See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/299986939

Discovering Useful Patterns from Multiple Instance Data

Article  in  Information Sciences · April 2016

DOI: 10.1016/j.ins.2016.04.007

CITATIONS READS
8 3,139

4 authors:

José María Luna Alberto Cano

University of Cordoba (Spain) Virginia Commonwealth University
42 PUBLICATIONS   814 CITATIONS    74 PUBLICATIONS   618 CITATIONS   

SEE PROFILE SEE PROFILE

Virgilijus Sakalauskas Sebastian Ventura

Vilnius University University of Cordoba (Spain)
66 PUBLICATIONS   144 CITATIONS    301 PUBLICATIONS   8,650 CITATIONS   

SEE PROFILE SEE PROFILE

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

artificial intelligence View project

Blind Source Separation View project

All content following this page was uploaded by Alberto Cano on 04 October 2017.

The user has requested enhancement of the downloaded file.

 Discovering Useful Patterns from Multiple Instance
Data
J. M. Lunaa , A. Canob , V. Sakalauskasc , S. Venturaa,d,∗
a
Dpt. of Computer Science and Numerical Analysis, University of Cordoba, Spain
b
Dpt. of Computer Science, Virginia Commonwealth University, Richmond, USA
c
Dpt. of Informatics. Kaunas Faculty of Humanities, Vilnius University, Lithuania
d
Dpt. of Information Systems, King Abdulaziz University, Jeddah, Saudi Arabia

Abstract
Association rule mining is one of the most common data mining techniques
used to identify and describe interesting relationships between patterns from
large datasets, the frequency of an association being defined as the num-
ber of transactions that it satisfies. In situations where each transaction
includes an undetermined number of instances (customers shopping habits
where each transaction represents a different customer having a varied num-
ber of instances), the problem cannot be described as a traditional association
rule mining problem. The aim of this work is to discover robust and use-
ful patterns from multiple instance datasets, that is, datasets where each
transaction may include an undetermined number of instances. We propose
a new problem formulation in the data mining framework: multiple-instance
association rule mining. The problem definition, an algorithm to tackle the
problem, the application fields, and the relations’ quality measures are for-
mally described. Experimental results reveal the scalability of the problem
on different data dimensionality. Finally, we apply it to two real-world ap-
plications field: (1) analysis of financial data gathered from one of the most
important banks in Lithuania; (2) study of existing relations between records
of unemployed gathered from the Spanish public employment service.
Keywords: Association Rules, Multiple-Instance Data, Data Mining

∗
Corresponding author
Email addresses: [email protected] (J. M. Luna), [email protected] (A. Cano),
[email protected] (V. Sakalauskas), [email protected] (S. Ventura)

Preprint submitted to Information Sciences February 24, 2016

1. Introduction
We are being involved in the digital revolution, which provides the storage
of more information on every single topic than ever before. This unrestrained
growth of data has led to a situation where the extraction, description and
analysis of hidden knowledge is a requirement to be able to take advantage of
this massively data storage. Association rule mining (ARM) [1], an unsuper-
vised learning task, is one of the most common data mining techniques [8, 9]
used to describe hidden knowledge in data [13]. ARM was proposed as a
way for seeking frequent, interesting and strong relations between products
in market basket analysis. In this regard, it aims to extract useful knowledge
about customers shopping habits to make right decisions.
To date, the market basket analysis has been carried out by considering
the whole set of transactions, extracting useful information about relations
between products like one of the most well known relationships in this field,
i.e., if diaper then beer. Nevertheless, it could easily happen that a specific
product could be bought only by a single customer, or even that this single
customer usually buys more often than everybody else. Thus, the relations
obtained describe a general scenario that provides information that is biased
against occasional customers. This issue plays an important role in decision
making since once the specific customer stops purchasing, the product will
no longer be sold.
In bank financial analysis, a similar analysis can be inferred, where each
customer might address a different and varied number of banking transac-
tions. In this field, a traditional ARM analysis provides relations between
banking transactions from a general scenario without distinction between fre-
quent and sporadic customers, or even without considering the time stamp in
which each transaction was carried out. The distinction between customers,
time stamps or any other specific attribute may provide an interesting and
quite different analysis with regard to the one inferred from traditional ARM.
This way of describing relations on different scenarios is somehow related
to multi-relational learning (MRL) [16] framework, where the scenarios to be
described are defined by primary keys. MRL refers to a context where the
examples may have a complex internal structure (multiple components and
there may be relationships between these components) and this structure
is fixed beforehand (primary keys cannot be modified). Nevertheless, the
only way of introducing ambiguity to MRL is by means of its primary key,
for example, a customer defined by a unique attribute (primary key) can

2
include many alternative instances. The ambiguity problem was introduced
by the multiple instance learning (MIL) [6]. In MIL, training transactions
are ambiguous and a single transaction may have many alternative instances
that describe it. The key challenge in MIL is to cope with the ambiguity,
receiving a growing attention in the machine learning community [5, 19].
The goal of this work is to introduce the ambiguity problem into the ARM
field in order to solve the problem of discovering relationships which may be
pointless in analysis like the previously described. Notice that the ambiguity
can be considered from an undefined number of perspectives, grouping the in-
stances by customers, time stamps, or any other different feature, depending
on the desired analysis. The proposed solution gives rise to the formulation
of a new task, association rule mining on multiple-instance data (MI-ARM),
which presents some similar points to MIL. This new task is able to extract
useful patterns, and relationships between that patterns, from datasets where
each transaction groups a set of instances. To demonstrate the strength and
usefulness of using the proposed new task, a sample dataset for the market
basket analysis has been considered. Different perspectives have been con-
sidered, grouping its records to form multiple instances according to different
goals. The knowledge extracted in this example does not describe a real sce-
nario but denotes the high utility of this type of association rules, discovering
knowledge that is hardly obtained with traditional ARM algorithms.
In this paper, a series of experiments have been carried out to justify and
demonstrate the usefulness of the proposed task, which differs from MRL [25]
in its ability to discern groups [26] by any specific feature instead of using a
unique attribute. Additionally, our proposed task enables to form group by
abstract features, which is impossible in MRL. It is not the intention to com-
pare the results for specific data but to provide an overview on the usefulness
of this new problem formulation. In this experimental analysis we also study
the scalability of this new problem considering a different number of both
transactions and attributes. Finally, two real-world application fields are
considered, making an analysis of interesting relations between patterns [4]
discovered from real-data.
This paper is arranged as follows. Section 2 presents the most relevant
definitions, related work and explains the novelty and contributions of our
work correspondingly. Section 3 formally describes the proposed new task.
Section 4 describes the proposed algorithm to be considered in this new
problem. Section 5 presents the datasets used in the experiments and the
results obtained. Finally, some concluding remarks are outlined in Section 6.

3
2. Preliminaries
In this section, we formally define both the association rule mining (ARM)
and multiple-instance learning (MIL) problems and then, we discuss the in-
tuition behind and justification for the new concept and highlight the novelty
of our work with regard to some related works.

2.1. Association Rule Mining

The extraction of association rules is considered as an important descrip-
tive task in the data mining field, and it has received enormous attention
since its introduction by Agrawal et al. [1] in the early 90s. Association
rules can be considered as a way to describe interesting and usually hidden
associations between items in data [22].

Definition 1 (Association rule). Let I = {i1 , i2 , ..., in } be the set of

items in a dataset, and a pattern P be formally defined as a subset of I, i.e.
{P = {ij , ..., ik } ⊆ I, 1 ≤ j, k ≤ n}, that describes valuable features of data.
An association rule, defined for a pattern P , is defined as implications of the
form IF Antecedent THEN Consequent, Antecedent (A) and Consequent
(C) being subsets of the pattern P , i.e. {A ⊂ P ⊆ I ∧ C = P \ A}, or also
{C ⊂ P ⊆ I ∧ A = P \ C}.

The meaning of an association rule is that if the antecedent is satisfied,

then it is highly probable that the consequent is also satisfied. Association
rules were originally designed for the market basket analysis as a way to
describe relationships between products like diapers → beer , denoting the
high probability of someone buying diapers also buying beer. It would allow
shop-keepers to exploit this relationship by locating the products in an ac-
curate place. Nowadays, more and more application domains [17, 11, 20] use
association rules to describe and analyse unknown relations in data, which
has given rise to the task of mining association rules.

Definition 2 (Association rule mining). Let T = {t1 , t2 , ..., tm } be the

set of transactions in a dataset, and each transaction ti ∈ T be a subset of
items such that tj ⊆ I. The association rule mining (ARM) task aims to find,
from the whole set of rules R, any association rule R that satisfies specific
user thresholds α of interest, i.e., ARM = {∀ R ∈ R : quality(R) ≥ α}.

4
 ARM suffers from different main problems that have been studied by
many researchers [7]. First, the discovery of patterns of interest in large
datasets can be computationally expensive [10], requiring a large amount of
memory [1, 18]. Second, in many application domains, some patterns simply
appear by chance [12] and they might be useless for the user. Third, the
mining of associations in continuous domains [14] is a hard task and many
algorithms, specially exhaustive search algorithms, require a preprocessing
step to discretize data.

2.2. Multiple-Instance Learning

The problem of multiple instance learning (MIL) was firstly introduced by
Dietterich et al. [6] in the context of drug activity prediction. This problem
arises in tasks where the training transactions are ambiguous, that is, a single
transaction may include many alternative instances that describe it.

Definition 3 (Multiple-instance learning). Let us assume that T be

the set of transactions in data, and each transaction ti ∈ T be composed of
a varied number k of instances, i.e. ti,1 , ti,2 , ..., ti,k . Each of these instances
is represented by a distinct feature vector V (ti,j ), j ≤ k, so a complete
transaction ti is therefore represented as {V (ti,1 ), ..., V (ti,k )}. The multiple-
instance learning is defined as the looking for a good approximation fˆ to f
by analysing a set of transactions represented as {V (ti,1 ), V (ti,2 ), ..., V (ti,k )}.

According to Dietterich et al. [6], the result of f is positive for a trans-

action ti if at least one of its instances ti,k has produced a positive result,
otherwise, the result of f is negative. Zucker et al. [24] determined that to
classify a new transaction in the multiple-instance problem by means of a
decision tree, the set of instances of the transaction is passed through the
tree. As soon as one positive leaf is reached by one of the instances, the
transaction is classified as positive, negative otherwise. For other authors,
instances are not necessarily alternative descriptions of a transaction but
descriptions of different parts of the transaction. Thus, two decision trees
(positive and negative) have to be learned. To classify a new transaction, its
set of instances is passed through the two trees.

2.3. Contribution and Related Work

The ultimate goal of ARM is the looking for frequent, interesting and
strong relations between patterns in a database. In this field, a specific

5
pattern P satisfies a transaction tj if and only if P ⊆ tj , and the frequency
of this pattern f (P ) is defined as the number of different transactions that it
satisfies, i.e. f (P ) = |{∀tj ∈ T : P ⊆ tj }|. A pattern P is defined as frequent
if and only if the number of transactions that it satisfies is greater or equal
to a minimum predefined threshold fmin , i.e. f (P ) ≥ fmin .
With the increasing interest in the data storage, the extraction of valu-
able and powerful knowledge is essential. Let us consider a supermarket
chain that wants to extract useful knowledge about their customers’ shop-
ping habits to make right decisions. It is interesting to obtain this knowl-
edge by analysing each customer regardless the number of transactions that
he represents. Thus, data is structured into transactions and each of these
transactions (one per customer) comprises an undetermined number of in-
stances (at least one per transaction). A specific pattern can be considered of
high interest depending on the number of customers satisfied, so the problem
cannot be considered as in classic ARM.
As mentioned in the previous section, this way of storing information
is related to multiple instance learning (MIL) [6]. Since many real-world
problems can be represented as multi-instance problems, MIL has received a
great attention in the machine learning community. MIL has been applied
successfully to different domains such as text categorization [2], drug activity
prediction [15], web index page recommendation [23] and predicting student
performance [21].
The main contribution of this paper is the use of rules to describe be-
haviours in multiple instance data. We formulate the multiple-instance asso-
ciation rule mining (MI-ARM) problem, which presents some similar points
to MIL. This new task can be considered, to some extent, as a simplification
of the multi-relational learning (MRL) [16] problem where only a single per-
spective is considered, that is, only one relation (transactions grouped by a
unique attribute) from the multi-relational database. Nevertheless, the com-
parison between MI-ARM and MRL is pointless since the former includes
a set of transactions comprising an undetermined number of instances, and
these transactions can be obtained by grouping instances by any feature on
a single scenario. On the contrary, MRL refers to a context where the trans-
actions may have a complex internal structure, which is fixed beforehand.
Besides, MRL does not allow to group transactions by abstract features as
in MI-ARM.

6
3. Problem Formulation: Multiple-Instance Association Rules
A transactional database comprises a set T = {t1 , t2 , ..., tn } of transac-
tions where each transaction ti denotes a specific record in the database.
Traditional algorithms for mining association rules [1] work on the transac-
tional database, discovering previously unknown patterns of high interest for
a predefined aim. Nevertheless, the increasing interest in data storage has
given rise to the extraction of valuable and powerful knowledge. Sometimes,
a transactional database can be considered from different perspectives, where
each specific transaction provides much richer information since it comprises
multiple records or instances, that is, each transaction ti ∈ T comprising
a varied number k of instances, i.e. ti,1 , ti,2 , ..., ti,k . Figure 1 abstractly
illustrates how a transactional database that comprises a set of single trans-
actions can be transformed into a varied set of multiple instance databases.
Each of these multiple instance databases can be considered as a perspective,
since each of its records can be grouped to for a multiple instance.
Traditional association rule mining cannot be applied to extract strong
relationships from multiple-instance databases. Hence, this task is only use-
ful for discovering rules in a single perspective, that is, the general scenario in

Transactional
database

Perspective1 Perspectiven
Perspective2

Multiple instance ...

database

Figure 1: Multiple-instance databases can be obtained from a transactional database by

considering different perspectives

7
which each transaction represents a single record. However, in many applica-
tion domains, the ability to represents the data into different perspectives is
essential, since the knowledge extracted is much richer and of higher interest
for the user.
Based on the both previously introduced concepts, we define a new prob-
lem that is formulated halfway between ARM and MIL. This new concept,
known as multiple-instance association rule mining (MI-ARM), defines the
discovery of interesting and strong relations between patterns from multiple-
instance data. MI-ARM is presented as an unsupervised learning task where
items are analysed from the perspective of transactions including an unde-
termined number of instances.
Definition 4 (Multiple-instance association rule). Let T be the set
of transactions in data, and each transaction ti ∈ T comprising a varied
number k of instances, i.e. ti,1 , ti,2 , ..., ti,k . Additionally, each single instance
ti,j be a subset of items such that ti,j ∈ ti : j ≤ k, ti,j ⊆ I, from the total
set of items I = {i1 , i2 , ..., in } in the dataset. A multiple-instance association
rule (MI-AR) is an implication of the type A → C, A and C being defined
from a pattern P = {A ∪ C}, i.e. {A ⊂ P ⊆ I ∧ C = P \ A}, or also
{C ⊂ P ⊆ I ∧ A = P \ C}.

According to the previous definition (see Definition 4), the frequency of

occurrence of a MI-AR obtained from a pattern P is defined as the number
of transactions satisfied by P . A transaction ti is satisfied if and only if at
least one instance ti,j is satisfied by P .
Definition 5 (Multiple-instance association rule mining). This task
is defined as an unsupervised learning problem. The aim of mining multiple-
instance association rules (MI-ARM) is to find, from a multiple instance
dataset, any MI-AR R that satisfies specific user thresholds α of interest,
i.e.MI-ARM = {∀ R : quality(R) ≥ α}.

MI-ARM (see Definition 5) considers a set of quality measures to deter-

mine the interest of the rules mined. The main difference between ARM and
MI-ARM falls upon the concept of transactions and instances. The support
quality measure (Equation 1) of a rule A → C obtained from a pattern P ,
i.e. P = {A ∪ C}, is defined as the number of transactions satisfied by P .
A transaction ti is satisfied if and only if at least one instance ti,j is satisfied
by P .

8
 |{∀ti ∈ T : ∃ti,j ∈ ti ∧ P = {A ∪ C} ⊆ ti,j }|
SupportM I−ARM (A ∪ C) =
|T |
(1)
A major feature of the support quality measure in the field of MI-ARM is
that it is possible to obtain A and C that both satisfy all the transactions by
isolation, but if they are analysed together they do not satisfy any transac-
tion, i.e. A ∩ C = ∅. On the contrary, this assertion is not possible in classic
ARM since it considers each transaction as a single instance.
As for the confidence measure (Equation 2), it is defined as the pro-
portion of the number of transactions that include A and C among all the
transactions that comprise A, which is determined by the support of A.

SupportM I−ARM (A ∪ C)
Conf idenceM I−ARM (A ∪ C) = (2)
SupportM I−ARM (A)
Support and confidence are broadly conceived as the finest quality mea-
sures in ARM and, consequently, a great variety of proposals make use of
them by using certain minimum thresholds. Nevertheless, many authors have
considered that the mere fact of exceeding these quality thresholds does not
guarantee that the rules are interesting at all [3], and some different quality
measures have been proposed.
The lift quality measure (see Equation 3) establishes how many times
A and C occur together more often than would be expected if they were
statistically independent.

SupportM I−ARM (A ∪ C)
Lif tM I−ARM (A ∪ C) = (3)
SupportM I−ARM (A) × SupportM I−ARM (C)

The conviction metric (see Equation 4) represents the ratio of the ex-
pected frequency that A occurs without C, considering A and C statistically
independent sets, divided by the observed frequency of incorrect predictions.

1 − Lif tM I−ARM (A ∪ C)
ConvictionM I−ARM (Y ) = (4)
1 − Conf idenceM I−ARM (A ∪ C)
In a similar way to the lift measure, Leverage (see Equation 5) calculates
the proportion of additional cases covered by both A and C above those
expected if A and C were independent of each other.

9
 LeverageM I−ARM (A ∪ C) = SupportM I−ARM (A ∪ C)−
(5)
(SupportM I−ARM (A) × SupportM I−ARM (C))

Let us consider a dataset (see Table 1) for the market basket analysis,
which comprises information about products purchased by customers in a su-
permarket. This sample dataset includes information about twelve different
transactions comprising a set of different products. As previously described,
the same dataset can be analysed on multiple perspectives and MI-ARs can
be applied to any of these scenarios. This issue implies the strength of this
new task, since the market basket analysis can be carried out by multiple
perspectives depending on the goal of the managers.
Imagine that the managers require to carry out a market basket analysis
to obtain information about how the products are related for the sake of
improving the sales. In this regard, a descriptive analysis is performed by
on the general scenario, discovering relations such as IF Teenager THEN
Tuna (Support=0.5, Confidence=1.0). This association rule states that every
Teenager purchases Tuna each time that he/she comes to this supermarket.
A manager could advertise Tuna between Teenagers to increase profits.
There are many other descriptive analyses that could be done with the
same dataset and none of the existing ARM algorithms could be carried
out. For instance, let us now consider that the managers want to analyse the
relationships by group age (see Table 2), in such a way that relations between
products are obtained regardless the age of people that usually buy in the
supermarket. Considering data in this new perspective new associations

Table 1: Sample dataset for the market basket analysis

Transaction Time stamp ID Age group Fresh fruit Seafood
1 April 1 Senior Orange Tuna
2 April 2 Teenager Banana Tuna
3 April 4 Teenager Banana Tuna
4 June 1 Senior Orange Anchovy
5 June 3 Adult Banana Tuna
6 July 1 Senior Orange Anchovy
7 July 2 Teenager Banana Tuna
8 July 2 Teenager Orange Tuna
9 July 4 Teenager Banana Tuna
10 December 2 Teenager Banana Tuna
11 December 3 Adult Orange Anchovy
12 December 1 Senior Orange Anchovy

10
Table 2: Sample multiple instance dataset for the market basket analysis by using the
perspective of Age group
Transaction Time stamp ID Age group Fresh fruit Seafood
1 April 1 Senior Orange Tuna
June 1 Senior Orange Anchovy
July 1 Senior Orange Anchovy
December 1 Senior Orange Anchovy
2 June 3 Adult Banana Tuna
December 3 Adult Orange Anchovy
3 April 2 Teenager Banana Tuna
April 4 Teenager Banana Tuna
July 2 Teenager Banana Tuna
July 2 Teenager Orange Tuna
July 4 Teenager Banana Tuna
December 2 Teenager Banana Tuna

can be obtained when the group age is considered to cluster instances into
transactions. In this regard, the rule IF April THEN Tuna (Support=0.66,
Confidence=1.0) determines that April and Tuna are features that usually
appears together regardless the group age of the customer. Similarly, the rule
IF July THEN Orange denotes that July is a good month to sell Orange.
Continuing with the previous sample dataset, it is also possible to form
groups of transactions by using abstract attributes that are not inherently
in the dataset. For instance, let us consider that managers want to obtain
information about existing relationships between products regardless the sea-
son (Spring, Summer or Winter), so the new multiple-instance database is
organized as shown in Table 3. Considering this perspective, it is possible to

Table 3: Sample multiple instance dataset for the market basket analysis formed by ab-
stract attributes that are not inherently in the dataset, for example, the season.
Transaction Time stamp ID Age group Fresh fruit Seafood
1 April 1 Senior Orange Tuna
April 2 Teenager Banana Tuna
April 4 Teenager Banana Tuna
2 June 1 Senior Orange Anchovy
June 3 Adult Banana Tuna
July 1 Senior Orange Anchovy
July 2 Teenager Banana Tuna
July 2 Teenager Orange Tuna
July 4 Teenager Banana Tuna
3 December 1 Senior Orange Anchovy
December 3 Adult Orange Anchovy
December 2 Teenager Banana Tuna

11
discover the rule IF Tuna THEN Teenager (Support=1.0, Confidence=1.0),
which determines that in all the seasons exist at least one customer who buys
tuna and he/she is a teenager. From this knowledge, it is stated that tuna
is not a seasonal product as, for example, anchovy is. Notice that anchovy
is not usually sold in Spring, presenting a support value of 0.66, that is, it is
sold in 2 out 3 seasons.
All of this demonstrate the usefulness of using multiple-instance in as-
sociation rule mining, discovering knowledge that is hardly obtained with
traditional ARM algorithms. Finally, it should be noted that we have con-
sidered a sample dataset to understand the strength of using MI-ARM. In
this regard, the knowledge extracted in this example does not describe a real
scenario but denotes the high utility of this type of association rules.

4. Algorithm: Apriori-MI
In this work, we have introduced a new problem that is formulated
halfway between ARM and MIL so that the goal is to deal with multiple-
instance datasets from an unsupervised and descriptive point of view. To
date, there is no definition in this field, which could be considered as an un-
explored area. In this regard, we consider that this new problem requires to
be solved and analysed by considering key algorithms like Apriori, setting the
basis for further algorithms and leaving it open to further research studies.
In the ARM field, the best-known algorithm was proposed by Agrawal [1]
at the beginning of the 90s. This algorithm, known as Apriori, is based on an
exhaustive search methodology and divides the ARM problem into two sub-
problems: (a) obtaining all the frequent patterns in data and (b) extracting
all the ARs — according to a predefined minimum confidence threshold β
— starting from the results obtained in the previous step. While the second
step is straight forward, the first step needs more attention since finding all
the frequent patterns in data is not a trivial task. It involves the searching
for any feasible pattern, so considering a dataset containing n single-items,
the set of available patterns is 2n − 1 (excluding the empty set which is
not a valid pattern). Although the size of the set grows exponentially in
the number of patterns, an efficient search can be considered by using the
downward-closure property of support (also known as anti-monotone). This
property determines that if a length-k pattern is not frequent in data, none
of its length-(k + 1) super patterns can be frequent.

12
 For a better understanding, let us consider a dataset comprising the trans-
actions {a,c,d}, {b,c,d}, {b,d}, {c,d}, {b,c,d}. The first step considered by
the Apriori algorithm is to calculate the number of transactions where each
single-item is satisfied, that is, the absolute support or frequency of occur-
rence of each item separately ({a} = 1, {b} = 3, {c} = 4 and {d} = 5).
In the next steps, Apriori generates a list of all 2-pairs, 3-pairs, and so on.
As depicted in Figure 2, the process of mining patterns can be very time-
consuming, so the downward-closure property plays an important role by
pruning branches of the tree. Let us assume that the minimum support α to
qualify a pattern as frequent is 2, which depends on the context. Thus, any
super-set that comprises the single-item {a}, which is satisfied by only one
transaction, will be infrequent and can be pruned.
This methodology cannot be directly applied to the multiple-instance
problem since it deals with a set of transactions T and each transaction
ti ∈ T having a varied number k of instances, i.e. ti,1 , ti,2 , ..., ti,k . Definitions
previously described for the MI-ARM concept determines that a transaction
ti is satisfied if and only if at least one instance ti,j ∈ ti is satisfied by
P = {A ∪ C}, which form the rule A → C. In this sense, the original
Apriori algorithm is modified to deal with multiple instance datasets, giving

{null}

a (1) b (3) c (4) d (5)

ab(0) ac (1) ad (1) bc (2) bd(3) cd (4)

abc(0) abd (0) acd (1) bcd(2)

abcd(0)

Figure 2: Steps of the Apriori algorithm for mining frequent patterns. Shaded area repre-
sents infrequent patterns that can be removed according to the downward-closure property
considering 2 as the minimum value of frequency

13
rise to the Apriori-MI algorithm. Focusing on the Apriori procedure (see
Algorithm 1), it requires to iterate every transaction in the dataset and each
of these transactions includes a different set of instances that also need to

Algorithm 1 Pseudo-code of the Apriori-MI algorithm

Input: E, α
Output: L
procedure Apriori algorithm for multiple-instance problems
1: k = 1
2: L1 ← {k − patterns}
3: C ← ∅
4: A ← ∅
5: k = 2
6: while Lk−1 6= ∅ do S
7: Ck ← {a ∪ {b}|a ∈ Lk−1 ∧ b ∈ Lk−1 ∧ b ∈ / a}
8: for all transaction ti ∈ T do
9: for all instance ti,j ∈ ti do
10: if ∃ c|c ∈ Ck ∧ c ⊆ ti,j then
11: Ae ← c
12: count[c] ← count[c] + 1
13: break
14: end if
15: end for
16: end for
17: Lk ← {c|c ∈ Ck ∧ count[c] ≥ α}
18: k ←k+1
19: end while
S
20: L ← Lk
k
21: return L
end procedure
Input: L, β
Output: R
procedure Rule generation procedure
22: R = ∅
23: for all l ∈ L do
24: for all x ⊂ l such that x 6= ∅ and x 6= l do
25: if support(l)/suppor(x) ≥ β then
26: R ← R ∪ {x → (l − x)}
27: end if
28: end for
29: end for
30: return R
end procedure

14
be analysed to determine its coverage (see lines 8 to 16, Algorithm 1). A
really important feature of this evaluation process is that Apriori-MI does
not require to check all the instances within each transaction. Once a specific
instance is satisfied, then the remaining ones are not analysed (see lines 10
to 14, Algorithm 1) and the algorithm continues with the next transaction.
For a better understanding, consider the transactions {{a,c,d}, {b,c,d}}
and {{b,d}, {c,d}, {b,c,d}}, and a threshold value of 1, the Apriori-MI mines
patterns as shown in Figure 3. In this algorithm the downward-closure prop-

{null}

a (1) b (1) c (2) d (2)

{b,d}
{a,c,d} {c,d}
ab(0) ac (1) ad (1) bc (1) bd(1) cd (2) {b,c,d}
{b,c,d}

abc(0) abd (0) acd (1) bcd(1)

abcd(0)

(a) Patterns obtained from the first transac- (b) Patterns obtained from the second
tion transaction

{b,d}
{c,d}
{b,c,d}

(c) Patterns obtained from the resulting

patterns of the set of transactions

Figure 3: Apriori-MI algorithm by using a minimum threshold value of one. Shaded area
represents infrequent patterns that can be removed according to the downward-closure
property

15
erty can be applied to discard misleading patterns and to reduce the com-
putational time. For each transaction, Apriori-MI carries out an exhaustive
search (Figures 3(a) and 3(b)), obtaining any pattern that satisfies the min-
imum threshold value. The resulting set of patterns is obtained from the
sum of the sets previously obtained for each transaction (see Figure 3(c)),
considering as frequent those that satisfy most transactions. It should be
noted that each specific pattern is computed only once for each transaction.
In a second step and similarly to Apriori, the proposed Apriori-MI algo-
rithm discovers any MI-AR from the set of frequent patterns. Thus, the set
of patterns ac, ad, bc, bd, cd, acd and bcd are used to obtain a resulting set
of rules. This set of rules comprises all the possible combinations from the
set of frequent patterns: a → c, c → a, a → d, d → a, b → c, c → b, d → b,
b → d, d → c, c → d, a → c ∧ d, c → a ∧ d, d → a ∧ c, c ∧ d → a, a ∧ d → c,
a ∧ c → d, b → c ∧ d, c → b ∧ d, d → b ∧ c, c ∧ d → b, b ∧ d → c, and b ∧ c → d.

5. Experimental Analysis
In this section, we demonstrate the importance of MI-ARM by applying
it to a series of datasets. It is not our intention to compare results of different
algorithms but to provide an overview of the use of multiple-instance as an
unsupervised and descriptive task. First, we study the utility of using MI-
ARM by comparing results with regard to classic ARM. Then, we analyse the
computational cost by considering a set of artificial data comprising different
number of transactions, instances per transaction and attributes.

5.1. Strength of using MI-ARM

In this section, we carry out an in depth description about the strength
of using the Apriori-MI algorithm by considering the same sample market
basket dataset used along this paper (see Table 1). Results are compared
to the traditional Apriori algorithm, which was proposed to be applied on
transactional datasets.
Let us consider now that the managers require to carry out a market
basket analysis to obtain information about how the products are related for
the sake of improving the sales. In this regard, a descriptive analysis is per-
formed by using either the traditional Apriori algorithm and the proposed
Apriori-MI algorithm for mining frequent relationships on the general sce-
nario (see Table 1). A minimum support threshold of 50% (6 transactions)

16
and a minimum confidence threshold of 70% are considered. Once the Apri-
ori algorithm is executed, two frequent patterns were discovered: {Teenager,
Tuna} and {Banana, Tuna}. From these two frequent patterns it is possible
to obtain four different association rules to describe the data behaviour:
Rule Support Confidence
IF Teenager THEN Tuna 50% 100%
IF Tuna THEN Teenager 50% 75%
IF Tuna THEN Banana 50% 75%
IF Banana THEN Tuna 50% 100%

Considering the most accurate association rules discovered, it is possible

to state that every teenager purchases tuna each time that he/she comes to
this supermarket. There is no restriction in this rule to be satisfied, so a
manager can advertise tuna between teenagers to increase profits. Managers
can also consider both bananas and tuna as a pack to be sold since everyone
that buys banana also buys tuna. As shown, this is the only information
that anybody could obtain considering the support restrictions since no more
relationships could be discovered according to the predefined thresholds.
Nevertheless, there are many other descriptive analyses that can be done
with the same dataset and none of the existing ARM algorithms could be
carried out. For instance, consider that the managers want to analyse the
relationships by group age (see Table 2), in such a way that relations be-
tween products are obtained by Apriori-MI regardless the age of people that
usually buy in the supermarket. Considering data in this new perspective
and using the same support threshold, that is, a minimum value of 50%,
the following patterns of size one are obtained: {April}, {June}, {July},
{December}, {Banana}, {Orange}, {Anchovy} and {Tuna}. In the same
way, the following patterns of size 2 are obtained: {April, Tuna}, {July,
Orange}, {December, Orange}, {December, Anchovy}, {Banana, Tuna},
{Orange, Tuna} and {Orange, Anchovy}. Finally, the following pattern of
size 3 is obtained: {December, Orange, Anchovy}. From this set of frequent
patterns, the following association rules to describe the data behaviour are
obtained:
Rule Support Confidence
IF April THEN Tuna 66% 100%
IF July THEN Orange 66% 100%
IF December THEN Orange 66% 100%
IF Anchovy THEN December 66% 100%
IF Banana THEN Tuna 66% 100%
IF Anchovy THEN Orange 66% 100%
IF Orange AND Anchovy THEN December 66% 100%
IF Orange AND December THEN Anchovy 66% 100%
IF Anchovy AND December THEN Orange 66% 100%
IF Anchovy THEN Orange AND December 66% 100%

17
 These results are quite interesting to understand the usefulness of the
analysis of the data by considering different perspectives. First, the rules
obtained are different to the ones obtained by the traditional Apriori algo-
rithm, and the knowledge extracted is also different. For instance, the rule
IF Tuna THEN Banana, obtained by Apriori, determined that both banana
and tuna were frequent when they appear together. However, this rule is not
satisfied when we analyse the dataset by group age and using Apriori-MI,
denoting the high correlation between these two products and the group age.
Thus, once one of the existing group age stop purchasing the products, the
relationship is altered. On the contrary, the rule IF Banana THEN Tuna is
satisfied regardless the age group of the customers with an accuracy of 100%
in both cases, that is, considering and without considering any group age.
Second, new associations are obtained by Apriori-MI when the group age
is considered to cluster instances into transactions. In this regard, the rule IF
April THEN Tuna determines that April is a good month to sell Tuna in this
supermarket. Similarly, the rule IF July THEN Orange denotes that July is
a good month to sell Orange. These two associations were not discovered by
the traditional Apriori algorithm since they were satisfied by only 25% and
16% of the transactions, respectively.
Similarly to the previous example, in which transactions were grouped
by a specific attribute (age), the multiple instance databases might be ob-
tained by considering other attributes or even by means of an abstract at-
tributes that are not inherently in the dataset. For instance, consider that
the managers want to obtain information about existing relationships be-
tween products regardless the season (Spring, Summer or Winter), so the
new multiple-instance database is organized as shown in Table 3.
Analysing the dataset by considering the predefined abstract attribute
(season) and considering the same support threshold, that is, a value of
50%, the following patterns of size 1 are obtained: {ID 1}, {ID 2}, {ID 3},
{ID 4}, {Senior}, {Teenager}, {Adult}, {Orange}, {Banana}, {Tuna} and
{Anchovy}. In the same way, the following patterns of size 2 are obtained:
{ID 1, Senior}, {ID 2, Teenager}, {ID 3, Adult}, {ID 4, Teenager}, {ID 1,
Orange}, {ID 1, Anchovy}, {ID 2, Banana}, {ID 2, Tuna} and {ID 4, Ba-
nana}, {ID 4, Tuna}, {Teenager, Banana}, {Teenager, Tuna}, {Senior, Or-
ange}, {Senior, Anchovy}, {Orange, Tuna}, {Orange, Anchovy}, {Banana,
Tuna}. Considering the patterns of size 3, the following sets are obtained:
{ID 1, Senior, Orange}, {ID 1, Senior, Anchovy}, {ID 2, Teenager, Banana},
{ID 2, Teenager, Tuna}, {ID 4, Teenager, Banana}, {ID 4, Teenager, Tuna},

18
{Teenager, Banana, Tuna} and {Senior, Orange, Anchovy}. Finally, the fol-
lowing patterns of size 4 are obtained: {ID 1, Senior, Orange, Tuna}, {ID 2,
Teenager, Banana, Tuna} and {ID 4, Teenager, Banana, Tuna}.
Considering the aforementioned group of frequent patterns, a set of as-
sociation rules are obtained. In this regard, we have chosen a subgroup of
association rules that could be of interest to be described:

Rule Support Confidence

IF Teenager THEN Tuna and Banana 100% 100%
IF Tuna THEN Teenager 100% 100%
IF Tuna THEN Banana 100% 100%
IF Anchovy THEN Orange 66% 100%

Analysing the aforementioned rules, it is discovered that, whereas tra-

ditional Apriori determines that 75% of the customers that buy tuna are
teenagers, the fact of considering the season of the year as an abstract at-
tribute to obtain transactions modifies the knowledge extracted. In this
regard, Apriori-MI determines that in all the seasons exist at least one cus-
tomer who buys tuna and he/she is a teenager. Thus, the aforementioned
rule calculates an accuracy of 100% instead of 75% as it was obtained by
Apriori. From this knowledge, it is stated that tuna is not a seasonal prod-
uct as, for example, anchovy is. Notice that anchovy is not usually sold in
Spring, presenting a support of 66%, that is, it is sold in 2 out 3 seasons.
To sum up, we have demonstrated the usefulness of using multiple-instance
in association rule mining, discovering knowledge that is hardly obtained with
traditional ARM algorithms.

5.2. Experimental results and computational cost

The aim of this section is to analyse the scalability of the MI-ARM task
once that its strength was described in the previous subsection. In this
regard, we have considered the use of a varied set of 30 datasets with different
features1 . In this study, we have included datasets that comprise between
200 and 5000 different transactions, and the average number of instances per
transaction varies between 3.5 to almost 15. Additionally, we have considered

1
The datasets and the description of their features (number of transactions, attributes,
average number of instances per transaction, etc) are available at http://www.uco.es/
grupos/kdis/kdiswiki/MI-ARM.

19
a different number of attributes and values per attribute, considering datasets
that comprise between 30 to 240 different values. Despite the fact that there
are many multiple-instance datasets in the literature, we have considered
this set of artificial datasets to analyse, in a correct way, the performance
of the algorithm when the number of transactions, attributes and values per
attribute varies. In this regard, the computational cost could be well studied
since the features of the data are accordingly predefined for the study.
Table 4 shows the results obtained for the set of datasets. For each
quality measure, the Apriori-MI algorithm is run using a minimum support
threshold value of 0.1, so no rule having a lower support value is considered
by the algorithm. As for the confidence value, no threshold is considered.
Analysing the average results depicted in Table 4, the number of rules

Table 4: Number of rules and average values for five quality measures when Apriori-MI is
applied to a varied set of data.
Dataset # Rules Support Confidence Lift Conviction Leverage
#1 686 0.155 0.319 0.653 0.760 -0.100
#2 192 0.156 0.408 1.010 1.019 0.013
#3 160 0.125 0.404 1.115 1.098 0.012
#4 1480 0.147 0.324 0.702 0.796 -0.086
#5 262 0.154 0.407 1.014 1.026 0.001
#6 208 0.128 0.406 1.112 1.101 0.012
#7 2978 0.127 0.307 0.626 0.766 0.093
#8 140 0.240 0.548 0.994 1.003 -0.002
#9 210 0.167 0.502 0.994 1.011 -0.001
#10 36788 0.147 0.244 0.365 ∞ -0.176
#11 13966 0.129 0.250 0.424 0.498 -0.211
#12 430 0.313 0.560 0.987 0.996 -0.009
#13 7392 0.213 0.359 0.567 ∞ -0.227
#14 294 0.294 0.542 0.993 1.016 -0.007
#15 276 0.299 0.586 1.008 1.012 0.004
#16 26670 0.172 0.263 0.372 0.392 -0.241
#17 12964 0.120 0.221 0.369 0.475 -0.216
#18 538 0.289 0.538 0.998 1.002 0.001
#19 19998 0.191 0.334 0.586 0.561 -0.193
#20 35979 0.118 0.342 0.993 1.321 0.000
#21 422 0.176 0.359 1.002 1.224 0.001
#22 15523 0.164 0.366 0.898 0.982 0.012
#23 12033 0.126 0.322 0.921 1.002 0.012
#24 528 0.194 0.542 1.001 1.005 0.023
#25 1916 0.122 0.304 0.623 0.764 -0.092
#26 100 0.243 0.552 1.002 1.017 -0.001
#27 150 0.169 0.507 1.002 1.019 -0.001
#28 4498 0.226 0.372 0.581 ∞ -0.232
#29 276 0.277 0.518 0.982 ∞ -0.021
#30 224 0.309 0.592 1.024 1.029 0.002

20
discovered is higher when the number of values per attribute is lower. This
is caused by the higher number of transactions that could be satisfied when
the attributes comprise a small number of values. On the contrary, the
mere fact of considering a higher number of values per attribute implies that
the attributes or patterns are hardly satisfied so the number of transactions
comprising specific patterns is lower. Additionally, in some datasets, e.g.
dataset #10, dataset #13, dataset #28 and dataset #29, the average value
computed for the conviction measure is denoted as infinity. This determines
that, in at least one of the discovered rules, the confidence value is maximum,
i.e. value 1, so according to the conviction measure (see Equation 4), the
denominator of the equation is zero and the result is ∞. Figure 4 includes a

0,4 0,6

0,3
Confidence
Support

0,4

0,2

0,2
0,1

Apriori-MI 5 Apriori-MI 10 Apriori-MI 15 Apriori-MI 5 Apriori-MI 10 Apriori-MI 15

(a) Support measure for all the datasets (b) Confidence measure for all the datasets

0,1
1,2

1 0
Leverage
Lift

0,8 −0,1

0,6
−0,2

0,4
−0,3

Apriori-MI 5 Apriori-MI 10 Apriori-MI 15 Apriori-MI 5 Apriori-MI 10 Apriori-MI 15

(c) Lift measure for all the datasets (d) Leverage measure for all the datasets

Figure 4: Boxplot of different quality measures for all the datasets, considering 5, 10 and
15 different values per attribute

21
set of boxplot graphics that show the average values obtained for the quality
measures when a different number of values per attribute (5, 10 and 15) is
considered.
Finally, we carry out an analysis of the scalability of the proposed task
(see Figure 5). First, it should be noted that the scalability of ARM algo-
rithms have been widely studied by different researchers [14], demonstrat-
ing that the runtime of exhaustive search algorithms exponentially increases
with the number of attributes. In this study, we aim to analyse the per-
formance of the proposed algorithm for mining association rules in multiple

100 5 values 5x106

5 values
10 values
10 values
15 values
15 values

95 4x10 6

75
Runtime (sec)

# Comparsions

3x106
60

45
2x106

30

1x106
15

1000 2000 3000 4000 5000 1000 2000 3000 4000 5000
# Transactions # Transactions

(a) Performance of Apriori-MI for a different number of transactions

45 5 values
1019 5 values

10 values 10 values

15 values 15 values

42 1016

1013
Runtime (sec)

39
# Rules

1010

15
107
10
104
5

0 101

2 4 6 8 10 12 2 4 6 8 10 12
# Attributes # Attributes

(b) Performance of Apriori-MI for a different number of attributes

Figure 5: Analysis of the scalability of the Apriori-MI algorithm for different number
of both transactions and attributes. The analysis is carried out by using datasets that
comprise attributes with 5, 10 and 15 values each one.

22
instance datasets. In this regard, Figure 5 shows the scalability of the pro-
posed exhaustive search MI-ARM algorithm when it is applied to datasets
with different number of both transactions (Figure 5(a)) and attributes (Fig-
ure 5(b)). As execution time can be affected by numerous parameters and
highly depends on the hardware used, we have illustrated the results by using
both the runtime in seconds and a different metric —the number of compar-
isons required to mine the association rules and the amount of rules to be
discovered.
Considering the performance of Apriori-MI when the number of transac-
tions increases (Figure 5(a)), a linear growth is obtained, and there is no high
differences among datasets with different number of values per attribute. The
increasing number of transactions affects the runtime due to an increment
in the number of comparisons required to calculate the measures for each
rule. Thus, Figure 5(a) also shows how the number of transactions affects
the number of comparisons required.
Finally, considering the variation of the number of attributes, Figure 5(b)
illustrates that the runtime exponentially increases with the increment of the
number of attributes. Additionally, the number of attributes also has a huge
influence on the amount of rules to be discovered. Figure 5(b) also shows the
number of feasible rules to be mined for a different number of both attributes
and values per attribute.

6. Real Cases of Use

Finally, in this section, we apply MI-ARM to two real-world problems,
describing the importance of MI-ARM in both fields. We first apply MI-
ARM to the financial field, considering data gathered from an important
bank in Lithuania. Second, we apply the new proposal in accordance with
the Spanish public employment service.

6.1. Application to the financial field

In this section, we apply the new problem formulation to a real-world
problem. The aim is to demonstrate the applicability of the proposed prob-
lem regardless the algorithm used to this end, so we use a multiple-instance
dataset including 13,811 transactions with information about banking oper-
ations where each user is defined as a transaction comprising different oper-
ations for a specific day.

23
Data description. Data have been gathered from an important bank in
Lithuania, and comprise 13,811 transactions with 61,346 instances in general,
so the average number of instances per transaction is 4.44. The dataset is or-
ganised in features or attributes including: currency of the operations (Swiss
franc, Danish krone, Pound sterling, Latvian Lat, Norwegian krone, Pol-
ish zloty, Lithuanian litas, Russian ruble and Euro); amount local currency;
product application (personal account, overdraft, special deposits, child, ac-
count, mortgage savings account, provisions for loans, etc); account officer
id; and sector (bank, large corporate, medium corporate, other financial in-
stitution, charitable institutions, private person, etc). The goal is to describe
the data to determine interesting and strong relations between operations to
detect frequent financial habitats in the population.
Data have been organized into transactions that represent different bank
customers. Each transaction (customer) comprises a different number of in-
stances, one per successful banking operation. The data organization enables
MI-ARM to be applied on it, providing a description based on the customers’
habits in such a way that sporadic customers are as important as habitual
customers.
By using MI-ARM, an institution is able to improve their needs and offer
specific products to specific clients. Additionally, the description could be
used to detect anomalous operations and to avoid banking frauds. The goal
of this analysis is just to demonstrate the usefulness of the proposed MI-ARM
formulation to a real-world field.

Results. Applying MI-ARM formulation to the aforementioned dataset we

obtained a series of accurate rules that relate interesting clients’ behaviours.
Table 5 shows some MI-ARs obtained, depicting the support, confidence
and lift values. It should be noted that a pattern is satisfied in a specific
transaction if and only if at least one instance within the transaction is
satisfied.
First, we obtain two interesting MI-ARs that are somehow related: (1) IF
(amountlcy > -172,034.22 AND currency = LTL) THEN (productCategory
= current account), and (2) IF (currency = LTL) THEN (productCategory
= current account). The first rule is a specialization of the second one, which
determines that 94.16% of the banking operations involving Lithuanian litas
(LTL) are related to the current account of the clients. Both rules are satisfied
in the 88.34% of the transactions with a confidence of 0.94 and a lift value
of 1.025. The first rule determines that those transactions that involve an

24
Table 5: Multiple-instance association rules obtained from data of a bank of Lithuania.
Support and confidence are represented in per unit basis.
Rule Support Confidence Lift
IF (amountlcy = [-172,034.22, max] 0.883 0.942 1.025
AND currency = LTL)
THEN (productCategory = current account)

IF (currency = LTL) 0.883 0.942 1.025

THEN (productCategory = current account)

IF (sector = individuals) 0.201 1.000 1.001

THEN (amountlcy = [-31,512,532.83, max])

IF (currency = LTL)
THEN (amountlcy = [min, 48,255.16]) 0.873 0.999 1.021

amount higher than -172,034.22 LTL are related to the current account of
the clients. In other words, it could be said that it is quite difficult to find a
Lithuanian litas account whose amount of currency is lower than -172,034.22
LTL.
Second, we discover a very strong MI-AR that relates a specific sector to
the amount of currency used in the transaction, i.e. IF (sector = individuals)
THEN (amountlcy > -31,512,532.83). This rule, which is satisfied in 20.11%
of the transactions, is obtained with a maximum confidence value, denoting
that when the client that has ordered the banking operation is an individual
(he/she does not act as a company or institution) then the amount of cur-
rency is greater than -31,512,532.83. This rule means than, in general, lower
amount of currency is always managed by institutions or companies instead
of individuals.
Finally, an interesting MI-AR is discovered, IF (currency = LTL) THEN
(amountlcy ≤ 48,255.16), denoting that the amount of money used in opera-
tions that involve LTL as currency is always lower than 48,255, and this rule
presents an accuracy of 99.92%. This rule could give rise to determine that a
specific operation involving more than 48,255 Lithuanian litas is considered
outrageous and should be carefully revised since it could be a fraud.
In this analysis, it is also possible to obtain negative association rules,
which denotes a negative relationship between the antecedent and conse-

25
quent of the rule. For instance, the three following sample rules are negative
associations: (1) IF(currency 6= LVL) THEN (productCategory 6= charge ac-
cumulation); (2) IF (currency 6= SEK) THEN (productCategory 6= charge
accumulation); (3) IF (currency 6= CHF) THEN (productCategory 6= charge
accumulation). All of these rule presents a confidence value of 1.00, denot-
ing that none of the operations that use Latvian lats (LVL) are related to
a charge accumulation. In the same way, none of the operations that use
Swedish krona (SEK) or Swiss franc (CHF) are related to a charge accumu-
lation. The support of the rules are also extremely high, obtaining a value
of 0.99, so they are satisfied in most of the transactions.
Finally, we have applied the classic Apriori algorithm to the same dataset
in order to demonstrate that the results and the information described by an
association rule on multiple instance data is quite different to the information
provided by classic association rules. The rule IF (amountlcy = [-172,034.22,
max] AND currency = LTL) THEN (productCategory = current account) is
satisfied by 88.3% of the transactions when the data is organized into multiple
instances where each transaction is a different client and each instance is a
single banking operation. On the contrary, using a classic data representation
where each transaction is a different banking operation, the aforementioned
rule is satisfied by 68% of the transactions, which is quite different to the
support obtained by MI-ARM. Additionally, the rule IF (currency = LTL)
THEN (productCategory = current account) is really accurate in MI-ARs,
since 94.2% of the clients whose operations involve Lithuanian litas (LTL)
are related to the current account of the clients. Nevertheless, if we analyse
the same rule in classic ARM, we obtain an accuracy of 75%, since the
information represented is completely different as previously described in
Section 5.1.

6.2. Application to the Spanish public employment service

In this section, we also apply the new problem formulation to another real-
world problem in accordance with the Spanish public employment service.
Data comprises information about the unemployed in Córdoba, Spain.

Data description. Data have been gathered from the Spanish public em-
ployment service, which has an automatic system that enables records of the
unemployed to be recognised. This system, known as SARE, was designed
and implemented by the computing service located in Córdoba, Spain. The

26
aim of the SARE system is to inform the Spanish citizens about the unem-
ployment benefits that they could receive from the State party. Data were
gathered from 169 working days, from January 2014 to September 2014, and
comprises information about 63,185 citizens having 133 attributes each one.
Data have been organized into transactions that represent different re-
gions. Each transaction comprises a different number of instances, one per
unemployed. This data organization enables MI-ARM to be applied on it,
providing a description based on the regions’ information in such a way that
those regions with a lower number of unemployed are as important as others
with a higher number. Finally, data have also been organized into transac-
tions that represent people from different ages, so the results are not biased
against ages with a lower rate of unemployed.

Results. Applying MI-ARM to the aforementioned dataset we obtained a

series of rules with different grades of accuracy (known as confidence in this
field). Table 6 shows the MI-ARs obtained from data organized into regions,
depicting the support, confidence and lift values. It should be noted that a
pattern is satisfied in a specific transaction if and only if at least one instance
within the transaction is satisfied.
The two first rules show the relation between the gender and the number
of working days (considering all the regions as transactions of the dataset).
As described by the confidence quality measure, only 25% of the women who
are close to the retirement age have been working for more than a year. On
the contrary, 100% of the men who are close to the retirement age have been
working for more than a year. It shows a reality in Spain, where the fact of
obtaining a job is easier for men than for women, specially when they are
older than 50. Continuing the analysis, the following three rules determine
how important the age is to obtain a job. They determine that, regardless
the gender, the fact of being working for more than a year is much more
probable if a person is younger than 30, i.e. an accuracy of 66.6%, than if a
person is older than 51 — only 25% of the people older than this age have
been working more than a year. It should be noted that all these rules were
obtained from people who have requested for a unemployment benefit.
Continuing the analysis, the two last rules shown in Table 6 are analysed
together with the rules shown in Table 7. These two rules show the rela-
tionship between the gender and the number of working days without any
distinction between the age of the worker. It should be noted that, whereas
the rules depicted in Table 6 were obtained considering the transactions or-

27
Table 6: Multiple-instance association rules obtained from data of the Spanish public
employment service. Transactions organized by regions. Support and confidence are rep-
resented in per unit basis.
Rule Support Confidence Lift
IF (age = [51, 65] 0.111 0.250 0.562
AND gender = female)
THEN (Number of working days = [1 year, max])

IF (age = [51, 65] 0.111 1.000 2.250

AND gender = male)
THEN (Number of working days = [1 year, max])

IF (age = [min, 30]) 0.222 0.666 1.500

THEN (Number of working days = [1 year, max])

IF (age = [31, 50]) 0.333 0.375 0.844

THEN (Number of working days = [1 year, max])

IF (age = [51, max]) 0.111 0.250 0.562

THEN (Number of working days = [1 year, max])

IF (gender = female) 1.000 1.000 1.000

THEN (Number of working days = [1 year, max])

IF (gender = male) 1.000 1.000 1.000

THEN (Number of working days = [1 year, max])

ganized by regions, the rules depicted in Table 7 were obtained considering

the transactions organized by ages. If we compare both, it is stated that the
rules are much more accurate, i.e. a reliability of 100%, when the transac-
tions are organized by regions than when the transactions are organized by
ages, i.e. reliabilities lower than 40%. This difference shows the usefulness of
applying multiple-instance in the field of association rule mining, obtaining
different perspectives depending on the data organization. Considering the
aforementioned rules, in any Spanish region it is possible to find at least a
man or a woman who has been working for more than a year. On the con-
trary, considering that the transactions are organized by ages, this assertion

28
Table 7: Multiple-instance association rules obtained from data of the Spanish public em-
ployment service. Transactions organized by ages. Support and confidence are represented
in per unit basis.
Rule Support Confidence Lift
IF (gender = female) 0.333 0.375 0.844
THEN (Number of working days = [1 year, max])

IF (gender = male) 0.222 0.333 0.750

THEN (Number of working days = [1 year, max])

is not true, since only 33% of the ages comprise at least a person who have
been working for more than a year. This analysis is quite interesting, since it
shows how important is the working age to obtain a position in a company.
Finally, we have applied the classic Apriori algorithm to the same dataset
in order to demonstrate that the results and the information provided are
different to the obtained by classic association rules. The rule IF (age = [31,
50]) THEN (Number of working days = [1 year, max]) is satisfied by 33.3%
of the transactions when the data is organized into multiple instances where
each transaction is a different region and each instance is an unemployed. On
the contrary, using a classic data representation where each transaction is a
single unemployed, the aforementioned rule is satisfied only by 3.06% of the
transactions, which is quite different to the support obtained by MI-ARM.
Additionally, the following two rules were previously analysed by using
two different perspectives (Tables 6 and 7): (1) IF (gender = female) THEN
(Number of working days = [1 year, max]); (2) IF (gender = male) THEN
(Number of working days = [1 year, max]). Analysing these two association
rules by a classic methodology, we discover that the first rule is satisfied in
1.92% of the transactions, and the second one in 5.95% of the transactions.
These results are completely different to the ones obtained when the rules are
applied to multiple instance data, and the meaning is quite different too. For
instance, the rule IF (gender = female) THEN (Number of working days =
[1 year, max]) applied to the whole dataset represents a relationship between
“gender = female” y “Number of working days = [1 year, max] ” for all the
unemployed. Nevertheless, if we describe this rule for a multiple instance
point of view where transactions are organized by regions (Table 6) or ages
(Table 7), the meaning is referred to regions or ages, not to individuals.

29
7. Concluding remarks
The problem of multiple-instance learning, where data are grouped into
transactions and each transaction comprises a set of instances, has been dealt
by different researchers from a supervised learning point of view. Neverthe-
less, multiple-instance data is more and more common so the analysis and
description of the behaviour of this type of data is a dare. In this sense
we have proposed the concept of multiple-instance association rule mining,
which is a new problem that enables descriptions in multiple-instance data
to be obtained.
In this paper, we have formally presented the new concept of multiple-
instance association rule mining, which could be approached from different
perspectives. Nevertheless, for the sake of describing and applying this new
concept, we have considered the analysis of an exhaustive search algorithm
that is based on the well-known Apriori algorithm in the ARM field. Addi-
tionally, we have applied the new problem to two real-world fields, obtaining
interesting relationships between patterns in data inherently multi-instance.
Authors are aware that this new problem could be considered from dif-
ferent perspectives not considered in this work or even in classic ARM yet.
Algorithms and formulations described in this paper serve as a starting point
for further researches.

Acknowledgements
This research was supported by the Spanish Ministry of Economy and
Competitiveness, project TIN-2014-55252-P, and by FEDER funds.

References
[1] R. Agrawal and R. Srikant. Fast algorithms for mining association rules
in large databases. In Proceedings of 20th International Conference on
Very Large Data Bases, VLDB’94, pages 487–499, Santiago de Chile,
Chile, 1994. Morgan Kaufmann.

[2] S. Andrews, I. Tsochantaridis, and T. Hofmann. Support vector ma-

chines for multiple-instance learning. In Advances in Neural Information
Processing Systems, pages 561–568. MIT Press, 2003.

30
 [3] F. Berzal, I. Blanco, D. Sánchez, and M. A. Vila. Measuring the accuracy
and interest of association rules: A new framework. Intelligent Data
Analysis, 6(3):221–235, 2002.

[4] L. Cagliero, S. Chiusano, P. Garza, and G. Bruno. Pattern set mining

with schema-based constraint. Knowledge-Based Systems, 84:224 – 238,
2015.

[5] J. Chevaleyre and D. Zucker. Learning rules from multiple instance data:
issues and algorithms. In Proceeding of the 9th Information Processing
and Management of Uncertainty in Knowledge-Based Systems, IPMU,
pages 455–459, Annecy, France, 2002.

[6] T. G. Dietterich, R. H. Lathrop, and T. Lozano-Perez. Solving the multi-

ple instance problem with axis-parallel rectangles. Artificial Intelligence,
89(1–2):31 – 71, 1997.

[7] J. Han, J. Pei, Y. Yin, and R. Mao. Mining frequent patterns without
candidate generation: A frequent-pattern tree approach. Data Mining
and Knowledge Discovery, 8:53–87, 2004.

[8] M. Hassan, A. Karim, J.-B. Kim, and M. Jeon. CDIM: Document clus-
tering by discrimination information maximization. Information Sci-
ences, 316:87–106, 2015.

[9] H. Liu, S. Zhang, and X. Wu. Mlslr: Multilabel learning via sparse
logistic regression. Information Sciences, 281:310–320, 2014.

[10] J. M. Luna, A. Cano, M. Pechenizkiy, and S. Ventura. Speeding-Up As-

sociation Rule Mining With Inverted Index Compression. IEEE Trans-
actions on Cybernetics, pp(99):1–14, 2016.

[11] J. M. Luna, C. Romero, J. R. Romero, and S. Ventura. An evolutionary

algorithm for the discovery of rare class association rules in learning
management systems. Applied Intelligence, 42(3):501–513, 2015.

[12] J. M. Luna, J. Romero, and S. Ventura. On the adaptability of G3PARM

to the extraction of rare association rules. Knowledge and Information
Systems, 38(2):391–418, 2014.

31
[13] J. M. Luna, J. R. Romero, C. Romero, and S. Ventura. On the use
of genetic programming for mining comprehensible rules in subgroup
discovery. IEEE Transactions on Cybernetics, 44(12):2329–2341, 2014.

[14] J. M. Luna, J. R. Romero, C. Romero, and S. Ventura. Reducing gaps

in quantitative association rules: a genetic programming free-parameter
algorithm. Integrated Computer Aided Engineering, 21(4):321–337, 2014.

[15] O. Maron and T. Lozano-Perez. A framework for multiple-instance

learning. In Advances in Neural Information Processing Systems, pages
570–576. MIT Press, 1998.

[16] E. Ng, A. W.-C. Fu, and K. Wang. Mining association rules from stars.
In Proceedings of the 2002 IEEE International Conference on Data Min-
ing (ICDM 2002), pages 322–329, Maebashi City, Japan, 2002.

[17] C. Ordoñez, N. Ezquerra, and C. Santana. Constraining and summa-

rizing association rules in medical data. Knowledge and Information
Systems, 9(3):259–283, 2006.

[18] A. K. H. Tung, H. Lu, J. Han, and L. Feng. Efficient mining of in-

tertransaction association rules. IEEE Transactions on Knowledge and
Data Engineering, 15(1):43–56, 2003.

[19] J. Wang. Solving the multiple-instance problem: a lazy learning ap-

proach. In Proceeding of the 17th International Conference on Machine
Learning, ICML, pages 1119–1125, Standford, CA, USA, 2000. Morgan
Kaufmann.

[20] X. Yan, C. Zhang, and S. Zhang. ARMGA: Identifying interesting as-

sociation rules with genetic algorithms. Applied Artificial Intelligence,
19(7):677–689, 2005.

[21] A. Zafra and S. Ventura. Multi-instance genetic programming for pre-

dicting student performance in web based educational environments.
Applied Soft Computing, 12(8):2693 – 2706, 2012.

[22] X. Zhang and Z. Deng. Mining summarization of high utility itemsets.

Knowledge-Based Systems, 84(0):67 – 77, 2015.

32
 [23] Z. Zhou, K. Jiang, and M. Li. Multi-instance learning based web mining.
Applied Intelligence, 22(2):135–147, 2005.

[24] J. Zucker and Y. Chevaleyre. Solving multiple-instance and multiple-

part learning problems with decision trees and decision rules. Applica-
tion to the mutagenesis problem. In Proceedings of the 14th Biennial
Conference of the Canadian Society on Computational Studies of Intel-
ligence: Advances in Artificial Intelligence, pages 204–214, 2001.

[25] J. Zucker and J. Ganascia. Changes of representation for efficient learn-

ing in structural domains. In Proceedings of the 13th International Con-
ference on Machine Learning (ICML 1996), pages 543–551, Bary, Italy,
1996.

[26] J. Zucker and J. Ganascia. Learning structurally indeterminate clauses.

In Proceedings of the 8th International Workshop on Inductive Logic
Programming (ILP 1998), pages 235–244, Berlin, Germany, 1998.

33

View publication stats