive coding”), about humans in the
loop, and about how to best support their work
I will be looking at scenarios in which
1 text classification technology is used for identifying documents belonging to a
given class / relevant to a given query ...
2 ... but the level of accuracy that can be obtained from the classifier is not
considered sufficient ...
3 ... with the consequence that one or more human assessors are asked to
inspect (and correct where appropriate) a portion of the classification
decisions, with the goal of increasing overall accuracy.
How can we support / optimize the work of the human assessors?
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 2 / 36
What I’ll be talking about
A talk about text classification (“predictive coding”), about humans in the
loop, and about how to best support their work
I will be looking at scenarios in which
1 text classification technology is used for identifying documents belonging to a
given class / relevant to a given query ...
2 ... but the level of accuracy that can be obtained from the classifier is not
considered sufficient ...
3 ... with the consequence that one or more human assessors are asked to
inspect (and correct where appropriate) a portion of the classification
decisions, with the goal of increasing overall accuracy.
How can we support / optimize the work of the human assessors?
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 2 / 36
What I’ll be talking about
A talk about text classification (“predictive coding”), about humans in the
loop, and about how to best support their work
I will be looking at scenarios in which
1 text classification technology is used for identifying documents belonging to a
given class / relevant to a given query ...
2 ... but the level of accuracy that can be obtained from the classifier is not
considered sufficient ...
3 ... with the consequence that one or more human assessors are asked to
inspect (and correct where appropriate) a portion of the classification
decisions, with the goal of increasing overall accuracy.
How can we support / optimize the work of the human assessors?
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 2 / 36
A worked out example
predicted
Y N
true
Y TP = 4 FP = 3
N FN = 4 TN = 9
F1 =
2TP
2TP + FP + FN
= 0.53
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 3 / 36
A worked out example (cont’d)
predicted
Y N
true
Y TP = 4 FP = 3
N FN = 4 TN = 9
F1 =
2TP
2TP + FP + FN
= 0.53
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 4 / 36
A worked out example (cont’d)
predicted
Y N
true
Y TP = 5 FP = 3
N FN = 3 TN = 9
F1 =
2TP
2TP + FP + FN
= 0.63
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 5 / 36
A worked out example (cont’d)
predicted
Y N
true
Y TP = 5 FP = 2
N FN = 3 TN = 10
F1 =
2TP
2TP + FP + FN
= 0.67
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 6 / 36
A worked out example (cont’d)
predicted
Y N
true
Y TP = 6 FP = 2
N FN = 2 TN = 10
F1 =
2TP
2TP + FP + FN
= 0.75
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 7 / 36
A worked out example (cont’d)
predicted
Y N
true
Y TP = 6 FP = 1
N FN = 2 TN = 11
F1 =
2TP
2TP + FP + FN
= 0.80
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 8 / 36
What I’ll be talking about (cont’d)
We need methods that
given a desired level of accuracy, minimize the assessors’ effort necessary to
achieve it; alternatively,
given an available amount of human assessors’ effort, maximize the accuracy
that can be obtained through it
This can be achieved by ranking the automatically classified documents in
such a way that, by starting the inspection from the top of the ranking, the
cost-effectiveness of the annotators’ work is maximized
We call the task of generating such a ranking Semi-Automatic Text
Classification (SATC)
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 9 / 36
What I’ll be talking about (cont’d)
We need methods that
given a desired level of accuracy, minimize the assessors’ effort necessary to
achieve it; alternatively,
given an available amount of human assessors’ effort, maximize the accuracy
that can be obtained through it
This can be achieved by ranking the automatically classified documents in
such a way that, by starting the inspection from the top of the ranking, the
cost-effectiveness of the annotators’ work is maximized
We call the task of generating such a ranking Semi-Automatic Text
Classification (SATC)
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 9 / 36
What I’ll be talking about (cont’d)
Previous work has addressed SATC via techniques developed for “active
learning”
In both cases, the automatically classified documents are ranked with the
goal of having the human annotator start inspecting/correcting from the top;
however
in active learning the goal is providing new training examples
in SATC the goal is increasing the overall accuracy of the classified set
We claim that a ranking generated “à la active learning” is suboptimal for
SATC1
1G Berardi, A Esuli, F Sebastiani. A Utility-Theoretic Ranking Method for Semi-Automated Text
Classification. Proceedings of the 35th Annual International ACM SIGIR Conference on Research and
Development in Information Retrieval (SIGIR 2012), Portland, US, 2012.
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 10 / 36
What I’ll be talking about (cont’d)
Previous work has addressed SATC via techniques developed for “active
learning”
In both cases, the automatically classified documents are ranked with the
goal of having the human annotator start inspecting/correcting from the top;
however
in active learning the goal is providing new training examples
in SATC the goal is increasing the overall accuracy of the classified set
We claim that a ranking generated “à la active learning” is suboptimal for
SATC1
1G Berardi, A Esuli, F Sebastiani. A Utility-Theoretic Ranking Method for Semi-Automated Text
Classification. Proceedings of the 35th Annual International ACM SIGIR Conference on Research and
Development in Information Retrieval (SIGIR 2012), Portland, US, 2012.
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 10 / 36
Outline of this talk
1 We discuss how to measure “error reduction” (i.e., increase in accuracy)
2 We discuss a method for maximizing the expected error reduction for a fixed
amount of annotation effort
3 We show some promising experimental results
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 11 / 36
Error Reduction, and How to Measure it
Outline
1 Error Reduction, and How to Measure it
2 Error Reduction, and How to Maximize it
3 Some Experimental Results
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 12 / 36
Error Reduction, and How to Measure it
Error Reduction, and how to measure it
Assume we have
1 class (or “query”) c;
2 classifier h for c;
3 set of unlabeled documents D that we have automatically classified by means
of h, so that every document in D is associated
with a binary decision (Y or N)
with a confidence score (a positive real number)
4 measure of accuracy A, ranging on [0,1]
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 13 / 36
Error Reduction, and How to Measure it
Error Reduction, and how to Measure it
(cont’d)
We will assume that A is
F1 =
2 · Precision · Recall
Precision + Recall
=
2 · TP
(2 · TP) + FP + FN
but any “set-based” measure of accuracy (i.e., based on a contingency table)
may be used
An amount of error, measured as E = (1 − A), is present in the automatically
classified set D
Human annotators inspect-and-correct a portion of D with the goal of
reducing the error present in D
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 14 / 36
Error Reduction, and How to Measure it
Error Reduction, and how to Measure it
(cont’d)
We will assume that A is
F1 =
2 · Precision · Recall
Precision + Recall
=
2 · TP
(2 · TP) + FP + FN
but any “set-based” measure of accuracy (i.e., based on a contingency table)
may be used
An amount of error, measured as E = (1 − A), is present in the automatically
classified set D
Human annotators inspect-and-correct a portion of D with the goal of
reducing the error present in D
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 14 / 36
Error Reduction, and How to Measure it
Error Reduction, and how to Measure it
(cont’d)
We define error at rank n (noted as E(n)) as the error still present in D after
the annotator has inspected the documents at the first n rank positions
E(0) is the initial error generated by the automated classifier
E(|D|) is 0
We define error reduction at rank n (noted as ER(n)) to be
ER(n) =
E(0) − E(n)
E(0)
the error reduction obtained by the annotator who inspects the docs at the
first n rank positions
ER(n) ∈ [0, 1]
ER(n) = 0 indicates no reduction
ER(n) = 1 indicates total elimination of error
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 15 / 36
Error Reduction, and How to Measure it
Error Reduction, and how to Measure it
(cont’d)
We define error at rank n (noted as E(n)) as the error still present in D after
the annotator has inspected the documents at the first n rank positions
E(0) is the initial error generated by the automated classifier
E(|D|) is 0
We define error reduction at rank n (noted as ER(n)) to be
ER(n) =
E(0) − E(n)
E(0)
the error reduction obtained by the annotator who inspects the docs at the
first n rank positions
ER(n) ∈ [0, 1]
ER(n) = 0 indicates no reduction
ER(n) = 1 indicates total elimination of error
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 15 / 36
Error Reduction, and How to Measure it
Error Reduction, and how to Measure it
(cont’d)
0.0 0.2 0.4 0.6 0.8 1.0
Inspection Length
0.0
0.2
0.4
0.6
0.8
1.0
ErrorReduction(ER)
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 16 / 36
Error Reduction, and How to Maximize it
Outline
1 Error Reduction, and How to Measure it
2 Error Reduction, and How to Maximize it
3 Some Experimental Results
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 17 / 36
Error Reduction, and How to Maximize it
Error Reduction, and how to Maximize it
Problem
How should we rank the documents in D so as to maximize the expected error
reduction?
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 18 / 36
Error Reduction, and How to Maximize it
A worked out example
predicted
Y N
true
Y TP = 4 FP = 3
N FN = 4 TN = 9
F1 =
2TP
2TP + FP + FN
= 0.53
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 19 / 36
Error Reduction, and How to Maximize it
A worked out example (cont’d)
predicted
Y N
true
Y TP = 4 FP = 3
N FN = 4 TN = 9
F1 =
2TP
2TP + FP + FN
= 0.53
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 20 / 36
Error Reduction, and How to Maximize it
A worked out example (cont’d)
predicted
Y N
true
Y TP = 5 FP = 3
N FN = 3 TN = 9
F1 =
2TP
2TP + FP + FN
= 0.63
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 21 / 36
Error Reduction, and How to Maximize it
A worked out example (cont’d)
predicted
Y N
true
Y TP = 5 FP = 2
N FN = 3 TN = 10
F1 =
2TP
2TP + FP + FN
= 0.67
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 22 / 36
Error Reduction, and How to Maximize it
A worked out example (cont’d)
predicted
Y N
true
Y TP = 6 FP = 2
N FN = 2 TN = 10
F1 =
2TP
2TP + FP + FN
= 0.75
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 23 / 36
Error Reduction, and How to Maximize it
A worked out example (cont’d)
predicted
Y N
true
Y TP = 6 FP = 1
N FN = 2 TN = 11
F1 =
2TP
2TP + FP + FN
= 0.80
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 24 / 36
Error Reduction, and How to Maximize it
Error Reduction, and how to Maximize it
Problem: how should we rank the documents in D so as to maximize the
expected error reduction?
Intuition 1: Documents that have a higher probability of being misclassified
should be ranked higher
Intuition 2: Documents that, if corrected, bring about a higher gain (i.e., a
bigger impact on A) should be ranked higher
Here, consider that a false positive and a false negative may have different
impacts on A (e.g., when A ≡ Fβ, for any value of β)
Bottom line
Documents that have a higher utility (= probability × gain) should be ranked
higher
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 25 / 36
Error Reduction, and How to Maximize it
Error Reduction, and how to Maximize it
Problem: how should we rank the documents in D so as to maximize the
expected error reduction?
Intuition 1: Documents that have a higher probability of being misclassified
should be ranked higher
Intuition 2: Documents that, if corrected, bring about a higher gain (i.e., a
bigger impact on A) should be ranked higher
Here, consider that a false positive and a false negative may have different
impacts on A (e.g., when A ≡ Fβ, for any value of β)
Bottom line
Documents that have a higher utility (= probability × gain) should be ranked
higher
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 25 / 36
Error Reduction, and How to Maximize it
Error Reduction, and how to Maximize it
(cont’d)
Given a set Ω of mutually disjoint events, a utility function is defined as
U(Ω) =
ω∈Ω
P(ω)G(ω)
where
P(ω) is the probability of occurrence of event ω
G(ω) is the gain obtained if event ω occurs
We can thus estimate the utility, for the aims of increasing A, of manually
inspecting a document d as
U(TP, TN, FP, FN) = P(FP) · G(FP) + P(FN) · G(FN)
provided we can estimate
If d is labelled with class c: P(FP) and G(FP)
If d is not labelled with class c: P(FN) and G(FN)
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 26 / 36
Error Reduction, and How to Maximize it
Error Reduction, and how to Maximize it
(cont’d)
Given a set Ω of mutually disjoint events, a utility function is defined as
U(Ω) =
ω∈Ω
P(ω)G(ω)
where
P(ω) is the probability of occurrence of event ω
G(ω) is the gain obtained if event ω occurs
We can thus estimate the utility, for the aims of increasing A, of manually
inspecting a document d as
U(TP, TN, FP, FN) = P(FP) · G(FP) + P(FN) · G(FN)
provided we can estimate
If d is labelled with class c: P(FP) and G(FP)
If d is not labelled with class c: P(FN) and G(FN)
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 26 / 36
Error Reduction, and How to Maximize it
Error Reduction, and how to Maximize it
(cont’d)
Estimating P(FP) and P(FN) (the probability of misclassification) can be
done by converting the confidence score returned by the classifier into a
probability of correct classification
Tricky: requires probability “calibration” via a generalized sigmoid function to
be optimized via k-fold cross-validation
Gains G(FP) and G(FN) can be defined “differentially”; i.e.,
The gain obtained by correcting a FN is (AFN→TP
− A)
The gain obtained by correcting a FP is (AFP→TN
− A)
Gains need to be estimated by estimating the contingency table on the
training set via k-fold cross-validation
Key observation: in general, G(FP) = G(FN)
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 27 / 36
Error Reduction, and How to Maximize it
Error Reduction, and how to Maximize it
(cont’d)
Estimating P(FP) and P(FN) (the probability of misclassification) can be
done by converting the confidence score returned by the classifier into a
probability of correct classification
Tricky: requires probability “calibration” via a generalized sigmoid function to
be optimized via k-fold cross-validation
Gains G(FP) and G(FN) can be defined “differentially”; i.e.,
The gain obtained by correcting a FN is (AFN→TP
− A)
The gain obtained by correcting a FP is (AFP→TN
− A)
Gains need to be estimated by estimating the contingency table on the
training set via k-fold cross-validation
Key observation: in general, G(FP) = G(FN)
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 27 / 36
Error Reduction, and How to Maximize it
Error Reduction, and how to Maximize it
(cont’d)
Estimating P(FP) and P(FN) (the probability of misclassification) can be
done by converting the confidence score returned by the classifier into a
probability of correct classification
Tricky: requires probability “calibration” via a generalized sigmoid function to
be optimized via k-fold cross-validation
Gains G(FP) and G(FN) can be defined “differentially”; i.e.,
The gain obtained by correcting a FN is (AFN→TP
− A)
The gain obtained by correcting a FP is (AFP→TN
− A)
Gains need to be estimated by estimating the contingency table on the
training set via k-fold cross-validation
Key observation: in general, G(FP) = G(FN)
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 27 / 36
Some Experimental Results
Outline
1 Error Reduction, and How to Measure it
2 Error Reduction, and How to Maximize it
3 Some Experimental Results
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 28 / 36
Some Experimental Results
Some Experimental Results
Learning algorithms: MP-Boost, SVMs
Datasets:
# Cats # Training # Test FM
1 MP-Boost FM
1 SVMs
Reuters-21578 115 9603 3299 0.608 0.527
OHSUMED-S 97 12358 3652 0.479 0.478
Baseline: ranking by probability of misclassification, equivalent to applying
our ranking method with G(FP) = G(FN) = 1
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 29 / 36
Some Experimental Results
0.0 0.2 0.4 0.6 0.8 1.0
Inspection Length
0.0
0.2
0.4
0.6
0.8
1.0
ErrorReduction(ER)Learner: MP-Boost; Dataset: Reuters-21578; Type: Macro
Random
Baseline
Utility-theoretic
Oracle
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 30 / 36
Some Experimental Results
0.0 0.2 0.4 0.6 0.8 1.0
Inspection Length
0.0
0.2
0.4
0.6
0.8
1.0
ErrorReduction(ER) Learner: SVMs; Dataset: Reuters-21578; Type: Macro
Random
Baseline
Utility-theoretic
Oracle
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 31 / 36
Some Experimental Results
0.0 0.2 0.4 0.6 0.8 1.0
Inspection Length
0.0
0.2
0.4
0.6
0.8
1.0
ErrorReduction(ER) Learner: MP-Boost; Dataset: Ohsumed-S; Type: Macro
Random
Baseline
Utility-theoretic
Oracle
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 32 / 36
Some Experimental Results
0.0 0.2 0.4 0.6 0.8 1.0
Inspection Length
0.0
0.2
0.4
0.6
0.8
1.0ErrorReduction(ER) Learner: SVMs; Dataset: Ohsumed-S; Type: Macro
Random
Baseline
Utility-theoretic
Oracle
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 33 / 36
Some Experimental Results
A few side notes
This approach allows the human annotator to know, at any stage of the
inspection process, what the estimated accuracy is at that stage
Estimate accuracy at the beginning of the process, via k-fold cross validation
Update after each correction is made
This approach lends itself to having more than one assessor working in
parallel on the same inspection task
Recent research I have not discussed today :
A “dynamic” SATC method in which gains are updated after each correction
is performed
“Microaveraging” and “Macroaveraging” -oriented methods
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 34 / 36
Some Experimental Results
Concluding Remarks
Take-away message: Semi-automatic text classification needs to be addressed
as a task in its own right
Active learning typically makes use of probabilities of misclassification but does
not make use of gains ⇒ ranking “à la active learning” is suboptimal for SATC
The use of utility theory means that the ranking algorithm is optimized for a
specific accuracy measure ⇒ Choose the accuracy measure the best mirrors
your applicative needs (e.g., Fβ with β > 1), and choose it well!
SATC is important, since in more and more application contexts the accuracy
obtainable via completely automatic text classification is not sufficient; more
and more frequently humans will need to enter the loop
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 35 / 36
Some Experimental Results
Thank you!
Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 36 / 36

More Related Content

Viewers also liked

գյուղատնտեսություն
գյուղատնտեսությունգյուղատնտեսություն
գյուղատնտեսությունYura Ohanyan
 
Acceptable Use Policies
Acceptable Use PoliciesAcceptable Use Policies
Acceptable Use Policies
Steve Meek
 
Text Quantification
Text QuantificationText Quantification
Text Quantification
Fabrizio Sebastiani
 
Format. condici
Format. condiciFormat. condici
Format. condici
jose_cantella4444
 
Borang markah final olahraga & sukaneka 2014
Borang markah final olahraga & sukaneka 2014Borang markah final olahraga & sukaneka 2014
Borang markah final olahraga & sukaneka 2014
Razita Hamid
 
JamesKey_MSN
JamesKey_MSNJamesKey_MSN
JamesKey_MSNJames Key
 
Day 1 LTE Technology Overview
Day 1 LTE Technology OverviewDay 1 LTE Technology Overview
Day 1 LTE Technology Overviewmahesh savita
 

Viewers also liked (10)

գյուղատնտեսություն
գյուղատնտեսությունգյուղատնտեսություն
գյուղատնտեսություն
 
Graf.excel
Graf.excelGraf.excel
Graf.excel
 
Dv
DvDv
Dv
 
Acceptable Use Policies
Acceptable Use PoliciesAcceptable Use Policies
Acceptable Use Policies
 
Text Quantification
Text QuantificationText Quantification
Text Quantification
 
Shahnawaz Doc.Updates jan2013
Shahnawaz Doc.Updates jan2013Shahnawaz Doc.Updates jan2013
Shahnawaz Doc.Updates jan2013
 
Format. condici
Format. condiciFormat. condici
Format. condici
 
Borang markah final olahraga & sukaneka 2014
Borang markah final olahraga & sukaneka 2014Borang markah final olahraga & sukaneka 2014
Borang markah final olahraga & sukaneka 2014
 
JamesKey_MSN
JamesKey_MSNJamesKey_MSN
JamesKey_MSN
 
Day 1 LTE Technology Overview
Day 1 LTE Technology OverviewDay 1 LTE Technology Overview
Day 1 LTE Technology Overview
 

Similar to Utility Theory, Minimum Effort, and Predictive Coding

Machine Learning and Automatic Text Classification: What's Next?
Machine Learning and Automatic Text Classification: What's Next?Machine Learning and Automatic Text Classification: What's Next?
Machine Learning and Automatic Text Classification: What's Next?
Fabrizio Sebastiani
 
Self-organisation of Knowledge in Socio-technical Systems: A Coordination Per...
Self-organisation of Knowledge in Socio-technical Systems: A Coordination Per...Self-organisation of Knowledge in Socio-technical Systems: A Coordination Per...
Self-organisation of Knowledge in Socio-technical Systems: A Coordination Per...
Andrea Omicini
 
Size-Based Disciplines for Job Scheduling in Data-Intensive Scalable Computin...
Size-Based Disciplines for Job Scheduling in Data-Intensive Scalable Computin...Size-Based Disciplines for Job Scheduling in Data-Intensive Scalable Computin...
Size-Based Disciplines for Job Scheduling in Data-Intensive Scalable Computin...
Mario Pastorelli
 
A N E XTENSION OF P ROTÉGÉ FOR AN AUTOMA TIC F UZZY - O NTOLOGY BUILDING U...
A N  E XTENSION OF  P ROTÉGÉ FOR AN AUTOMA TIC  F UZZY - O NTOLOGY BUILDING U...A N  E XTENSION OF  P ROTÉGÉ FOR AN AUTOMA TIC  F UZZY - O NTOLOGY BUILDING U...
A N E XTENSION OF P ROTÉGÉ FOR AN AUTOMA TIC F UZZY - O NTOLOGY BUILDING U...
ijcsit
 
Daniel Samaan: ChatGPT and the Future of Work
Daniel Samaan: ChatGPT and the Future of WorkDaniel Samaan: ChatGPT and the Future of Work
Daniel Samaan: ChatGPT and the Future of Work
Edunomica
 
Grouping techniques for facing Volume and Velocity in the Big Data
Grouping techniques for facing Volume and Velocity in the Big DataGrouping techniques for facing Volume and Velocity in the Big Data
Grouping techniques for facing Volume and Velocity in the Big Data
Facultad de Informática UCM
 
Frankie Rybicki slide set for Deep Learning in Radiology / Medicine
Frankie Rybicki slide set for Deep Learning in Radiology / MedicineFrankie Rybicki slide set for Deep Learning in Radiology / Medicine
Frankie Rybicki slide set for Deep Learning in Radiology / Medicine
Frank Rybicki
 
Intra Cranial Pressure ( Icp ) Measurements Are Taken Via...
Intra Cranial Pressure ( Icp ) Measurements Are Taken Via...Intra Cranial Pressure ( Icp ) Measurements Are Taken Via...
Intra Cranial Pressure ( Icp ) Measurements Are Taken Via...
Michelle Love
 
“Towards Multi-Step Expert Advice for Cognitive Computing” - Dr. Achim Rettin...
“Towards Multi-Step Expert Advice for Cognitive Computing” - Dr. Achim Rettin...“Towards Multi-Step Expert Advice for Cognitive Computing” - Dr. Achim Rettin...
“Towards Multi-Step Expert Advice for Cognitive Computing” - Dr. Achim Rettin...
diannepatricia
 
A Clustering Method for Weak Signals to Support Anticipative Intelligence
A Clustering Method for Weak Signals to Support Anticipative IntelligenceA Clustering Method for Weak Signals to Support Anticipative Intelligence
A Clustering Method for Weak Signals to Support Anticipative Intelligence
CSCJournals
 
Probabilistic Modular Embedding for Stochastic Coordinated Systems
Probabilistic Modular Embedding for Stochastic Coordinated SystemsProbabilistic Modular Embedding for Stochastic Coordinated Systems
Probabilistic Modular Embedding for Stochastic Coordinated Systems
Stefano Mariani
 
DOJProposal7.doc
DOJProposal7.docDOJProposal7.doc
DOJProposal7.docbutest
 
On Machine Learning and Data Mining
On Machine Learning and Data MiningOn Machine Learning and Data Mining
On Machine Learning and Data Miningbutest
 
IRJET- Survey for Amazon Fine Food Reviews
IRJET- Survey for Amazon Fine Food ReviewsIRJET- Survey for Amazon Fine Food Reviews
IRJET- Survey for Amazon Fine Food Reviews
IRJET Journal
 
Semantic-Driven Design and Management of KDD Processes
Semantic-Driven Design and Management of KDD ProcessesSemantic-Driven Design and Management of KDD Processes
Semantic-Driven Design and Management of KDD Processes
Emanuele Storti
 
DOJProposal7.doc
DOJProposal7.docDOJProposal7.doc
DOJProposal7.docbutest
 
Hybrid Approach for Brain Tumour Detection in Image Segmentation
Hybrid Approach for Brain Tumour Detection in Image SegmentationHybrid Approach for Brain Tumour Detection in Image Segmentation
Hybrid Approach for Brain Tumour Detection in Image Segmentation
ijtsrd
 
Computational model for artificial learning using formal concept analysis
Computational model for artificial learning using formal concept analysisComputational model for artificial learning using formal concept analysis
Computational model for artificial learning using formal concept analysisAboul Ella Hassanien
 

Similar to Utility Theory, Minimum Effort, and Predictive Coding (20)

Machine Learning and Automatic Text Classification: What's Next?
Machine Learning and Automatic Text Classification: What's Next?Machine Learning and Automatic Text Classification: What's Next?
Machine Learning and Automatic Text Classification: What's Next?
 
Self-organisation of Knowledge in Socio-technical Systems: A Coordination Per...
Self-organisation of Knowledge in Socio-technical Systems: A Coordination Per...Self-organisation of Knowledge in Socio-technical Systems: A Coordination Per...
Self-organisation of Knowledge in Socio-technical Systems: A Coordination Per...
 
Size-Based Disciplines for Job Scheduling in Data-Intensive Scalable Computin...
Size-Based Disciplines for Job Scheduling in Data-Intensive Scalable Computin...Size-Based Disciplines for Job Scheduling in Data-Intensive Scalable Computin...
Size-Based Disciplines for Job Scheduling in Data-Intensive Scalable Computin...
 
A N E XTENSION OF P ROTÉGÉ FOR AN AUTOMA TIC F UZZY - O NTOLOGY BUILDING U...
A N  E XTENSION OF  P ROTÉGÉ FOR AN AUTOMA TIC  F UZZY - O NTOLOGY BUILDING U...A N  E XTENSION OF  P ROTÉGÉ FOR AN AUTOMA TIC  F UZZY - O NTOLOGY BUILDING U...
A N E XTENSION OF P ROTÉGÉ FOR AN AUTOMA TIC F UZZY - O NTOLOGY BUILDING U...
 
Daniel Samaan: ChatGPT and the Future of Work
Daniel Samaan: ChatGPT and the Future of WorkDaniel Samaan: ChatGPT and the Future of Work
Daniel Samaan: ChatGPT and the Future of Work
 
Grouping techniques for facing Volume and Velocity in the Big Data
Grouping techniques for facing Volume and Velocity in the Big DataGrouping techniques for facing Volume and Velocity in the Big Data
Grouping techniques for facing Volume and Velocity in the Big Data
 
DEFENSE
DEFENSEDEFENSE
DEFENSE
 
Frankie Rybicki slide set for Deep Learning in Radiology / Medicine
Frankie Rybicki slide set for Deep Learning in Radiology / MedicineFrankie Rybicki slide set for Deep Learning in Radiology / Medicine
Frankie Rybicki slide set for Deep Learning in Radiology / Medicine
 
Intra Cranial Pressure ( Icp ) Measurements Are Taken Via...
Intra Cranial Pressure ( Icp ) Measurements Are Taken Via...Intra Cranial Pressure ( Icp ) Measurements Are Taken Via...
Intra Cranial Pressure ( Icp ) Measurements Are Taken Via...
 
“Towards Multi-Step Expert Advice for Cognitive Computing” - Dr. Achim Rettin...
“Towards Multi-Step Expert Advice for Cognitive Computing” - Dr. Achim Rettin...“Towards Multi-Step Expert Advice for Cognitive Computing” - Dr. Achim Rettin...
“Towards Multi-Step Expert Advice for Cognitive Computing” - Dr. Achim Rettin...
 
A Clustering Method for Weak Signals to Support Anticipative Intelligence
A Clustering Method for Weak Signals to Support Anticipative IntelligenceA Clustering Method for Weak Signals to Support Anticipative Intelligence
A Clustering Method for Weak Signals to Support Anticipative Intelligence
 
Probabilistic Modular Embedding for Stochastic Coordinated Systems
Probabilistic Modular Embedding for Stochastic Coordinated SystemsProbabilistic Modular Embedding for Stochastic Coordinated Systems
Probabilistic Modular Embedding for Stochastic Coordinated Systems
 
DOJProposal7.doc
DOJProposal7.docDOJProposal7.doc
DOJProposal7.doc
 
On Machine Learning and Data Mining
On Machine Learning and Data MiningOn Machine Learning and Data Mining
On Machine Learning and Data Mining
 
IRJET- Survey for Amazon Fine Food Reviews
IRJET- Survey for Amazon Fine Food ReviewsIRJET- Survey for Amazon Fine Food Reviews
IRJET- Survey for Amazon Fine Food Reviews
 
Semantic-Driven Design and Management of KDD Processes
Semantic-Driven Design and Management of KDD ProcessesSemantic-Driven Design and Management of KDD Processes
Semantic-Driven Design and Management of KDD Processes
 
final seminar
final seminarfinal seminar
final seminar
 
DOJProposal7.doc
DOJProposal7.docDOJProposal7.doc
DOJProposal7.doc
 
Hybrid Approach for Brain Tumour Detection in Image Segmentation
Hybrid Approach for Brain Tumour Detection in Image SegmentationHybrid Approach for Brain Tumour Detection in Image Segmentation
Hybrid Approach for Brain Tumour Detection in Image Segmentation
 
Computational model for artificial learning using formal concept analysis
Computational model for artificial learning using formal concept analysisComputational model for artificial learning using formal concept analysis
Computational model for artificial learning using formal concept analysis
 

Recently uploaded

一比一原版(UofS毕业证书)萨省大学毕业证如何办理
一比一原版(UofS毕业证书)萨省大学毕业证如何办理一比一原版(UofS毕业证书)萨省大学毕业证如何办理
一比一原版(UofS毕业证书)萨省大学毕业证如何办理
v3tuleee
 
原版制作(Deakin毕业证书)迪肯大学毕业证学位证一模一样
原版制作(Deakin毕业证书)迪肯大学毕业证学位证一模一样原版制作(Deakin毕业证书)迪肯大学毕业证学位证一模一样
原版制作(Deakin毕业证书)迪肯大学毕业证学位证一模一样
u86oixdj
 
Learn SQL from basic queries to Advance queries
Learn SQL from basic queries to Advance queriesLearn SQL from basic queries to Advance queries
Learn SQL from basic queries to Advance queries
manishkhaire30
 
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
ahzuo
 
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
John Andrews
 
Levelwise PageRank with Loop-Based Dead End Handling Strategy : SHORT REPORT ...
Levelwise PageRank with Loop-Based Dead End Handling Strategy : SHORT REPORT ...Levelwise PageRank with Loop-Based Dead End Handling Strategy : SHORT REPORT ...
Levelwise PageRank with Loop-Based Dead End Handling Strategy : SHORT REPORT ...
Subhajit Sahu
 
一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理
一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理
一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理
ahzuo
 
一比一原版(Bradford毕业证书)布拉德福德大学毕业证如何办理
一比一原版(Bradford毕业证书)布拉德福德大学毕业证如何办理一比一原版(Bradford毕业证书)布拉德福德大学毕业证如何办理
一比一原版(Bradford毕业证书)布拉德福德大学毕业证如何办理
mbawufebxi
 
一比一原版(UniSA毕业证书)南澳大学毕业证如何办理
一比一原版(UniSA毕业证书)南澳大学毕业证如何办理一比一原版(UniSA毕业证书)南澳大学毕业证如何办理
一比一原版(UniSA毕业证书)南澳大学毕业证如何办理
slg6lamcq
 
一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理
一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理
一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理
oz8q3jxlp
 
Adjusting primitives for graph : SHORT REPORT / NOTES
Adjusting primitives for graph : SHORT REPORT / NOTESAdjusting primitives for graph : SHORT REPORT / NOTES
Adjusting primitives for graph : SHORT REPORT / NOTES
Subhajit Sahu
 
原版制作(swinburne毕业证书)斯威本科技大学毕业证毕业完成信一模一样
原版制作(swinburne毕业证书)斯威本科技大学毕业证毕业完成信一模一样原版制作(swinburne毕业证书)斯威本科技大学毕业证毕业完成信一模一样
原版制作(swinburne毕业证书)斯威本科技大学毕业证毕业完成信一模一样
u86oixdj
 
Everything you wanted to know about LIHTC
Everything you wanted to know about LIHTCEverything you wanted to know about LIHTC
Everything you wanted to know about LIHTC
Roger Valdez
 
ViewShift: Hassle-free Dynamic Policy Enforcement for Every Data Lake
ViewShift: Hassle-free Dynamic Policy Enforcement for Every Data LakeViewShift: Hassle-free Dynamic Policy Enforcement for Every Data Lake
ViewShift: Hassle-free Dynamic Policy Enforcement for Every Data Lake
Walaa Eldin Moustafa
 
Data_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptx
Data_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptxData_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptx
Data_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptx
AnirbanRoy608946
 
My burning issue is homelessness K.C.M.O.
My burning issue is homelessness K.C.M.O.My burning issue is homelessness K.C.M.O.
My burning issue is homelessness K.C.M.O.
rwarrenll
 
Ch03-Managing the Object-Oriented Information Systems Project a.pdf
Ch03-Managing the Object-Oriented Information Systems Project a.pdfCh03-Managing the Object-Oriented Information Systems Project a.pdf
Ch03-Managing the Object-Oriented Information Systems Project a.pdf
haila53
 
Analysis insight about a Flyball dog competition team's performance
Analysis insight about a Flyball dog competition team's performanceAnalysis insight about a Flyball dog competition team's performance
Analysis insight about a Flyball dog competition team's performance
roli9797
 
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
Timothy Spann
 
The Building Blocks of QuestDB, a Time Series Database
The Building Blocks of QuestDB, a Time Series DatabaseThe Building Blocks of QuestDB, a Time Series Database
The Building Blocks of QuestDB, a Time Series Database
javier ramirez
 

Recently uploaded (20)

一比一原版(UofS毕业证书)萨省大学毕业证如何办理
一比一原版(UofS毕业证书)萨省大学毕业证如何办理一比一原版(UofS毕业证书)萨省大学毕业证如何办理
一比一原版(UofS毕业证书)萨省大学毕业证如何办理
 
原版制作(Deakin毕业证书)迪肯大学毕业证学位证一模一样
原版制作(Deakin毕业证书)迪肯大学毕业证学位证一模一样原版制作(Deakin毕业证书)迪肯大学毕业证学位证一模一样
原版制作(Deakin毕业证书)迪肯大学毕业证学位证一模一样
 
Learn SQL from basic queries to Advance queries
Learn SQL from basic queries to Advance queriesLearn SQL from basic queries to Advance queries
Learn SQL from basic queries to Advance queries
 
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
 
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
 
Levelwise PageRank with Loop-Based Dead End Handling Strategy : SHORT REPORT ...
Levelwise PageRank with Loop-Based Dead End Handling Strategy : SHORT REPORT ...Levelwise PageRank with Loop-Based Dead End Handling Strategy : SHORT REPORT ...
Levelwise PageRank with Loop-Based Dead End Handling Strategy : SHORT REPORT ...
 
一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理
一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理
一比一原版(UIUC毕业证)伊利诺伊大学|厄巴纳-香槟分校毕业证如何办理
 
一比一原版(Bradford毕业证书)布拉德福德大学毕业证如何办理
一比一原版(Bradford毕业证书)布拉德福德大学毕业证如何办理一比一原版(Bradford毕业证书)布拉德福德大学毕业证如何办理
一比一原版(Bradford毕业证书)布拉德福德大学毕业证如何办理
 
一比一原版(UniSA毕业证书)南澳大学毕业证如何办理
一比一原版(UniSA毕业证书)南澳大学毕业证如何办理一比一原版(UniSA毕业证书)南澳大学毕业证如何办理
一比一原版(UniSA毕业证书)南澳大学毕业证如何办理
 
一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理
一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理
一比一原版(Deakin毕业证书)迪肯大学毕业证如何办理
 
Adjusting primitives for graph : SHORT REPORT / NOTES
Adjusting primitives for graph : SHORT REPORT / NOTESAdjusting primitives for graph : SHORT REPORT / NOTES
Adjusting primitives for graph : SHORT REPORT / NOTES
 
原版制作(swinburne毕业证书)斯威本科技大学毕业证毕业完成信一模一样
原版制作(swinburne毕业证书)斯威本科技大学毕业证毕业完成信一模一样原版制作(swinburne毕业证书)斯威本科技大学毕业证毕业完成信一模一样
原版制作(swinburne毕业证书)斯威本科技大学毕业证毕业完成信一模一样
 
Everything you wanted to know about LIHTC
Everything you wanted to know about LIHTCEverything you wanted to know about LIHTC
Everything you wanted to know about LIHTC
 
ViewShift: Hassle-free Dynamic Policy Enforcement for Every Data Lake
ViewShift: Hassle-free Dynamic Policy Enforcement for Every Data LakeViewShift: Hassle-free Dynamic Policy Enforcement for Every Data Lake
ViewShift: Hassle-free Dynamic Policy Enforcement for Every Data Lake
 
Data_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptx
Data_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptxData_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptx
Data_and_Analytics_Essentials_Architect_an_Analytics_Platform.pptx
 
My burning issue is homelessness K.C.M.O.
My burning issue is homelessness K.C.M.O.My burning issue is homelessness K.C.M.O.
My burning issue is homelessness K.C.M.O.
 
Ch03-Managing the Object-Oriented Information Systems Project a.pdf
Ch03-Managing the Object-Oriented Information Systems Project a.pdfCh03-Managing the Object-Oriented Information Systems Project a.pdf
Ch03-Managing the Object-Oriented Information Systems Project a.pdf
 
Analysis insight about a Flyball dog competition team's performance
Analysis insight about a Flyball dog competition team's performanceAnalysis insight about a Flyball dog competition team's performance
Analysis insight about a Flyball dog competition team's performance
 
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
 
The Building Blocks of QuestDB, a Time Series Database
The Building Blocks of QuestDB, a Time Series DatabaseThe Building Blocks of QuestDB, a Time Series Database
The Building Blocks of QuestDB, a Time Series Database
 

Utility Theory, Minimum Effort, and Predictive Coding

  • 1. Utility Theory, Minimum Effort, and Predictive Coding Fabrizio Sebastiani Istituto di Scienza e Tecnologie dell’Informazione Consiglio Nazionale delle Ricerche 56124 Pisa, Italy DESI V – Roma, IT, 14 June 2013
  • 2. What I’ll be talking about A talk about text classification (“predictive coding”), about humans in the loop, and about how to best support their work I will be looking at scenarios in which 1 text classification technology is used for identifying documents belonging to a given class / relevant to a given query ... 2 ... but the level of accuracy that can be obtained from the classifier is not considered sufficient ... 3 ... with the consequence that one or more human assessors are asked to inspect (and correct where appropriate) a portion of the classification decisions, with the goal of increasing overall accuracy. How can we support / optimize the work of the human assessors? Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 2 / 36
  • 3. What I’ll be talking about A talk about text classification (“predictive coding”), about humans in the loop, and about how to best support their work I will be looking at scenarios in which 1 text classification technology is used for identifying documents belonging to a given class / relevant to a given query ... 2 ... but the level of accuracy that can be obtained from the classifier is not considered sufficient ... 3 ... with the consequence that one or more human assessors are asked to inspect (and correct where appropriate) a portion of the classification decisions, with the goal of increasing overall accuracy. How can we support / optimize the work of the human assessors? Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 2 / 36
  • 4. What I’ll be talking about A talk about text classification (“predictive coding”), about humans in the loop, and about how to best support their work I will be looking at scenarios in which 1 text classification technology is used for identifying documents belonging to a given class / relevant to a given query ... 2 ... but the level of accuracy that can be obtained from the classifier is not considered sufficient ... 3 ... with the consequence that one or more human assessors are asked to inspect (and correct where appropriate) a portion of the classification decisions, with the goal of increasing overall accuracy. How can we support / optimize the work of the human assessors? Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 2 / 36
  • 5. A worked out example predicted Y N true Y TP = 4 FP = 3 N FN = 4 TN = 9 F1 = 2TP 2TP + FP + FN = 0.53 Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 3 / 36
  • 6. A worked out example (cont’d) predicted Y N true Y TP = 4 FP = 3 N FN = 4 TN = 9 F1 = 2TP 2TP + FP + FN = 0.53 Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 4 / 36
  • 7. A worked out example (cont’d) predicted Y N true Y TP = 5 FP = 3 N FN = 3 TN = 9 F1 = 2TP 2TP + FP + FN = 0.63 Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 5 / 36
  • 8. A worked out example (cont’d) predicted Y N true Y TP = 5 FP = 2 N FN = 3 TN = 10 F1 = 2TP 2TP + FP + FN = 0.67 Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 6 / 36
  • 9. A worked out example (cont’d) predicted Y N true Y TP = 6 FP = 2 N FN = 2 TN = 10 F1 = 2TP 2TP + FP + FN = 0.75 Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 7 / 36
  • 10. A worked out example (cont’d) predicted Y N true Y TP = 6 FP = 1 N FN = 2 TN = 11 F1 = 2TP 2TP + FP + FN = 0.80 Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 8 / 36
  • 11. What I’ll be talking about (cont’d) We need methods that given a desired level of accuracy, minimize the assessors’ effort necessary to achieve it; alternatively, given an available amount of human assessors’ effort, maximize the accuracy that can be obtained through it This can be achieved by ranking the automatically classified documents in such a way that, by starting the inspection from the top of the ranking, the cost-effectiveness of the annotators’ work is maximized We call the task of generating such a ranking Semi-Automatic Text Classification (SATC) Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 9 / 36
  • 12. What I’ll be talking about (cont’d) We need methods that given a desired level of accuracy, minimize the assessors’ effort necessary to achieve it; alternatively, given an available amount of human assessors’ effort, maximize the accuracy that can be obtained through it This can be achieved by ranking the automatically classified documents in such a way that, by starting the inspection from the top of the ranking, the cost-effectiveness of the annotators’ work is maximized We call the task of generating such a ranking Semi-Automatic Text Classification (SATC) Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 9 / 36
  • 13. What I’ll be talking about (cont’d) Previous work has addressed SATC via techniques developed for “active learning” In both cases, the automatically classified documents are ranked with the goal of having the human annotator start inspecting/correcting from the top; however in active learning the goal is providing new training examples in SATC the goal is increasing the overall accuracy of the classified set We claim that a ranking generated “à la active learning” is suboptimal for SATC1 1G Berardi, A Esuli, F Sebastiani. A Utility-Theoretic Ranking Method for Semi-Automated Text Classification. Proceedings of the 35th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2012), Portland, US, 2012. Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 10 / 36
  • 14. What I’ll be talking about (cont’d) Previous work has addressed SATC via techniques developed for “active learning” In both cases, the automatically classified documents are ranked with the goal of having the human annotator start inspecting/correcting from the top; however in active learning the goal is providing new training examples in SATC the goal is increasing the overall accuracy of the classified set We claim that a ranking generated “à la active learning” is suboptimal for SATC1 1G Berardi, A Esuli, F Sebastiani. A Utility-Theoretic Ranking Method for Semi-Automated Text Classification. Proceedings of the 35th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2012), Portland, US, 2012. Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 10 / 36
  • 15. Outline of this talk 1 We discuss how to measure “error reduction” (i.e., increase in accuracy) 2 We discuss a method for maximizing the expected error reduction for a fixed amount of annotation effort 3 We show some promising experimental results Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 11 / 36
  • 16. Error Reduction, and How to Measure it Outline 1 Error Reduction, and How to Measure it 2 Error Reduction, and How to Maximize it 3 Some Experimental Results Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 12 / 36
  • 17. Error Reduction, and How to Measure it Error Reduction, and how to measure it Assume we have 1 class (or “query”) c; 2 classifier h for c; 3 set of unlabeled documents D that we have automatically classified by means of h, so that every document in D is associated with a binary decision (Y or N) with a confidence score (a positive real number) 4 measure of accuracy A, ranging on [0,1] Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 13 / 36
  • 18. Error Reduction, and How to Measure it Error Reduction, and how to Measure it (cont’d) We will assume that A is F1 = 2 · Precision · Recall Precision + Recall = 2 · TP (2 · TP) + FP + FN but any “set-based” measure of accuracy (i.e., based on a contingency table) may be used An amount of error, measured as E = (1 − A), is present in the automatically classified set D Human annotators inspect-and-correct a portion of D with the goal of reducing the error present in D Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 14 / 36
  • 19. Error Reduction, and How to Measure it Error Reduction, and how to Measure it (cont’d) We will assume that A is F1 = 2 · Precision · Recall Precision + Recall = 2 · TP (2 · TP) + FP + FN but any “set-based” measure of accuracy (i.e., based on a contingency table) may be used An amount of error, measured as E = (1 − A), is present in the automatically classified set D Human annotators inspect-and-correct a portion of D with the goal of reducing the error present in D Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 14 / 36
  • 20. Error Reduction, and How to Measure it Error Reduction, and how to Measure it (cont’d) We define error at rank n (noted as E(n)) as the error still present in D after the annotator has inspected the documents at the first n rank positions E(0) is the initial error generated by the automated classifier E(|D|) is 0 We define error reduction at rank n (noted as ER(n)) to be ER(n) = E(0) − E(n) E(0) the error reduction obtained by the annotator who inspects the docs at the first n rank positions ER(n) ∈ [0, 1] ER(n) = 0 indicates no reduction ER(n) = 1 indicates total elimination of error Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 15 / 36
  • 21. Error Reduction, and How to Measure it Error Reduction, and how to Measure it (cont’d) We define error at rank n (noted as E(n)) as the error still present in D after the annotator has inspected the documents at the first n rank positions E(0) is the initial error generated by the automated classifier E(|D|) is 0 We define error reduction at rank n (noted as ER(n)) to be ER(n) = E(0) − E(n) E(0) the error reduction obtained by the annotator who inspects the docs at the first n rank positions ER(n) ∈ [0, 1] ER(n) = 0 indicates no reduction ER(n) = 1 indicates total elimination of error Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 15 / 36
  • 22. Error Reduction, and How to Measure it Error Reduction, and how to Measure it (cont’d) 0.0 0.2 0.4 0.6 0.8 1.0 Inspection Length 0.0 0.2 0.4 0.6 0.8 1.0 ErrorReduction(ER) Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 16 / 36
  • 23. Error Reduction, and How to Maximize it Outline 1 Error Reduction, and How to Measure it 2 Error Reduction, and How to Maximize it 3 Some Experimental Results Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 17 / 36
  • 24. Error Reduction, and How to Maximize it Error Reduction, and how to Maximize it Problem How should we rank the documents in D so as to maximize the expected error reduction? Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 18 / 36
  • 25. Error Reduction, and How to Maximize it A worked out example predicted Y N true Y TP = 4 FP = 3 N FN = 4 TN = 9 F1 = 2TP 2TP + FP + FN = 0.53 Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 19 / 36
  • 26. Error Reduction, and How to Maximize it A worked out example (cont’d) predicted Y N true Y TP = 4 FP = 3 N FN = 4 TN = 9 F1 = 2TP 2TP + FP + FN = 0.53 Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 20 / 36
  • 27. Error Reduction, and How to Maximize it A worked out example (cont’d) predicted Y N true Y TP = 5 FP = 3 N FN = 3 TN = 9 F1 = 2TP 2TP + FP + FN = 0.63 Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 21 / 36
  • 28. Error Reduction, and How to Maximize it A worked out example (cont’d) predicted Y N true Y TP = 5 FP = 2 N FN = 3 TN = 10 F1 = 2TP 2TP + FP + FN = 0.67 Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 22 / 36
  • 29. Error Reduction, and How to Maximize it A worked out example (cont’d) predicted Y N true Y TP = 6 FP = 2 N FN = 2 TN = 10 F1 = 2TP 2TP + FP + FN = 0.75 Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 23 / 36
  • 30. Error Reduction, and How to Maximize it A worked out example (cont’d) predicted Y N true Y TP = 6 FP = 1 N FN = 2 TN = 11 F1 = 2TP 2TP + FP + FN = 0.80 Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 24 / 36
  • 31. Error Reduction, and How to Maximize it Error Reduction, and how to Maximize it Problem: how should we rank the documents in D so as to maximize the expected error reduction? Intuition 1: Documents that have a higher probability of being misclassified should be ranked higher Intuition 2: Documents that, if corrected, bring about a higher gain (i.e., a bigger impact on A) should be ranked higher Here, consider that a false positive and a false negative may have different impacts on A (e.g., when A ≡ Fβ, for any value of β) Bottom line Documents that have a higher utility (= probability × gain) should be ranked higher Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 25 / 36
  • 32. Error Reduction, and How to Maximize it Error Reduction, and how to Maximize it Problem: how should we rank the documents in D so as to maximize the expected error reduction? Intuition 1: Documents that have a higher probability of being misclassified should be ranked higher Intuition 2: Documents that, if corrected, bring about a higher gain (i.e., a bigger impact on A) should be ranked higher Here, consider that a false positive and a false negative may have different impacts on A (e.g., when A ≡ Fβ, for any value of β) Bottom line Documents that have a higher utility (= probability × gain) should be ranked higher Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 25 / 36
  • 33. Error Reduction, and How to Maximize it Error Reduction, and how to Maximize it (cont’d) Given a set Ω of mutually disjoint events, a utility function is defined as U(Ω) = ω∈Ω P(ω)G(ω) where P(ω) is the probability of occurrence of event ω G(ω) is the gain obtained if event ω occurs We can thus estimate the utility, for the aims of increasing A, of manually inspecting a document d as U(TP, TN, FP, FN) = P(FP) · G(FP) + P(FN) · G(FN) provided we can estimate If d is labelled with class c: P(FP) and G(FP) If d is not labelled with class c: P(FN) and G(FN) Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 26 / 36
  • 34. Error Reduction, and How to Maximize it Error Reduction, and how to Maximize it (cont’d) Given a set Ω of mutually disjoint events, a utility function is defined as U(Ω) = ω∈Ω P(ω)G(ω) where P(ω) is the probability of occurrence of event ω G(ω) is the gain obtained if event ω occurs We can thus estimate the utility, for the aims of increasing A, of manually inspecting a document d as U(TP, TN, FP, FN) = P(FP) · G(FP) + P(FN) · G(FN) provided we can estimate If d is labelled with class c: P(FP) and G(FP) If d is not labelled with class c: P(FN) and G(FN) Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 26 / 36
  • 35. Error Reduction, and How to Maximize it Error Reduction, and how to Maximize it (cont’d) Estimating P(FP) and P(FN) (the probability of misclassification) can be done by converting the confidence score returned by the classifier into a probability of correct classification Tricky: requires probability “calibration” via a generalized sigmoid function to be optimized via k-fold cross-validation Gains G(FP) and G(FN) can be defined “differentially”; i.e., The gain obtained by correcting a FN is (AFN→TP − A) The gain obtained by correcting a FP is (AFP→TN − A) Gains need to be estimated by estimating the contingency table on the training set via k-fold cross-validation Key observation: in general, G(FP) = G(FN) Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 27 / 36
  • 36. Error Reduction, and How to Maximize it Error Reduction, and how to Maximize it (cont’d) Estimating P(FP) and P(FN) (the probability of misclassification) can be done by converting the confidence score returned by the classifier into a probability of correct classification Tricky: requires probability “calibration” via a generalized sigmoid function to be optimized via k-fold cross-validation Gains G(FP) and G(FN) can be defined “differentially”; i.e., The gain obtained by correcting a FN is (AFN→TP − A) The gain obtained by correcting a FP is (AFP→TN − A) Gains need to be estimated by estimating the contingency table on the training set via k-fold cross-validation Key observation: in general, G(FP) = G(FN) Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 27 / 36
  • 37. Error Reduction, and How to Maximize it Error Reduction, and how to Maximize it (cont’d) Estimating P(FP) and P(FN) (the probability of misclassification) can be done by converting the confidence score returned by the classifier into a probability of correct classification Tricky: requires probability “calibration” via a generalized sigmoid function to be optimized via k-fold cross-validation Gains G(FP) and G(FN) can be defined “differentially”; i.e., The gain obtained by correcting a FN is (AFN→TP − A) The gain obtained by correcting a FP is (AFP→TN − A) Gains need to be estimated by estimating the contingency table on the training set via k-fold cross-validation Key observation: in general, G(FP) = G(FN) Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 27 / 36
  • 38. Some Experimental Results Outline 1 Error Reduction, and How to Measure it 2 Error Reduction, and How to Maximize it 3 Some Experimental Results Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 28 / 36
  • 39. Some Experimental Results Some Experimental Results Learning algorithms: MP-Boost, SVMs Datasets: # Cats # Training # Test FM 1 MP-Boost FM 1 SVMs Reuters-21578 115 9603 3299 0.608 0.527 OHSUMED-S 97 12358 3652 0.479 0.478 Baseline: ranking by probability of misclassification, equivalent to applying our ranking method with G(FP) = G(FN) = 1 Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 29 / 36
  • 40. Some Experimental Results 0.0 0.2 0.4 0.6 0.8 1.0 Inspection Length 0.0 0.2 0.4 0.6 0.8 1.0 ErrorReduction(ER)Learner: MP-Boost; Dataset: Reuters-21578; Type: Macro Random Baseline Utility-theoretic Oracle Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 30 / 36
  • 41. Some Experimental Results 0.0 0.2 0.4 0.6 0.8 1.0 Inspection Length 0.0 0.2 0.4 0.6 0.8 1.0 ErrorReduction(ER) Learner: SVMs; Dataset: Reuters-21578; Type: Macro Random Baseline Utility-theoretic Oracle Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 31 / 36
  • 42. Some Experimental Results 0.0 0.2 0.4 0.6 0.8 1.0 Inspection Length 0.0 0.2 0.4 0.6 0.8 1.0 ErrorReduction(ER) Learner: MP-Boost; Dataset: Ohsumed-S; Type: Macro Random Baseline Utility-theoretic Oracle Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 32 / 36
  • 43. Some Experimental Results 0.0 0.2 0.4 0.6 0.8 1.0 Inspection Length 0.0 0.2 0.4 0.6 0.8 1.0ErrorReduction(ER) Learner: SVMs; Dataset: Ohsumed-S; Type: Macro Random Baseline Utility-theoretic Oracle Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 33 / 36
  • 44. Some Experimental Results A few side notes This approach allows the human annotator to know, at any stage of the inspection process, what the estimated accuracy is at that stage Estimate accuracy at the beginning of the process, via k-fold cross validation Update after each correction is made This approach lends itself to having more than one assessor working in parallel on the same inspection task Recent research I have not discussed today : A “dynamic” SATC method in which gains are updated after each correction is performed “Microaveraging” and “Macroaveraging” -oriented methods Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 34 / 36
  • 45. Some Experimental Results Concluding Remarks Take-away message: Semi-automatic text classification needs to be addressed as a task in its own right Active learning typically makes use of probabilities of misclassification but does not make use of gains ⇒ ranking “à la active learning” is suboptimal for SATC The use of utility theory means that the ranking algorithm is optimized for a specific accuracy measure ⇒ Choose the accuracy measure the best mirrors your applicative needs (e.g., Fβ with β > 1), and choose it well! SATC is important, since in more and more application contexts the accuracy obtainable via completely automatic text classification is not sufficient; more and more frequently humans will need to enter the loop Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 35 / 36
  • 46. Some Experimental Results Thank you! Fabrizio Sebastiani (ISTI-CNR Pisa (Italy)) Utility Theory, Minimum Effort, and Predictive Coding DESI V – Roma, IT, 14 June 2013 36 / 36