Efficiently finding the parameter for emergign pattern based classifier

Efficiently Finding the Best Parameter
for the Emerging Pattern-Based Classifier PCL

Thanh-Son Ngo1 , Mengling Feng2 Guimei Liu1 , and Limsoon Wong1
1
National University of Singapore
{ngothanh,liugm,wongls}@comp.nus.edu.sg
2
Institute for Infocomm Research
mfeng@i2r.a-star.edu.sg

Abstract. Emerging patterns are itemsets whose frequencies change
sharply from one class to the other. PCL is an example of efficient clas-
sification algorithms that leverage the prediction power of emerging pat-
terns. It first selects the top-K emerging patterns of each class that match
a testing instance, and then uses these selected patterns to decide the
class label of the testing instance. We study the impact of the parameter
K on the accuracy of PCL. We have observed that in many cases, the
value of K is critical to the performance of PCL. This motivates us to de-
velop an algorithm to find the best value of K for PCL. Our results show
that finding the best K can improve the accuracy of PCL greatly, and
employing incremental frequent itemset maintenance techniques reduces
the running time of our algorithm significantly.

1 Introduction

Classification is the task of learning from one data set and making predictions
in another data set. The learning data is called training data which consists of
entities with their labels. The other data set is called testing data which consists
of entities without labels, and the classifying task is to make prediction of their
labels based on what we learn from the training data. Many classifiers have been
proposed in the literature. Here, we focus on pattern-based classifiers.
Frequent patterns are patterns that appear frequently in the dataset. Fre-
quent patterns whose supports in the training data set change significantly from
one class to the other are used to construct the classifiers. These patterns are
called emerging patterns (EP) [9,10,16] and these patterns are applied to test-
ing instances to predict their class memberships. As a classifier may generate
many rules for the same class, aggregating their discrimination power gives bet-
ter prediction results [4,8,10,20]. Using EPs has the advantage that they not
only predict the class labels but also provide explanation for the decision.
Jumping emerging patterns (JEP) are a special type of EPs that have non-zero
occurrence in one class and zero occurrence in all other classes [10,11]. PCL [9,10]
is a classifier based on aggregating the prediction power of frequent JEPs. Given
Supported by A*STAR grant (SERC 072 101 0016).

M.J. Zaki et al. (Eds.): PAKDD 2010, Part I, LNAI 6118, pp. 121–133, 2010.
c Springer-Verlag Berlin Heidelberg 2010

122 T.-S. Ngo et al.

a test instance t, PCL selects the top-K JEPs with the highest supports that
are contained in t from each class and then computes the score for each class
using these selected patterns. PCL assigns the class label with the highest score
to the testing instance. The choice of a good value for K is tricky and its optimal
value varies on different datasets, as shown in Fig. 1. Remarkably, the problem
of choosing the best value of K has not been investigated previously.

Fig. 1. Accuracy of PCL on four datasets with different values of K. X-axis shows
different values of K and Y-axis shows the accuracy. When K increases, the accuracy
does not always increase.

Here, to avoid overfitting K to the data, we sample many subsets of the train-
ing data to get the value of K that appears the best on average. By this method,
we maximize the likelihood that the chosen K will produce the best results in the
whole dataset. Our main contributions are summarized as follows: (i) We revisit
the PCL algorithm [9,10] and propose a method to find the most appropriate
parameters to improve the performance of PCL. (ii) We introduce a method
to speed up the proposed algorithm as well as cross-validation methodology for
frequent-pattern-based classification algorithms in general.

2 Related Work

Emerging patterns are a special type of frequent patterns. They occur frequently
in one class but rarely in other classes. Emerging pattern-based classification ben-
efits enormously from the advancement of frequent pattern mining algorithms
and better understanding of the patterns space. Many methods have been pro-
posed to select a good set of patterns for constructing classifiers [15]. Most of
these algorithms first generate frequent patterns satisfying a certain minimum
support constraint, and then select patterns based on some criteria and make

Finding Best Parameter for PCL 123

predictions based on the selected patterns. These algorithms mainly differ in
how patterns are selected and how class labels are determined.
The first class of methods pick the best pattern to classify testing instances,
like CBA [14] and MMAC [18]. CBA first derives classification rules from fre-
quent patterns, and then ranks the rules in descending order of their confidence.
A very small set of rules is then selected to maximize the classification accuracy
on the training data. The class label of a new instance is decided by the first rule
that matches the instance. MMAC extends CBA to multiple classes and labels.
Only a small set of patterns are selected for classification. So it is possible that
a new testing instance does not contain any EP that is selected.
The second class of methods treats frequent patterns as additional features
and then use classical algorithms to classify the data [2,3]. Cheng et al. [2,3]
build a connection between pattern frequency and discriminative measures such
as information gain score, and develop a strategy to set minimum support in
frequent pattern mining for generating useful patterns for classification. Based
on this strategy, coupled with a proposed feature selection algorithm, selected
EPs are used as additional features to build high quality classifiers. The results
show that using EPs as additional features can improve the accuracy of C4.5
and SVM. The main drawback of this approach is that the intuitiveness of the
pattern-based classifiers may be lost.
The last class of methods selects the top-K patterns and aggregates the pre-
dictive power of the selected patterns. They include CAEP [4], iCAEP [20],
PCL [9,10], CEP [1], CPAR [17], CMAR [8] and HARMONY [19]. CAEP uses
EPs with high growth rate to do classification. iCAEP aggregates the prediction
power of EPs based on information theory. PCL uses only JEPs. JEPs occur in
one and only one class, which may be too restrictive in some cases. CEP relaxes
this constraint by putting an upper bound on the number of occurrences of the
EPs in other classes. CPAR uses the expected accuracy to select classification
rules. It compares the average expected accuracy of the best K rules of each class
and chooses the class with the highest expected accuracy as the predicted class.
CMAR uses a weighted measure to select rules, and the score of a class is also
calculated using this weighted measure. HARMONY directly mines the final set
of classification rules by using an instance-centric rule generation approach.
The aggregation-based algorithms generally show better performance than
other algorithms. However, the value of K is critical to their performance in many
cases. Here, we use PCL as an example to study how to select proper values for
parameter K to maximize the accuracy of the aggregation-based algorithms.

3 Preliminaries

Let I = {i1 , i2 , .., in } be a set of distinct literals called items. An itemset or
a pattern is a subset of I. A transaction is a non-empty set of items. Let
C = {C1 , C2 , .., Ck } be a set of distinct labels called class labels. A transac-
tional database consists of a set of transactions associated with their labels. A
classification task involves two phases: training and testing. In training, the class


labels are revealed and in testing, class labels are hidden. The classifier builds a
model based on the training set and uses this model to predict the class labels
of transactions in the testing set. The accuracy of a classifier A is the proportion
of test-instances which are correctly classified by A.
A pattern P covers a transaction t if P ⊆ t. The support of P is the number
of transactions that are covered by P . We use sup(P, D) to indicate the support
of P in the dataset D. We use P |D to indicate the set of transactions in D
that contain P . Given a 2-class dataset, jumping emerging patterns (JEP) are
patterns whose frequency in one class is non-zero and in other class is zero. An
EPi is called a JEP from class A to class B if its support in class A is zero. A
pattern P is a generator if and only if for every P ⊂ P , sup(P , D) > sup(P, D).
A JEP generator is both a generator and a JEP.

4 Algorithms

4.1 PCL

We present here an overview of PCL [9,10]. The dataset D is divided into
positive and negative classes. Given a test-instance t, two sets of JEPs that
cover t are used: EPt1 , EPt2 , .., EPtn from negative to positive classes and
+ + +
− − −
EPt1 , EPt2 , .., EPtn from positive to negative classes. The K most frequent
JEPs that cover t are sorted in descending order of their supports. Suppose
the set of JEPs (based on the training set) from negative to positive classes
are EP1 , EP2 , .., EPn in descending order of their supports. Similarly,
+ + +
− − −
EP1 , EP2 , .., EPn is the set of JEPs from positive to negative. It is hypoth-
esized that if t belongs to one class, it should contain more JEPs of this class
than the other class. So Li and Wong [9] formulated the scores of t with respect
to the two classes as below.
k k −
i=1 sup(EPti , D)
+
i=1 sup(EPti , D)
Score(t, +) = k
Score(t, −) = k
i=1 sup(EPi+ , D) i=1 sup(EPi− , D)

4.2 PSM

Maintenance of EPs was first introduced in [11], though a complete solution
for insertion and deletion was not discussed. Here, we describe a more efficient
method called PSM for complete maintenance introduced recently in [5].
PSM is a maintenance algorithm for frequent pattern space. It is based on the
GE-tree, an effective data structure described in [6,7] for enumerating genera-
tors. Frequent generators are stored in the GE-tree and the tree is incrementally
adjusted when new transactions are added or existing transactions are removed.
Each node in the tree represents a generator. To find EPs, we modify the GE-tree
so that each node stores both positive and negative supports of the correspond-
ing generator. In addition to frequent generators, GE-tree maintains a negative
border which comprises infrequent generators whose immediate subsets are all


frequent [5,6]. The negative border helps generate new frequent generators and
equivalence classes efficiently when transactions are added or removed.
Computating frequent generators is expensive. The benefit of PSM is that a
new set of frequent generators does not need to be computed from scratch when
the data is modified. As a small change in dataset seldom causes a big change
in the set of frequent patterns, PSM is effective for pattern space maintenance.

4.3 rPCL and ePCL
A good choice of K has a big impact on prediction results in PCL. We propose a
method to tackle this problem as follows. According to the Central Limit The-
orem, the distribution of accuracy will behave like normal distribution. Indeed,
Fig. 2 suggests the convergence of average classification accuracy in training data
to the real value of accuracy in the whole dataset. So we simulate the actual pro-
cess of classification in training set and choose the value of K that maximizes
the mean accuracy. The simulation is run repeatedly to determine which value
of K appears as the best on average. By the Central Limit Theorem, the average
accuracy of each K approaches the true accuracy of that value of K, given suf-
ficient number of simulation runs. Thus, the value of K that has the best mean
accuracy in the simulation runs will also perform well in the whole set of data.

Fig. 2. Distribution of accuracies across 100 runs given a fixed K=10 over training set.
The vertical red line indicates the actual value of accuracy when we evaluate the same
K in the whole dataset. The means are shown to be close to the actual accuracy.

We describe two algorithms rPCL and ePCL. Algorithm rPCL is a direct but
naive solution for the method above. We use it as a benchmark to understand
where inefficiencies might be and to assess the improvement of a more efficient
method using maintenance. ePCL is the fast version where PSM is used. It
produces the same results as rPCL but runs many times faster.


Fig. 3. The rPCL algorithm

In rPCL (Fig. 3) we use a technique called repeated random sub-sampling
validation as a framework to assess the best parameter for PCL. It involves
several rounds, in each round the training set is randomly divided into new
complementary testing and training set. For each K, we perform this procedure
to determine the expected accuracy of the classifier wrt this K. The chosen K is
the one which returns the best average accuracy over these runs. 10-fold cross
validation is popular for accessing classifier performance. So, we subsample 10%
of training set as testing subsample and 90% as training subsample.
In rPCL, makeDecision simply decides if the classifier with ScoreK [t, +] and
ScoreK [t, −] correctly predicts the class of t, wrt to the current K. At step 9, we
can compute ScoreK [t, +] from Score( K − 1)[t, +] in constant time. To achieve
this, we use a vector to store ScoreK [t, +] and ScoreK [t, −] for all t ∈ Dt. To
determine a good value of K, a significantly large number of iterations maxtime
is performed and all values of K in the range 1..maxK are considered.
Constructing a set of frequent EPs from the training set (step 4) is compu-
tationally very expensive. It involves frequent pattern mining. And this step is
repeated many times with different Dt and Dn. Because the testing-training sep-
arations are largely similar among all the runs, we should not need to construct
the set of EPs from scratch in such a naive fashion. PSM [5,6] is used to achieve
this purpose. PSM allows us to adjust the set of frequent patterns when the
training fold is changed. We only need to construct the frequent patterns space
at the beginning and repeatedly use PSM to maintain the set of rules when the
sub-sampling folds are changed. That is, we only need to mine frequent patterns
once at the beginning. With this innovation, we present an improved algorithm
called ePCL (Fig. 4) which stands for enhanced PCL.


Fig. 4. ePCL algorithms. The highlighted codes indicate the involvement of PSM.

Fig. 5. Workﬂow for rPCL and ePCL. EP set indicates the set of EPs used to construct
PCL.


The set of frequent generators are incrementally maintained. Line 5 and 17
show the execution of PSM: PSM.delete is invoked when a fold is removed from
the training set and PSM.add is invoked to add this fold into original set. That
eliminates building a new set of frequent patterns when we run the simulation
with a new testing-training separation. Fig. 5 is the workflow of rPCL and ePCL.

4.4 Complexity Analysis
We compare the theoretical improvement of ePCL from rPCL. In rPCL, the
outer loop repeats maxtime, which is the number of runs needed to find the
best K. In each loop, we need to run PCL classifier which involves a frequent
pattern mining step (F P M ) and evaluation for all K from 1 to maxK. The
total time is maxtime ∗ (F P M + P CL(maxK)). P CL(maxK) is the time to
compute scores for K = 1..maxK. This step evolves visiting entire frequent pat-
tern set to get the top K patterns and compute scores accordingly for each test
instance. In ePCL, we only need to build frequent patterns once and incremen-
tally adjust the pattern space. The total time is F P M + maxtime ∗ (|D|/10 ∗
maintenance + P CL(maxK)), where maintenance is the running time for main-
taining one transaction, |D| is the size of the dataset. According to [5], PSM is
much more efficient than mining-from-scratch algorithms. At 10% of the data
size, the speed-up is at least 3 times faster for incremental and 5 times faster for
decremental maintenance than mining from scratch.

5 Experimental Studies
5.1 Experiment Setup
Experiments are conducted using 18 data sets of various sizes from UCI reposi-
tory. Continuous attributes are discretized by the entropy method. Table 1 gives
details of these data sets. We evaluate the performance on two-class datasets. For
multi-class datasets, we select one class as positive and the remaining as negative.
The process is done for all classes. PCL can be extended to handle multiple-class
datasets; however, the method of choosing parameter is still applicable. 10-fold
cross validation is used to assess the efficiency and average accuracies are re-
ported. The datasets are separated in such a way that the proportion of each
class is consistent in both testing and training.
For the original PCL, we use the frequent pattern mining algorithm provided
by [13]. The experiments are done in Windows machine with 2.6 GHz processor
and 1G memory. We assess the accuracy improvement of rPCL over PCL and
running time comparison of ePCL and rPCL. Time is measured in seconds.
When we compute the score of an instance in two classes, it is possible that
the scores are both zero. Such a test-instance is not covered by any pattern
and therefore unclassifiable by both PCL and ePCL. The average percentage of
unclassified data is 11% and there is no dataset with more than 20%. We do not
include these cases in our accuracy computation. The purpose is to evaluate the
improvement of rPCL over the original PCL only.


Table 1. Datasets information

Dataset # attributes # instances # continuous attributes # nomial attributes
mushroom 4 8124 4 0
iris 8 150 8 0
Glass 22 214 0 22
zoo 16 101 0 16
Promoter 18 106 0 18
Vote 10 435 10 0
Splice 19 3190 0 19
Hepatitis 59 155 0 59
Pima 61 768 0 61
Hypothyroid 25 3163 7 18
Breast 16 3190 0 16
Cleve 10 2727 0 10
German 12 2700 0 12
Lymph 19 1332 0 19
Vehicle 19 3809 0 19
Waveform 20 9000 0 20
Wine 14 178 0 14
Tic-tac-toe 10 4312 0 10

5.2 Parameters Setting

All the patterns used in PCL and its improved versions are JEP generators [10].
The exception is an early paper [8]; it uses the “boundary” JEPs, which are
defined in [9] as those JEPs where none of their proper subsets are JEPs. Here,
we use JEP generators as discriminative patterns because:

– It is known that the JEP space is partitioned into disjoint equivalence
classes [9,10]. So the set of JEP generators (which correspond to the most
general patterns in each equivalence class) adequately characterizes the data.
In contrast, boundary JEPs are insufficient to capture all equivalence classes.
– The minimum description length principle is a well-accepted general solution
for model selection problems. The choice of JEP generators is in accord with
this principle [12]. Also, our experiments show that JEP generators are less
noise-sensitive than boundary JEPs.
– JEP generators are efficient to compute. We can make use of many efficient
frequent patterns mining algorithms [12].

For rPCL and ePCL, we set maxtime to 50 and we run K in a range from 1
to 50. We limit the range to 50 because small-frequency patterns have minimal
prediction power over the top ones. Patterns ranked higher than 50 generally
have low support. As shown in Fig. 1, the predictions stabilize after K = 50,
so no need to consider K > 50. For original PCL, we use K = 10 as suggested
in [9]. Minimum support is set to make sure enough EPs are found.


5.3 Efficiency
PCL with generators and boundary EPs. For PCL, we have choices over
which type of frequent patterns are used to make predictions. We have done
experiments to justify our choice of JEP generators, as suggested in [9], rather
than boundary JEPs. We want to test the robustness of PCL in the case of noise
with these two types of patterns. When a part of the original dataset is missing,
the noisy dataset reduces the accuracy of the classifier. Fig. 6 shows accuracy
of PCL with generators (gen PCL) and boundary JEPs (boundary PCL) for
different levels of missing rate (We assume items in the dataset are missing with
certain rate): 0%, 10%, 20%, 30%, 40% and 50%. The accuracies are average
over 18 datasets. Gen PCL is less noise-sensitive than boundary PCL.

Fig. 6. Accuracy of gen PCL and boundary PCL in the presence of missing information

PCL and ePCL. We compare the accuracy of the original PCL and ePCL.
Since rPCL and ePCL give the same results, we do not show the accuracy of
rPCL here. Table 2 shows accuracy of ePCL and the original PCL in 18 datasets.
Overall, the improvement is 3.19%. In some datasets like Promoters, Wine and
Hepatitis, we get improvement of 26%, 11% and 15% respectively.

Table 2. Accuracy comparison of ePCL and PCL

Datasets ePCL PCL Improve (%) Datasets ePCL PCL Improve (%)
Mushroom 1 1 0 Iris 0.9999 0.9658 3.534591
Glass 0.9714 0.9675 0.404967 Zoo 0.9997 0.9528 4.925911
Promoter 0.8216 0.6525 25.92549 Vote 0.9439 0.9492 -0.55892
Splice 0.6667 0.6667 0 Hepatitis 0.8872 0.7716 14.98068
Pima 0.8192 0.8331 -1.67043 Hypothyroid 0.9962 0.9861 1.017527
Breast 0.9912 0.9907 0.050572 Cleve 0.9644 0.9959 -3.16567
German 0.9644 0.994 -2.97462 Lymph 1 1 0
Vehicle 0.9569 0.9437 1.390222 Waveform 0.8347 0.8053 3.643507
Wine 0.97892 0.8789 11.3941 Tic-tac-toe 0.9512 0.965 -1.42591


Table 3. Accuracy comparison of ePCL and PCL in difficult cases

Datasets Difficult PCL ePCL Datasets Difficult PCL ePCL
cases (%) cases (%)
mushroom 0 - - Iris 0 - -
Glass 33.94 0.9071 0.9183 zoo 26.53 0.8274 0.9989
Promoter 49.02 0.3123 0.647 Vote 16.31 0.6978 0.6663
Splice 100 0.6667 0.6667 Hepatitis 100 0.7716 0.8872
Pima 12.10 0 0 Hypothyroid 20 0.9327 0.9813
Breast 10.90 0.917 0.9214 Cleve ¡1 - -
German 10.60 0.9449 0.6745 Lymph 0 - -
Vehicle 3.7453 0 0 Waveform ¡2 - -
Wine 31 0.8788 0.9789 Tic-tac-toe 30.2 0.553 0.578

Recall in PCL, to classify one test instance, one score is computed for each
class. Some instances received zero score in one class and non-zero score in the
other class; thus the association rules vote unanimously to one class. We call these
the easy cases and the rest are difficult cases. The value of K has a much greater
impact on the difficult cases. Table 3 shows the accuracy of original PCL and
ePCL in difficult cases. Although we do not show statistics for easy cases here,
the accuracy is close to 100% for most easy cases. We do not show datasets with
too few difficult cases since there are not enough samples to evaluate accurately.

Performance. We compare the running time of rPCL, ePCL and original PCL.
The performance of ePCL compared to rPCL is good, demonstrating the supe-
riority of using pattern maintenance. Table 4 shows that ePCL is an order of
magnitude faster in many cases. In one case (vehicle.dat), rPCL could not com-
plete after more than one day, but ePCL completed within 4 hours. Nevertheless,
ePCL takes more time than PCL due to the repeating part; fortunately, this is
compensated by the much better accuracy of ePCL.

Table 4. Performance comparison

Datasets PCL rPCL ePCL Speed up Datasets PCL rPCL ePCL Speed up
(rPCL/ (rPCL/
ePCL) ePCL)
mushroom 2.75 84 35 2.4 Iris 2 99 3 33
glass 8.5 120 2 60 zoo 5 291 7 41.5
promoters 1 51 7 7.2 Vote 0.75 47 11 4.2
splice 2.5 129 4 32.2 hepatitis 0.5 38 3 12.6
pima 0.75 42 9 4.6 hypothyroid 2 105 43 2.4
breast 2.5 109 60 1.8 cleve 4.25 181 123 1.4
german 11 513 611 0.8 lymph 27.5 1224 512 2.3
vehicle 2518 too long 6966 A lot waveform 117 5738 2507 2.3
wine 3.25 188 17 11.1 tictactoe 13 88 43 2


6 Discussion and Conclusion

Cross-validation is a technique to assess the performance of classification algo-
rithms. Typically the testing set is much smaller than the training set and a
classifier is built from the training set which is largely similar between runs. Our
framework of maintenance can be applied to efficiently perform cross-validation
for pattern-based classification. Recall PSM allows us to maintain the set of
frequent generators while adding or deleting a relatively small portion of the
dataset. We only need to maintain a single frequent pattern set and incremen-
tally adjust using PSM. In addition, for more efficient computation, PSM can be
extended to handle multiple additions/deletions at the same time by organizing
transactions in a prefix tree.
We showed the importance of finding the top K patterns for PCL. We intro-
duced a method to perform the task efficiently, making use of the PSM main-
tenance algorithm [5]. The maintenance framework can be applied to general
cross-validation tasks which involve frequent update of the pattern space. Our
experiments showed improvement in accuracy. However, we compromised on
complexity. We hope to get a better implementation for ePCL, thus minimiz-
ing the running time to close to original PCL and implement the mentioned
extensions.

References

1. Bailey, J., Manoukian, T., Ramamohanarao, K.: Classification using constrained
emerging patterns. In: Dong, G., Tang, C., Wang, W. (eds.) WAIM 2003. LNCS,
vol. 2762, pp. 226–237. Springer, Heidelberg (2003)
2. Cheng, H., et al.: Discriminative frequent pattern analysis for effective classifica-
tion. In: Proc 23th ICDE, pp. 716–725 (2007)
3. Cheng, H., et al.: Direct discriminative pattern mining for effective classification.
In: Proc 24th ICDE, pp. 169–178 (2008)
4. Dong, G., et al.: CAEP: Classification by Aggregating Emerging Patterns. In: Proc.
2nd Intl. Conf. on Discovery Science, pp. 30–42 (1999)
5. Feng, M.: Frequent Pattern Maintenance: Theories and Algorithms. PhD thesis,
Nanyang Technological University (2009)
6. Feng, M., et al.: Negative generator border for effective pattern maintenance. In:
Proc 4th Intl. Conf. on Advanced Data Mining and Applications, pp. 217–228
(2008)
7. Feng, M., et al.: Evolution and maintenance of frequent pattern space when trans-
actions are removed. In: Zhou, Z.-H., Li, H., Yang, Q. (eds.) PAKDD 2007. LNCS
(LNAI), vol. 4426, pp. 489–497. Springer, Heidelberg (2007)
8. Li, W., Han, J., Pei, J.: CMAR: Accurate and efficient classification based on
multiple class-association rules. In: Proc 1st ICDM, pp. 369–376 (2001)
9. Li, J., Wong, L.: Solving the fragmentation problem of decision trees by discovering
boundary emerging patterns. In: Proc. 2nd ICDM, pp. 653–656 (2002)
10. Li, J., Wong, L.: Structural geography of the space of emerging patterns. Intelligent
Data Analysis 9(6), 567–588 (2005)


11. Li, J., Ramamohanarao, K., Dong, G.: The space of jumping emerging patterns
and its incremental maintenance algorithms. In: Proc. of 17th ICML, pp. 551–558
(2000)
12. Li, J., et al.: Minimum description length principle: Generators are preferable to
closed patterns. In: Proc. 21st Natl. Conf. on Artificial Intelligence, pp. 409–415
(2006)
13. Li, H., et al.: Relative risk and odds ratio: A data mining perspective. In: Proc
24th PODS, pp. 368–377 (2005)
14. Liu, B., Hsu, W., Ma, Y.: Integrating classification and association rule mining.
In: Proc. 4th KDD, pp. 80–86 (1998)
15. Ramamohanarao, K., Fan, H.: Patterns based classifiers. In: Proc. 16th WWW,
pp. 71–83 (2007)
16. Ramamohanarao, K., Bailey, J.: Discovery of emerging patterns and their use in
classification. In: Gedeon, T(T.) D., Fung, L.C.C. (eds.) AI 2003. LNCS (LNAI),
vol. 2903, pp. 1–12. Springer, Heidelberg (2003)
17. Thabtah, F.A., Cowling, P., Peng, Y.: MMAC: A new Multi-class Multi-label As-
sociative Classification approach. In: Proc. 4th ICDM, pp. 217–224 (2004)
18. Yin, X., Han, J.: CPAR: Classification based on Predictive Association Rules. In:
Proc. 3rd SDM, pp. 331–335 (2003)
19. Wang, J., Karypis, G.: HARMONY: Efficiently mining the best rules for classifi-
cation. In: Proc. 5th SDM, pp. 205–216 (2005)
20. Zhang, X., Dong, G., Ramamohanarao, K.: Information-based classification by
aggregating emerging patterns. In: Proc. 2nd Intl. Conf. on Intelligent Data Engi-
neering and Automated Learning, pp. 48–53 (2000)

Efficiently finding the parameter for emergign pattern based classifier

More Related Content

What's hot (16)

Viewers also liked (8)

Similar to Efficiently finding the parameter for emergign pattern based classifier (20)

Recently uploaded (20)

Efficiently finding the parameter for emergign pattern based classifier