See discussions, stats, and author profiles for this publication at: [Link]

net/publication/322315484

Multi-label Classification With Weighted Labels Using Learning Classifier

Systems

Conference Paper · December 2017

DOI: 10.1109/ICMLA.2017.0-147

CITATIONS READS
5 832

3 authors, including:

Shabnam Nazmi Mohammad Razeghi Jahromi

North Carolina Agricultural and Technical State University ABB
14 PUBLICATIONS   18 CITATIONS    13 PUBLICATIONS   62 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Genetic-based machine learning for multi-label classification View project

Incorporating expert knowledge into machine learning using possibility theory View project

All content following this page was uploaded by Mohammad Razeghi Jahromi on 26 March 2018.

The user has requested enhancement of the downloaded file.

 2017 16th IEEE International Conference on Machine Learning and Applications

Multi-label Classification With Weighted Labels

Using Learning Classifier Systems
Shabnam Nazmia , Mohammad Razeghi-Jahromib , Abdollah Homaifarc *
School of Electrical and Computer Engineering
North Carolina A&T State University
a snazmi@[Link], b mrazeghijahromi@[Link], c homaifar@[Link]

Abstract—In this work the Michigan style strength-based confidence levels that are possibly provided by the expert or
learning classifier system, which is a rule-based supervised a known probability distribution. Thus, the first contribution
learning algorithm, is extended by changing its action space rep- of this work is a learning algorithm that learns the labels
resentation to handle multi-label classification tasks. Moreover, it
is assumed that the class membership for training data is partially with their respective confidence values. This goal is achieved
known and the uncertainty is represented by confidence values by employing a learning classifier system (LCS) [6], [7]
that reflects the probability of each label being true. Here, it that is a rule-based evolutionary classifier system. In learning
is assumed that confidence values are provided by the expert classifiers, the complete model is formed of a population of
or generated from a probability distribution function. Necessary rules in the form of IF condition-THEN action that solve
parameters are introduced, and learning classifiers are modified
to learn simultaneously the confidence level and multi-label in the classification task cooperatively. Each rule (also called
the training data. Therefore, the classifier’s output for unseen classifier) has some parameters that evolve during the learning
data is a set of labels and their respective confidence values. process to model the problem. The genetic algorithm (GA)
Additionally, to quantify the classifier performance, a novel loss helps the algorithm to enumerate the population with the fittest
measure is introduced that generalizes the well-known Hamming and most useful rules. LCSs use a scheme of reinforcement
loss criteria to takes into account the classification error and
confidence estimation error simultaneously. The algorithm is learning [8] that guides the population towards the solution by
tested on one real-world data and two synthetic data sets. Results rewarding and punishing its actions.
show the ability of the model in learning multi-class and multi- When dealing with multi-label data, association of a sample
label data with low confidence estimation error. to multiple classes simultaneously becomes more uncertain
and employing a confidence-based labeling helps to train a
Keywords: Learning classifier systems, Multi-label classi- more reliable multi-label classifier. Multi-label classification
fication, Uncertain labels, Evolutionary based learning (MLC) problems have a broad range of approaches that solve
I. I NTRODUCTION the MLC task from different perspectives. There are a number
of methods that transform the original multi-label data set
In supervised classification tasks, usually it is assumed that
into single-label data sets to facilitate the use of existing
available data is confidently labeled by a human. Nonetheless,
classification techniques. Binary relevance (BR) [9], [10] is a
in many applications human supervision is imprecise, diffi-
popular problem transformation method that trains k separate
cult or expensive, therefore it is useful that if available, to
binary classifiers, one for each class. Label powerset (LP)
incorporate some level of confidence on the labels assigned to
[10] deals with the unique combinations of labels that exist
the samples while training the model. Some of the supervised
in data as a new label. To improve its efficiency, [11] has
learning approaches have been adapted to handle uncertainty
considered a random combination of labels with different sizes
in labels. In many studies the Dempster-Shafer (DS) belief
as label sets to train classifier. The second family of MLC
theory is employed to process the uncertainty in labels, which
approaches are algorithm adaptation methods, that modify
can include probabilistic, possibilistic or imprecise labels. In
the existing classification algorithms to adapt to the MLC
[1] and [2] belief decision trees [3] are used to handle multiple
task. For instance, in [12] a lazy multi-label classification
label data. In [4] and [5], each sample is generated from a
algorithm using k-NN method is proposed. In [13], the MLC
mixture model assigned with a ”soft” label defined as a DS
task is formalized within the Bayes probability setting that
basic belief and maximum likelihood estimation is extended
explicitly takes into account the label dependence. In [14],
to solve the problem.
using boosting of decision tree models, each weak hypothesis
The sparse literature that deals specifically with partial
produces not only a class label, but also its confidence in
knowledge provided for sample-label association, has moti-
the prediction. However, this approach does not consider
vated the authors to try and find a model that takes the
the possible, a-priori information about label confidences in
maximum benefit from the information contained in the data
training data.
set. In this work, the objective is to represent the partial
In [15], classifier action is adapted to learn BR vectors and
information contained in uncertainly labeled part of data with
evolve default hierarchies within multi-label data using orga-
*Corresponding author. nization classifier systems, though results are reported only

0-7695-6321-X/17/31.00 ©2017 IEEE 275

DOI 10.1109/ICMLA.2017.0-147
for a small binary data set with three bits and three classes. based on its accuracy in labeling instances. UCS [17], [18]
In this work, the same representation is used to extend the is derived from XCS that works in a supervised learning
LCS model to deal with multi-label data. Thus, in the multi- framework, and ExSTraCS [19], [20] that has shown to solve
label learning classifier systems (MLCS), algorithm training heterogeneous supervised learning problems with noisy data.
instances can belong to more than one class with respective In this work strength-based classifier systems [6] are chosen
confidence levels. In contrast to most problem transformation that are primarily developed for binary-valued problems and
methods, a single model is trained and is able to predict a set have shown to successfully solve the Boolean Multiplexer
of labels for unseen samples, as well as its confidence in each benchmark problem [16] up to 20-multiplexer (multiplexer
of those labels. Note that, for a certain sample the sum of the with four address bits). Nonetheless, strength-based classifiers
values in the confidence vector does not have to sum up to have not been proven to be efficient is solving complex
one. This is because the sample can belong to more than one problems with real-valued attributes or problems with large
class with complete confidence (i.e. one) for each label. number of data. However, their simpler architecture compared
Finally, a novel accuracy measure is proposed that reflects to XCS or ExSTraCS, lets us interpret the impact of proposed
the classification error of the model and its confidence esti- extension on the process of learning. In the proposed method,
mation error simultaneously. classifier system is extended to solve problems with real-
The rest of this paper is organized as follows. In the next valued attributes to show the actual potential of the proposed
section, a summary of the notations and abbreviations that weighted-label multi-label classifier structure. This idea can
are used in the paper is provided. Next, learning classifier be extended to XCS or UCS classifier systems to benefit
systems are introduced and the proposed structure and required their powerful inference capability, specially for problems with
modifications to handle multi-label data and confidence level mixed attributes and larger number of samples [19], [21].
in labels is explained in detail. Next, training results for three In this part LCS is formulated and discussed to solve the
data sets are provided and are compared to other techniques. problem of MLC, but it is easily applicable to multi-class
Finally conclusions and future work are provided. problems. In LCS each rule has a condition that represents
the input space, an action that represents the labels, a strength
II. N OTATIONS value (S) that shows its usefulness, and finally an experience
The notations that are used in this paper are as follows. that is a measure of the rule’s contribution in training. In
It is assumed that X ⊂ Rm represents the input space and addition to the parameters original to strength-based classifiers,
Y = {y1 , y2 , . . . , yk } is the finite set of class labels. Each a confidence estimate value and its respective error (ε) that
multi-label instance x ∈ X is associated with a subset of will effect the amount of rule’s participation in learning and
classes as y ⊂ Y (for single label data y is a single label), its reward from the environment is introduced to each rule.
that may have its confidence levels (C ∈ Rk ) assigned to each For binary features, the condition part is composed of
label. If no confidence is assigned to samples, it is assumed to ternary alphabet with ‘#‘ representing a wild card that matches
be one by default and for each label-set assigned to instances, with either 1 or 0. For real-valued features, the center-spread
a binary relevance vector λ ∈ Rk is created. Therefore, representation [22] is employed, which represents the ith
D = {(x1 , λ1 , C1 ), (x2 , λ2 , C2 ), . . . , (xn , λn , Cn )} is the set attribute in the classifier condition by a pair (ci , si ). In other
of training instances such that for every λi and j = 1, . . . , k word, each attribute covers the interval (ci − si , ci + si ) of
we have, its respective feature space. In contrast to widely used binary
 coding, the action of classifier is the binary relevance of the
1 yj ∈ y
λi (j) = (1) class labels and can represent association to more than one
0 yj ∈ /y
class at a time [15].
If a sample does not belong to a specific class, its respective LCS starts with an empty population and gradually learns
λ(j) value is zero and thus no confidence is defined for it. the problem by generating new rules to cover a new instance.
If it belongs to a specific class, its λ(j) value will be one For every new rule, its strength initiates with a small value
and its confidence value can take a value between zero and (S0 ) and for correctly classified instances, its strength will
one. |A| represents the cardinality of set A, dH (a, b) is the increase through the external reward. Confidence estimates
hamming distance between two sets, and dEH (a, b) stands will initialize with zero and by means of a simple update
for the extended hamming distance between two sets that is rule, gradually approaches the correct value provided for the
introduced in this work. Therefore, the objective is to find a instances. In addition, error and experience are initialized at
hypothesis H : X → (Y ×C ), that for each instance generates zero.
a set of predicted label binary relevance vector (λ̂), along with The process of learning begins by providing one instance at
their respective confidence values (Ĉ ). a time (xl , λl , Cl ). Classifiers in the population are scanned to
find rules that have matching conditions with the sample. A
III. L EARNING CLASSIFIER SYSTEM WITH WEIGHTED classifier with binary attributes will match an input if for every
LABELS FOR MULTI - LABEL DATA
attribute it has either identical values with the given instance or
Learning classifiers have many variations that aim to solve the wild card (#), which matches every value of the attribute.
different problem domains. In XCS [16], classifier fitness is For real-valued attributes the condition ci − si < xi < ci + si

276
must satisfy for every attribute. If there is no match for the Next, classifier confidence value is updated using the delta
current sample, the covering process [16] creates a new rule rule
with a random condition that matches the sample and has its
Ĉi := Ĉi + β|Cl − Ĉi |. (4)
correct label as its action. For this purpose, with probability
equal to 1 − P# the attribute in the condition is specified, i.e. Classifier confidence estimate is updated using the mayonne
for binary-valued attributes the value of the instance (xl ) at adaptive modifée [23], according to which parameters are set
the ith position is considered as the ith bit of condition and to the average of the so far encountered cases as long as the
for real-valued attribute as the center and a random spread average update causes a stronger change than the update due
is generated with respect to the range of each attribute. With to the learning rate β. If the winning action is correct, an
probability equal to P# , the attribute will be a wild card. The external reward will be distributed among classifiers in [A]
new rule is placed into [P ] until the population size reaches proportional to their strength and error values as
its limit. Si e−αεi
Rules with matching conditions constitute the match list Ri =  R .
−αεi 0
(5)
([M ]) and bid a portion of their strength to compete for i∈[A] Si e

labeling the sample and possibly win the external reward. A R0 is the total reward provided by the environment. Here the
classifiers with higher strength and lower error should have resource sharing scheme is an extension of FPRS [24] to reflect
more potential to outbid others, therefore the bid (B) that was classifier confidence estimate accuracy as well. Thus, classifier
originally a function of strength and specificity (μ), has been strength is a measure of its correct prediction of class label,
modified to reflect classifier error as well. Therefore for rulei in addition to its correct estimation of confidence value.
the bid value is Finally, classifier strength is updated following (6) by
adding a reward and deducting taxes. Also, from the rest of
Bi = Cbid Si μi e−αεi . (2) the population that are a part of [M ] or remained in [P ], some
tax values are deducted that are proportional to their strengths
In this equation, 0 < α < 1 is a constant that defines the
accepted threshold for classifier error to be effective and Ci is Si := Si (1 − Ctax − Cbid ) + Ri . (6)
the bid coefficient. The motivation for integrating the classifier
Therefore, one iteration of learning finishes and LCS waits
error in its bid is that, it will create a pressure toward a
for the next training instance. This process continues for the
population with smaller confidence estimation error. Without
number of defined iterations.
this parameter during the bidding process, the model might
Moreover, at an initially defined frequency, genetic algo-
converge to a population with rules that are equally incorrect
rithm selects two individuals from [A] (niche GA [7]) and
in confidence estimation and their survival is only affected by
creates two off-springs after cross-over and mutation. Created
their strength. Therefore, there is a probability that the model
off-springs are replaced into [P ] for rules with the lowest
may trap in local solutions.
strength. Note that in interaction with genetic algorithm,
After the winner is selected, an action list ([A]) is formed
classifier strength acts as its fitness value, therefore GA tends
from those classifiers in [M ] that advocate exactly the same
to favor classifiers with high strength and low error. Procedure
action as the winning action. The winning action is then
represented in (1) shows the flow of the learning algorithm
executed, a possible reward is received and the parameters
during training.
of the classifiers in [A] are updated. Following (3), classifier
In the test stage, for each test data, the population of
errors is updated to reflect its confidence estimate distance
classifiers is scanned to find those with a matching condi-
with the actual confidence of the sample as
tion part. Assume that there are sub-populations offering r
εi = Ĉi − Cl 1 . (3) different label-sets, such that M = {m1 , . . . , mr } represents
the number of classifiers advocating each of those label-
The parameter that controls the minimum effective value of sets. Class labels predicted by the model in the union of
ε for each rule is α; the larger this coefficient is, the less the labels existing in r unique label set. To calculated the
chance a rule has to win the bid and has a smaller share collective confidence votes of the model for each of the labels,
of external reward. Since GA uses strength as the fitness first the average confidence estimate of each subpopulation is
measure of rules, selecting a larger value for α, decreases calculated as, C̄ = {C̄1 , . . . , C̄r }. Then to combine the effect
the selection probability of rules that have a relatively larger of all subpopulation, weighted average of the C̄i ’s is calculated
error but have matching conditions and correct actions, as r
i=1 C̄i mi
result delays discovery and reproduction of correct classifiers. C̃ =  r . (7)
Accordingly, a trade off between small confidence estimation i=1 mi

error and faster convergence is desirable. One can start training Thus C̃ is the confidence level vector of the model for each
the model in a more relaxed condition with smaller α value label. For a class that is suggested at least by one rule, the
to let the algorithm discover useful rules, and continue the respective C̃ value is nonzero, and for the the rest of the
training by increasing the α value to decrease the confidence classes, C̃ contains zeros. Procedure represented in (2) shows
estimation error even more. the flow of the learning algorithm during test.

277
Algorithm 1 MLCS-training
1: Initialize parameters;
2: while Maximum iteration not reached do
3: for i = 1:Number of training instances do
4: Samplei = (xi , λi , Ci );
5: Create [M ];
6: if [M ] = ∅ then
7: Do covering;
8: end if
9: for j = 1:|[M ]| do
10: Calculate bids (Bj ) using (2);
11: end for Fig. 1. Locus of label-confidence value pairs in two-dimensional space per
12: Select the maximum bid; class.
13: Create [A] and determine if the winning action is
correct; estimation error as well. Assume that label prediction and
14: if If winning action is correct then confidence estimation, constitute a two dimensional space. The
15: for j = 1:|[A]| do locus of the label vector λ (and its prediction λ̂) and confidence
16: Calculate reward share (Rj ) using (5); vector C (and its estimation Ĉ) is represented by the solid line
17: Update error (εj ) using (3); in figure (1) for one class. Considering class i of a k class data
18: Update confidence estimate (Ĉj ) using (4); set, if the label value is shown as ‘t‘ and its prediction shown
19: Update classifier strength using (6); as ‘o‘, the classification error for this class can be calculated
20: end for as the Euclidean distance (di (o, t)) between ‘t‘ and ‘o‘, shown
21: end if by the dashed line. Finally, the collective error considering all
22: Deduct taxes from [P ] and [M ]; classes will be the sum of each individual distance.
23: Apply GA according to the defined frequency;
|D|
24: Update accuracy of the model; 1 1
LEH ((λ × C ), H(x)) = dEH ((λi , Ci ), Hi (x))
25: end for |D| i=1 k
26: end while
(9)

Algorithm 2 MLCS-test In which, dEH is calculated as

1: for i = 1:Number of test instances do 
k
2: Samplei = (xi , λi , Ci ); dEH = di (o, t). (10)
3: Create [M ]; i=1
4: Extract the union of all label sets as predicted labels;
5: Extract number of rules advocating each label set (M); To have a better understanding, one can consider an instance
6: Calculate C̄ for each label set; from a three class data set such that λ = [0, 0, 1] with C =
7: Calculate weighted average of C̄ using (7) [0, 0, 1], while model has predicted λ̂ = [1, 0, 1] with Ĉ =
8: end for
[0.4, 0, 0.95]. Then for the first class, error is the collective
outcome of misclassification and missed prediction value, as
shown in figure (1) by the dotted line. For the third class, the
IV. P ERFORMANCE MEASURE error is only because of error in confidence estimation, that
lies on the vertical line corresponding to λ = 1, and collective
The performance of classification in MLC is mostly reported error is the average of these two distances.
in terms of Hamming loss [13], which is defined as the number
(or fraction) of labels whose relevance is incorrectly predicted. V. S IMULATION RESULTS AND DISCUSSION
Using the notation introduced in section II, the Hamming loss To examine the performance of the proposed learning al-
LH (·) for a hypothesis is defined as, gorithm, three experiments are conducted using leave-one-out
|D| (LOO) cross-validation technique which is an almost unbiased
1 1
LH (λ, H(x)) = dH (λi , Hi (x)) (8) estimator of the true error rate of a classifier [25]. Results are
|D| i=1 k reported in terms of Hamming loss and extended Hamming
In which k is the total number of classes in D. loss for multi-label data sets, and the standard accuracy for
In this work, to assess the accuracy of the model, a novel a multi-class data set. To show how well classifier population
measure is proposed that takes into account both classification learns the confidence levels, the average confidence estimation
accuracy and confidence estimation accuracy, that can be as- error (ε̄) of an entire population is also reported for each
sumed as an extension to the hamming loss. More specifically, experiment. Parameter specifications used in the proposed
hamming distance is extended to take into account confidence method are listed in Table (I).

278
 TABLE I where subscript index bits from left to right and primes denote
L IST OF PARAMETERS USED TO TRAIN THE MLCS MODEL .
negation. To show the proposed algorithm’s ability to learn
Population size limit 150 multi-label data, the class definition of the multiplexer problem
α 3 is modified. In the modified definition, the address bits of a
β 0.2 string are concatenated with the label specified by F6 and form
S0 = R0 100 the label for that string. The new labeling scheme, generally
P# 0.6 can contain more than one non-zero bits, therefor represents a
Cbid 0.1 multi-label sample. In the first test it is assumed that all labels
are confidently assigned, i.e. C = 1 for each class. Then the
Pcrossover 0.8
experiment is repeated for the same data set when confidence
Pmutation 0.04
value for class one in equal to 0.5 for all samples. The results
are compared to support vector machine (SVM) and decision
tree (DT) algorithms as reported in Table (III).
1) Experiment 1: The first experiment is on a multi-class
real-world data set. For this purpose, iris data set [26] is se-
lected which has three features, four classes and 150 samples. TABLE III
LOO CROSS - VALIDATION RESULTS FOR F6 PROBLEM .
For training the model, labels are coded into BR vectors and
since the data set is multi-class, each vector has only one HL EHL ε̄ Label confidence
non-zero element. For the sake of simplicity, it is assumed MLCS 0 0.0565 0.028 C=1
MLCS 0 0.0165 0.012 C(Class 1) = 0.5
that confidence values are all one, i.e. all data is confidently SVM 0.077 - - -
labeled, but it is also possible to assign lower confidences to DT 0 - - -
labels. Results are compared to the decision tree classifier and
also to the results reported in [27] for fuzzy rule mining using According to the results, the proposed algorithm is perform-
GA and multi-rule-table method [28] as shown in Table (II). ing as good as decision tree with zero HL, which means that
the class labels are learned completely, while support vector
TABLE II classifier has some misclassification and does not perform as
LOO CROSS - VALIDATION RESULTS FOR IRIS DATA SET.
well as MLCS and decision tree. EHL is nonzero due to the
Rate % ε̄
Average number of small confidence estimation error of individual rules. When the
rules data is labeled with partial confidence, i.e. confidence values
MLCS 97.57 0.0073 91.33
Fuzzy+GA 96 - 12.13 are smaller than one, ε reaches to zero in few number of
Multi-rule 94.67 - 691.11 iterations. This can be inferred from the two simulation results
DT 100 - - with complete confidence and partial confidence for class one
(Ci (1) = 0.5). Note that when samples in training data have
Results show that the proposed method has a slightly better different label confidence values, the solution search space
performance in learning a multi-class problem compared to becomes larger and the number of rules that are generated
other reported rule-based methods, although the number of is larger than simple case of equal confidence. A larger
generated rules is larger than the fuzzy rule-based classifier population size, will affect the convergence of the solution and
with random search. Note that, in this experiment reducing may decrease the computational efficiency of the algorithm.
the final number of rules was not our objective, though 3) Experiment 3: In the third experiment, a real-valued
it is possible to employ appropriate rule condensation [16] multi-label data set is tested and for this purpose ”real 6-
techniques to reduce the number of rules in the trained multiplexer” (RF6) [22] benchmark problem is selected. In
model. Decision tree algorithm outperforms all other methods this problem for each feature of the vector x = [x1 , . . . , x6 ],
which is not unexpected due to its strong induction capability. if xi < θ, the feature is interpreted as 0 and 1 otherwise.
Moreover, according to the value of ε̄, MLCS has learned the Therefore, the value of RF6 is the values of F6 applied to the
confidence values of labels with a small average error over bit string. It is assumed that features xi are scaled to [0, 1].
the population of rules. This means rules that are generated In order to obtain a multi-label data set, the label extension
with zero confidence, have gradually learned the objective defined in experiment (II) is employed here. A randomly
confidence levels of the data with small error. created data set with 150 samples and θ = 0.5 is provided
2) Experiment 2: In the second experiment, a Boolean to the algorithm for training and results are compare to SVM
multi-label data is used that is an extension of the binary and DT as shown in Table (IV).
multiplexer benchmark problem. For this purpose, the Boolean Results show that the proposed method performs better than
6-multiplexer is selected that is a binary-class problem taking a SVM with smaller HL value, in addition to the fact that when
six-bit string as input. The two left-most bits can be considered using SVMs to solve multi-label data a separate model is
as the address to a position in the remaining four bits. In trained for each class, while MLCS trains only a single model.
disjunctive normal form, the function that specifies the class DT has a better performance with zero hamming loss. Also,
label is given by F6 = b́0 b́1 b2 + b́0 b1 b3 + b0 b́1 b4 + b0 b1 b5 , the algorithm is performing well in learning the confidence

279
 TABLE IV [7] R. J. Urbanowicz and J. H. Moore, “Learning classifier systems: a com-
LOO CROSS - VALIDATION RESULTS FOR RF6 PROBLEM . plete introduction, review, and roadmap,” Journal of Artificial Evolution
and Applications, vol. 2009, p. 1, 2009.
HL EHL ε̄ [8] L. P. Kaelbling, M. L. Littman, and A. W. Moore, “Reinforcement
MLCS 0.011 0.0205 0.0004 learning: A survey,” Journal of artificial intelligence research, vol. 4,
SVM 0.08 - - pp. 237–285, 1996.
DT 0 - - [9] S. Godbole and S. Sarawagi, “Discriminative methods for multi-labeled
classification,” in Pacific-Asia Conference on Knowledge Discovery and
Data Mining. Springer, 2004, pp. 22–30.
[10] G. Tsoumakas and I. Katakis, “Multi-label classification: An overview,”
values and the estimation error is very small for the entire International Journal of Data Warehousing and Mining, vol. 3, no. 3,
population. 2006.
[11] G. Tsoumakas and I. Vlahavas, “Random k-labelsets: An ensemble
VI. C ONCLUSION method for multilabel classification,” in European Conference on Ma-
chine Learning. Springer, 2007, pp. 406–417.
In this paper action space representation of rules in strength- [12] M.-L. Zhang and Z.-H. Zhou, “A k-nearest neighbor based algorithm
based learning classifier systems is modified to handle multi- for multi-label classification,” in Granular Computing, 2005 IEEE
label classification tasks. Moreover, it is assumed that sample- International Conference on, vol. 2. IEEE, 2005, pp. 718–721.
[13] W. Cheng, E. Hüllermeier, and K. J. Dembczynski, “Bayes optimal mul-
label association is partially known and confidence values are tilabel classification via probabilistic classifier chains,” in Proceedings
used to represent this membership. During training, classifier of the 27th international conference on machine learning (ICML-10),
learns the degree of confidence and during the test, generates 2010, pp. 279–286.
[14] R. E. Schapire and Y. Singer, “Improved boosting algorithms using
a predicted label set and an average confidence level for each confidence-rated predictions,” in Proceedings of the eleventh annual
label. To measure the performance of the model, a novel loss conference on Computational learning theory. ACM, 1998, pp. 80–91.
function is proposed that is an extension to the well-known [15] R. M. Vallim, D. E. Goldberg, X. Llorà, T. S. Duque, and A. C.
Carvalho, “A new approach for multi-label classification based on default
Hamming loss measure and considers classification error and hierarchies and organizational learning,” in Proceedings of the 10th
confidence estimation error at the same time. annual conference companion on Genetic and evolutionary computation.
Simulation results show that for the iris data set, which is a ACM, 2008, pp. 2017–2022.
[16] S. Wilson, “Classifier fitness based on accuracy,” Evolutionary Compu-
multi-class data set, the accuracy of the model is slightly better tation, 1995.
than the other rule-based classification methods and label [17] E. Bernadó-Mansilla and J. M. Garrell-Guiu, “Accuracy-based learning
confidences are learned with a small error. For the multi-label classifier systems: models, analysis and applications to classification
tasks,” Evolutionary computation, vol. 11, no. 3, pp. 209–238, 2003.
Boolean and real-valued benchmark data sets, the Hamming [18] A. Orriols-Puig and E. Bernadó-Mansilla, “A further look at ucs classi-
loss performance is better than support vector classifiers and fier system,” GECCO06, pp. 8–12, 2006.
confidence levels are learned with small estimation error. [19] R. Urbanowicz, N. Ramanand, and J. Moore, “Continuous endpoint
data mining with exstracs: A supervised learning classifier system,”
In the future, this work will be extended to more efficient in Proceedings of the Companion Publication of the 2015 Annual
learning classifier algorithms such as U CS, to achieve a better Conference on Genetic and Evolutionary Computation. ACM, 2015,
performance and solve data sets with a larger number of pp. 1029–1036.
[20] R. Urbanowicz and J. Moore, “Retooling fitness for noisy problems in a
samples and features. Moreover, the results of the proposed supervised michigan-style learning classifier system,” in Proceedings of
algorithm in learning partial knowledge in sample-label asso- the 2015 Annual Conference on Genetic and Evolutionary Computation.
ciation, will be compared to other approaches that are able to ACM, 2015, pp. 591–598.
[21] M. V. Butz, T. Kovacs, P. L. Lanzi, and S. W. Wilson, “Toward a theory
handle label uncertainty. of generalization and learning in xcs,” IEEE transactions on evolutionary
computation, vol. 8, no. 1, pp. 28–46, 2004.
ACKNOWLEDGMENT [22] S. W. Wilson, “Get real! xcs with continuous-valued inputs,” in Learning
This work is supported by Air Force Research Laboratory Classifier Systems. Springer, 2000, pp. 209–219.
[23] G. Venturini, “Adaptation in dynamic environments through a minimal
and OSD under agreement number FA8750-15-2-0116. probability of exploration,” in Proceedings of the third international
conference on Simulation of adaptive behavior: from animals to animats
R EFERENCES 3: from animals to animats 3. MIT Press, 1994, pp. 371–379.
[1] P. Vannoorenberghe and T. Denoeux, “Handling uncertain labels in [24] A. Workineh and A. Homaifar, “Fitness proportionate niching: Main-
multiclass problems using belief decision trees,” in Proceedings of taining diversity in a rugged fitness landscape,” in Proceedings of the
IPMU, vol. 3, 2002, pp. 1919–1926. International Conference on Genetic and Evolutionary Methods (GEM).
[2] T. Fenoeux and M. S. Bjanger, “Induction of decision trees from partially The Steering Committee of The World Congress in Computer Science,
classified data using belief functions,” in Systems, Man, and Cybernetics, Computer Engineering and Applied Computing (WorldComp), 2012,
2000 IEEE International Conference on, vol. 4. IEEE, 2000, pp. 2923– p. 1.
2928. [25] S. M. Weiss and C. A. Kulikowski, Computer systems that learn: clas-
[3] Z. Elouedi, K. Mellouli, and P. Smets, “Belief decision trees: theoretical sification and prediction methods from statistics, neural nets, machine
foundations,” International Journal of Approximate Reasoning, vol. 28, learning, and expert systems. Morgan Kaufmann Publishers Inc., 1991.
no. 2-3, pp. 91–124, 2001. [26] R. A. Fisher, “The use of multiple measurements in taxonomic prob-
[4] E. Côme, L. Oukhellou, T. Denoeux, and P. Aknin, “Learning from lems,” Annals of human genetics, vol. 7, no. 2, pp. 179–188, 1936.
partially supervised data using mixture models and belief functions,” [27] Y.-C. Hu, R.-S. Chen, and G.-H. Tzeng, “Finding fuzzy classification
Pattern recognition, vol. 42, no. 3, pp. 334–348, 2009. rules using data mining techniques,” Pattern Recognition Letters, vol. 24,
[5] T. Denoeux, “Maximum likelihood estimation from uncertain data in the no. 1, pp. 509–519, 2003.
belief function framework,” IEEE Transactions on knowledge and data [28] H. Ishibuchi, K. Nozaki, N. Yamamoto, and H. Tanaka, “Selecting fuzzy
engineering, vol. 25, no. 1, pp. 119–130, 2013. if-then rules for classification problems using genetic algorithms,” IEEE
[6] L. Bull, “A brief history of learning classifier systems: from cs-1 to xcs Transactions on fuzzy systems, vol. 3, no. 3, pp. 260–270, 1995.
and its variants,” Evolutionary Intelligence, vol. 8, no. 2-3, pp. 55–70,
2015.

280

View publication stats