Association Rule Learning Based Approach to Automatic Generation of

Feature Model Conﬁgurations

Olfa Ferchichi

1 a

, Raoudha Beltaifa

2,4 b

and Lamia Labed

3 c

Laboratoire Riadhi Ecole Nationale des Sciences de l’Informatiques, Tunisia

Sesame University, Tunisia

Carthage University, Tunisia

Tunis University, Tunisia

Keywords:

Software Product Lines, Variability, Feature Models, Conﬁguration, Advanced Apriori Algorithms,

Association Rule Learning.

Abstract:

The evolution the Software Product Line processes requires targeted support to address emerging customer

functionals and non functionals requirements, evolving technology platforms, and new business strategies. En-

hancing the features of core assets is a particularly promising avenue in Software Product Line evolution.The

manual conﬁguration of SPLs is already highly complex and error-prone. The key challenge of using feature

models is to derive a product conﬁguration that satisﬁes all business and customer requirements. However,

proposing a unsupervised learning-based solution to facilitate this evolution is a growing challenge. To address

this challenge, in this paper we use association rules learning to support business during product conﬁguration

in SPL. Based on extended feature models, advanced apriori algorithm automatically ﬁnds an optimal product

conﬁguration that maximizes the customer satisfaction. Our proposal is applied on a practical case involving

the feature model of a Mobile Phone Product Line.

1 INTRODUCTION

Software Product Line (SPL), a widely recognized

approach in software engineering (Clements, 2001)

(Ferchichi et al., 2020) (Pohl et al., 2005) involves

generating a range of interconnected products by

blending reusable core elements with custom assets

tailored to each speciﬁc product. Demonstrated as

an efﬁcient strategy for leveraging architecture-level

reuse, software product lines have established their

effectiveness in the ﬁeld. The SPL engineering

framework consists of two main processes: domain

engineering and application engineering. The domain

engineering process begins with domain analysis

to identify both common and variable features.

These features serve as the foundation for designing

and implementing the domain. Through domain

activities, software assets known as core assets are

created. Core assets are reusable components utilized

in developing Product Line (PL) products during the

https://orcid.org/0000-0003-2520-7588

https://orcid.org/0000-0003-4096-5010

https://orcid.org/0000-0001-7842-0185

application engineering process. These assets can

encompass various elements such as feature models,

architectures, components, or any other reusable

outcomes of domain activities.

One of the key aspects of product lines is variabil-

ity. Variability (Metzger and Pohl, 2014) (Bashroush

et al., 2017) (El-Sharkawy et al., 2019) is deﬁned as

the ability of a software system or artifact to be efﬁ-

ciently extended, changed, customized, or conﬁgured

(White et al., 2014) for use in a particular context.

In this paper, we propose an approach for automat-

ically deriving PL conﬁgrations based on association

rules learning.

The remainder of this paper is divided into six sec-

tions. in section 2 we present the fundamental con-

cepts of SPL. In Section 3, we present the problem

description. In Section 4, we present related work.

Section 5 describes the proposed advanced apriori al-

gorithms approach for conﬁguration SPL. The experi-

ment and the interpretation of the result are presented

in section 6. We give a conclusion in Section 7.

262

Ferchichi, O., Beltaifa, R. and Labed, L.

Association Rule Learning Based Approach to Automatic Generation of Feature Model Conﬁgurations.

DOI: 10.5220/0012742000003687

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 19th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2024), pages 262-272

ISBN: 978-989-758-696-5; ISSN: 2184-4895

2 BACKGROUND

2.1 Feature Model

Variability is expressed in a software variability

model, which such as Feature Model (FM). A FM

(Lee et al., 2002) is a tree or a directed acyclic graph

of features. A FM is organized hierarchically. A

mandatory feature has to be selected if its parent fea-

ture is mandatory or if its parent feature is optional

and has been selected. Mandatory features deﬁne

commonalities. Optional, alternative, and ‘or’ fea-

tures deﬁne variability in feature models. As a re-

sult, a feature model is a compact representation of all

mandatory and optional features of a software product

line. Each valid combination of features represents a

potential product line application.

In Figure 1, we present a Feature-Oriented Do-

main Analysis (FODA) model associated with the

’Mobile Phone’ SPL. In this model, we identify the

following elements: the root feature ’Mobile Phone,’

optional features such as ’stores’ and ’Bluetooth,’

mandatory features like ’keyboard’ and ’camera’. The

group_or relationship between ’Connectivity’ and its

children ’Mobile-Network’, ’Bluetooth’ and ’NFC’.

The alternative relationship with cardinality

∧

1..1

∨

is represented between the feature ’Sensors’ and

its children ’Indicative’ and ’Capactive’. The inter-

val noted [1..2] which determine the possible number

of instances of a feature ’Sim’. Extensions of FMs

can be grouped into three categories: 1) basic fea-

ture models (offering mandatory, alternative and ‘or’

features, as well as ‘requires’ and ‘excludes’ cross-

tree constraints), 2) cardinality-based feature mod-

els (Sreekumar, 2023) and 3) extended feature mod-

els (adding arbitrary feature attributes; e.g., to ex-

press variation in non_functionals requirements such

as quality (Ferchichi et al., 2021).

2.2 Non_Functional Requirements

Non_Functional requirements (NFR) are classiﬁed as

quantitative when they involve measurable informa-

tion (e.g., cost) and qualitative when the information

lacks measurability (e.g., customer preference level)

as discussed by (Gérard et al., 2007). FMs with Car-

dinalities address the challenges of SPL by modeling

the variability in terms of choices in Features. The

variability in the product family is represented by Fea-

ture cardinality, a cardinality group of features.

The concepts of conﬁguration and conﬁguration

processes were initially used in the ﬁelds of artiﬁcial

intelligence (Kumar, 2017) and problem-solving in

operational research, primarily to enable the conﬁg-

uration of physical products. (Faltings and Freuder,

1998) introduced a special issue of ’Intelligent Sys-

tems and their Applications’ focused on conﬁguration

processes. The two authors deﬁne this conﬁguration

process as a means of customizing parts of products to

meet speciﬁc consumer needs. They particularly em-

phasize three criteria that a conﬁguration must meet:

(i) it must be correct so that the company can deliver

the product; (ii) it must be produced quickly to avoid

losing the customer to the competition; (iii) it must be

optimal to convince the consumer. They also note that

these criteria strongly favor the automation of the con-

ﬁguration process, and the progress of e-commerce

tends to amplify this trend. In SPL,the conﬁguration

process is a major activity in application engineering,

involving the selection of desired features within the

PL before the product is realized. A FM represents

all possible product conﬁgurations in an SPL deﬁned

in Figure 1 . As an example, the mobile phone fea-

ture model can generate up to 1232 different product

variants (conﬁgurations). A sample product conﬁgu-

ration for the mobile phone product line is illustrated

in Figure 2 with FeatureIDE (Kaur and Kumar, 2014).

In SPLs, a signiﬁcant hurdle is the task of toggling

features within a feature model to craft new software

product conﬁgurations (Machado et al., 2014). As the

number of features increases in the model, so does the

array of potential product options. This process re-

sembles an optimal feature selection challenge within

an Extended Feature Model (EFM) when conﬁgur-

ing products within an SPL. However, without auto-

mated assistance, optimizing this selection becomes

cumbersome. It involves addressing various objec-

tives concurrently, such as aligning with user pref-

erences and maximizing feature selection. Identify-

ing the ’optimal’ products within these vast and con-

strained parameter spaces exceeds human intuition.

Thus, automated techniques for feature selection are

essential to streamline conﬁguration efforts.

3 PROBLEM DESCRIPTION

The primary concern with SPL is determining how

to conﬁgure an optimized feature set that meets the

customer’s requirements. In Figure 1, features cor-

respond to the functional requirements of the Mobile

Phone SPL. However, features may also be linked to

non-functional requirements. (Ferchichi et al., 2021),

have highlighted the importance of considering non-

functional requirements. For example, in the con-

text of the Mobile Phone SPL, it’s possible to iden-

tify non-functional requirements associated with each

feature, such as quality. This implies that every prod-

Association Rule Learning Based Approach to Automatic Generation of Feature Model Conﬁgurations

263

Figure 1: A Feature Model of the Mobile Phone SPL.

uct differs not only in terms of its functional features

but also due to its non-functional requirements.

Therefore,in our previous work (Ferchichi et al.,

2021), we proposed an extension of feature models

with non-functional requirements in the quality

aspect.

Figure 3 illustrates the non-functional require-

ments to extend the FM of Figure 1. In Figure 3, all

features have the quality attribute (see table 1). As a

motivating example, given a Mobile Phone PL that in-

cludes a variety of varying features, what is the prod-

uct that best meets the customer requirements limited

by a given quality? The challenge is that with hun-

dreds or thousands of features, it is hard to analyze

all different product conﬁgurations to ﬁnd an optimal

conﬁguration.

When it comes to manually conﬁguring extensive

features of interdependent SPLs, it might become im-

possible, especially if numerous features and depen-

dencies between features are involved. Additionally,

the conﬁguration process must be reiterated when-

ever the conﬁguration or implementation of an SPL

changes.

Ideally, a user should only have to conﬁgure a SPL

that encompasses the entire application scenario. The

user should not need to be concerned with the imple-

mentation details of underlying SPLs.

In this papier, we will discuss association rules,

a fundamental concept in data mining that allows us

to identify relationships among elements in a dataset.

We will explain the basic principles of association

rule mining, the various measures used to evaluate

association rules, as well as the different algorithms

Table 1: Non functionals requirements.

Quality attribute

Hardware Refers to the physical components

of a mobile phone, such as the pro-

cessor, memory, display, battery,

cameras, speakers, ports

Software Refers to the programs, applica-

tions, and operating systems that

run on a mobile phone, enabling

various functions and features be-

yond the hardware components.

Security Refers to feature designed to en-

hance the security of a system, de-

vice, application, or component

Behavioral Refers to a feature or attribute of a

system, organism, or entity that per-

tains to its behavior or the way it

acts or operates.

Structural Refers to a feature or attribute of an

object, system, or entity that per-

tains to its physical or organiza-

tional structure.

used to generate these rules.

Therefore, the main goal of the papier is to pro-

pose an automatic product conﬁguration approach

(see Figure 4) based on unsupervised learning.

Based on the association rule learning, our ap-

proach derives all the valid conﬁgurations of a PL,

which is modeled by an extended feature model. Then

the user can ﬁx some fucntional and non-fucntional

requirements and automatically the most satisfying

conﬁguration of a product is extracted.

ENASE 2024 - 19th International Conference on Evaluation of Novel Approaches to Software Engineering

264

Figure 2: Product Line Conﬁguration ’Mobile Phone’.

Figure 3: Extended FM by non-functional requirements.

4 RELATED WORK

Many past works were devoted to automatic conﬁgu-

ration of SPL In this section, we present approaches

related to our proposal.

(Batory, 2004) outlines AHEAD, a method for

conﬁguring Software Product Lines (SPLs) employ-

ing step-wise reﬁnement, where conﬁgurations are

iteratively reﬁned. Our approach shares similari-

ties, as it also involves selecting additional features

across multiple steps to achieve a desired conﬁgura-

tion (Hubaux et al., 2012)

introduced a formalism for establishing the work-

ﬂow necessary to conﬁgure a feature model in sev-

eral steps. The MUSCLES approach also emphasizes

conﬁguring a model through multiple steps. Never-

theless, (Hubaux et al., 2009)’s study does not explore

feature model drifts or the automated derivation of a

conﬁguration path from an initial to a ﬁnal conﬁgu-

ration. Additionally, MUSCLES includes support for

optimizations.

Various techniques for conﬁguring and validating

feature models in a single step have been proposed

((Benavides et al., 2013) (Heradio et al., 2022) (Be-

navides et al., 2005) (Ochoa et al., 2019))

These approaches leverage Constraint Satisfaction

Problems (CSPs) and propositional logic to generate

feature model conﬁgurations in a single stage and en-

sure their validity. Such techniques are valuable in

addressing the high complexity associated with deter-

mining a valid feature selection for a feature model

that satisﬁes a set of intricate constraints.

Association Rule Learning Based Approach to Automatic Generation of Feature Model Conﬁgurations

265

Figure 4: Overview of the Approach of Automatic generation of Product Conﬁgurations.

(Elsner et al., 2010) have examined variability

over time and the challenges associated with under-

standing the relationships between variability points.

MUSCLES concentrates on automating three essen-

tial tasks identiﬁed by Elsner et al. for managing vari-

ability over time. Speciﬁcally, MUSCLES offers ca-

pabilities for automating and optimizing tasks termed

by Elsner et al. as: (1) proactive planning, (2) track-

ing, and (3) analysis. While Elsner et al. focus on the

general identiﬁcation of issues in managing variabil-

ity over time, MUSCLES provides a framework for

automating the speciﬁc tasks outlined by Elsner et al.

as necessary in this context.

(Machado et al., 2014), present the SPLConﬁg, to

support business during product conﬁguration in SPL.

Based on feature models, SPLConﬁg automatically

ﬁnds an optimal product conﬁguration that maximizes

the customer satisfaction.

In their studies, (Berger et al., 2013) and (Men-

donca et al., 2009) utilized propositional logic to au-

tomate the validation of FM, depict staged feature

conﬁgurations, and offer assistance for manual fea-

ture selection. Nevertheless, their methods offer only

semi-automatic support for product conﬁguration, re-

lying solely on Feature Properties (FPs). Further-

more, these methodologies rely on exact exponential-

time algorithms such as Satisﬁability (SAT) or bi-

nary decision diagrams, which are unsuitable for SPL

applications due to their extensive computational re-

quirements and cumbersome nature.

Our proposal is deﬁned by an approach for auto-

matically deriving PL conﬁgurations based on the as-

sociating rule learning. From the generated conﬁgu-

rations the user can choose the quality attribute to be

considered in the conﬁguration of his need. In fact,

based the quality attributes deﬁned in the extended

feature model, the extracted conﬁguration, from all

the derived conﬁgurations, may satisfy the user.

5 AUTOMATIC

CONFIGURATIONS OF SPL

In Figure 4, we present an approach of deriving auto-

matically all the valid conﬁgurations of a PL deﬁned

by an extended feature model, as a ﬁrst step. We note

that the extended FM is a FM enriched with quality

steretypes. Having all the conﬁgurations enables the

automatic detection of dead features and falsely op-

tional ones. It’s important to note that a feature is

considered dead if it doesn’t appear in any conﬁgura-

tion of the SPL. A conﬁguration is considered falsely

optional if it appears in all conﬁgurations even though

it’s optional. With all the conﬁgurations of an SPL,

if these situations occur, the evolution process auto-

matically initiates the removal of the dead feature and

changes the optionality (from optional to mandatory)

of the falsely optional feature.

Then, having the valid conﬁgurations implies re-

moving automatically all conﬁgurations having dead

and optional features. In the next step, the customer

inserts a set of functional and nonfunctional features

to deﬁne his requirements, and automatically he gets

the better conﬁguration satisfying his requirements.

6 SPL CONFIGURATIONS

BASED ON ASSOCIATION

RULE MINING

Association rule mining is one of the most important

application ﬁelds of the data mining tasks. Associa-

tion rule is used to ﬁnd out the dependency of among

multiple domains based on the given degree of sup-

port and conﬁdence.

In this section, we present the unsupervised

Learning approach, and then we present our motiva-

tions to introduce this approach in improving EFM

conﬁguration.

Association Rule Mining (ARM) is a data min-

ing (Saxena and Rajpoot, 2021) technique that aims

to discover interesting relationships or associations

ENASE 2024 - 19th International Conference on Evaluation of Novel Approaches to Software Engineering

266

among variables in large datasets. These associa-

tions are often expressed in the form of rules that de-

scribe the co-occurrence patterns of items or attributes

within transactions.

ARM is unsupervised learning approach and is

a category of machine learning that uses labeled

datasets to train algorithms to predict outcomes and

recognize patterns. In our approach, we used the

advanced apriori algorithms (Suneetha and Krish-

namoorti, 2010) which is based on an association rule

learning.

6.1 Advanced Apriori Algorithms

An advanced apriori algorithm is used for frequent

item set mining and association rule learning over

databases. It proceeds by identifying the frequent in-

dividual items in the database and extending them to

larger and larger item sets as long as those item sets

appear sufﬁciently often in the database. In classical

Apriori algorithm, when candidate itemsets are gen-

erated, the algorithm needs to test their occurrence

frequencies. The manipulation with redundancy will

result in high frequency in querying, so tremendous

amount of resources will be expended in time or in

space. Therefore advanced algorithm was proposed

for mining the association rules in generating frequent

k-item sets.

Instead of judging whether these candidates are

frequent item sets after generating new candidates,

this new algorithm ﬁnds frequent item sets directly

and removes the subset that is not frequent, based on

the classical Apriori algorithm.

6.1.1 Key Metrics for Apriori Algorithm

The potential number of associations can be exten-

sive, particularly with a large number of items. The

challenge lies in determining which associations are

most meaningful. When employing the Apriori algo-

rithm, three essential metrics (Alsanad and Altuwai-

jri, 2022) are used : Support, Conﬁdence and Lift.

They are which are applied on rules.

• Support: The support of a rule or a set of at-

tributes (items) indicates the percentage of records

that satisfy the rule. (see formula 1)

Supp(X ⇒ Y ) =

X&Y

(1)

We denote:

– |X| : as the number of records containing at-

tributes X in the database (BD)

– |BD| as the total number of records.

• Conﬁdence: The conﬁdence of a rule measures

the validity of the rule: the percentage of exam-

ples that verify the conclusion among those that

verify the premise. (see formula 2)

Con f (X ⇒ Y ) =

X&Y

Con f (X ⇒ Y ) = Supp(X&Y )/Supp(X)

(2)

• Lift: is a measure that tells us whether the prob-

ability of an item Y increases or decreases given

item X. (see formula 3)

Li f t(X ⇒ Y ) =

Con f (X ⇒ Y )

Supp(Y )

(3)

In this paper we describe the improvement of clas-

sical Apriori algorithm in the following two aspects:

a) Reducing the passes of DB scan. b) Reducing the

unnecessary features generation.

6.1.2 Advanced Apriori Algorithm

The advanced apriori algorithm we applied in our ap-

proach is deﬁned in the following:

• Algorithm

Input : transactional database D and minimum

support threshold min_sup.

Output: L, frequent item-sets in D.

Method:

1) L

= frequent item-set of length 1,

2) Generate power set of L

and named as SPL

initialize with item-set_count=0, it will global for

entire algorithm.

3) For each transaction t in database Do.

a) For each item I in t Do,

Compare I with L

If (not match) then delete item from transaction t.

End Do.

b) Generate power set of t and named as items feaure

c) Compare item-sets of SPL with items feaure

d) If (item-set match) increase the item-set_count by

1 of SPL.

4) End Do.

Pruning phase:

5) For each item-set I in SPL Do,

6) If item-set_count of I is less than to min_sup

threshold the delete I

7) End Do,

8) Remaining item-set in SPL will be frequent

item-sets which holds min_sup threshold.

6.1.3 Description of the Algorithm Steps

The adavanced Apriori algo is described by the fol-

lowing steps

Association Rule Learning Based Approach to Automatic Generation of Feature Model Conﬁgurations

267

• (step 1): Initializing candidate sets items (cardi-

nality k = 1)

• (step 2 ): Filtering: Retain only frequent sets L,

i.e., with cardinality k whose support is ≥ supmin

(ﬁxed)

• (step 3): Joining: Combine all frequent sets of

cardinality k to calculate sets of cardinality k1 -

- Search for frequent items F1 :item-sets1

- Among F1, search for frequent item pairs F2:

item-sets2

- From F2, search for frequent item triplets F3:

item-sets3

• (steps 4) Testing: If the set of sets of cardinality

k+1 is empty, stop; otherwise, k = k+1, return to

step 2."

6.2 Conﬁguration of SPL Based on the

Advanced Apriori Algorithm

In this section, we present the application of the Ad-

vaced apriori algorithm for automatic generation of

SPL conﬁgurations. The database D is an Extented

Feature Model (EFM) of a SPL. An item is a feature

deﬁned in the EFM. A rule is deﬁned by a relationship

between two features. The support of a rule is deﬁned

according to the type of the relationship beween two

features(items).

Each relationship bewteen two features indicates

if they can be included in the same conﬁguration or

not, which has an inﬂuence on the support value of

associated rule.

• Mandatory

When a sub-feature has a mandatory relationship

with its parent, the conﬁguration cannot exhibit

the child feature unless it also exhibits the parent

feature. The support of this relationship is equal

to 100%. So, every mandatory feature must be in-

cluded in all conﬁgurations of the line (see Figure

5).

Figure 5: Support of the mandatory relationships.

• Optional

In an optional relationship, the feature can either

be chosen or not, so the probability of being cho-

sen is 1/2 and the probability of not being chosen

is also 1/2, thus the supmin=50.

• Alternative Group , cardinality group of features

∧

1..1

∨

and Constraint X-or

Figure 6: Support of the optional relationships.

With in the Basic and Extended Feature Models,

alternative Group, cardinality group of features

∧

1..1

∨

and Constraint X-OR have the same seman-

tics. In this relation, the minsupp is equal to 1/3,

and we can only choose a single feature, which

implies that we can extract only one rule.

Figure 7: Support of the alternative relationships.

• Or Group , cardinality group of features

∧

n..m

∨

and Constraint Or

Within the Basic and Extended Feature Models,

Or Group , cardinality group of features

∧

n..m

∨

and Constraint have the same semantics. The set

of sub-features C, D, and B are linked to the par-

ent feature A through a multiple-choice relation-

ship if one or more sub-features can be included

in a product presenting the parent feature. In this

example, we will apply in detail all the steps of

the advanced Apriori algorithm .

In this case, the supmin is the support of a feature

equal to 1/3 .We follow the next three steps of the

algorithm to compute the frequent features.

• Searching for Frequent Items F1 Item-Set 1

In this section, we will calculate the support of

each feature which belongs to item-set1. Sup-

port (B)= 1/3, support (C)=1/3 support (D)= 1/3

, Thus, the supmin is 1/3, and we apply the for-

mula support ≥ supmin so F1={B,C,D}

• Among F1, Searching for Frequent Items F2:

Item-Set 2

The itemsets are BC, BD, CD, support(BC)=

2/3, support(BD)= 2/3 ,support(CD)= 2/3, so sup-

Figure 8: Representation of the Or Group and cardinality

group of features

∧

n..m

∨

relationships.

ENASE 2024 - 19th International Conference on Evaluation of Novel Approaches to Software Engineering

268

port(BD),support(DC) and support (BC) =2/3 ≥

1/3 . F2= {BC, BD, CD}.

• From F2, Search for Frequent Item Triplets

Items F3: Item-Set 3

F3= {BCD}, support(BCD)=1 ≥ 1/3.

The Figure 9 illustrates the process of extract-

ing frequent itemsets. This involves traversing a lat-

tice and calculating the supports associated with each

combination. The number of conﬁgurations quickly

becomes very high, and each conﬁguration would re-

quire scanning the database. It is essential to consider

the minsupp parameter.

6.2.1 The Selected Rules

"Considering all the retained association rules in the

following table based on support calculations. In the

table, we have extracted all frequent rules across the

different relations of the ’Group’ or group.

Table 2: Extraction of Association Rules.

N° Rules Conﬁdence

1 A =⇒ B 100%

2 A =⇒ C 100%

3 A =⇒ D 100%

4 A =⇒ BC 100%

5 A =⇒ BD 100%

6 A =⇒ CD 100%

4 A =⇒ BCD 100%

The table represents the number of associated con-

ﬁgurations for these relationships.

Table 3: Numbers of conﬁgurations.

Features C1 C2 C3 C4 C5 C6 C7

B 1 0 0 1 0 1 1

C 0 1 0 1 1 0 1

D 0 0 1 0 1 1 1

6.2.2 Representation of Lift

The enhanced algorithm is outlined in the following

steps:

Input: Data: a transaction database

Min_sup: the minimum support count threshold

• Initially, each item is considered a member of the

set of feature item-set C1. The algorithm scans

all transactions to count the occurrences of each

item.

• The set of frequent item-sets, L1, is determined

by comparing the feature count with the minimum

support count, containing candidate 1-itemsets

satisfying the minimum support.

• To generate the set of frequent item-sets 2, L2,

the algorithm generates a feature set of item-sets

2. Then, the transactions in Data are scanned, and

the support count of each feature item set in C2 is

accumulated, repeating step 2.

• C2 is determined from L2.

• Generate C3 features from L2 and scan C2 for the

count of each feature, then repeat step 2.

• At the end of the pass, determine which feature

item sets are actually large, and those become the

seed for the next pass.

• This process continues until no new large item

sets are found (see Figure 10).

7 APPROACH EXPERIMENT

In this section, we present an application of the ap-

proach proposed for the Mobile Phone extended fea-

ture model (as depicted in Figure 3). We imple-

mented the approach using Python. Following this,

we demonstrate the application of the advanced Apri-

ori algorithm.

The extended feature model created for the SPL

mobile phone using FeatureIDE is stored in an XML

format. We edit the feature model both graphi-

cally and textually, simultaneously identifying fea-

tures. The feature model is stored in an XML ﬁle,

which we then imported into Python. As illustrated in

Figure 11 below.

Subsequently, we tested the algorithm with 6

transactions, selecting the ﬁrst three features with the

FeatureIDs F, F1, F2. We then applied the algorithm

to extract frequent rules. Through these rules, we aim

to retrieve the possible conﬁgurations’ versions.

The Figure 13 depicts a simulation of the example

of EFM with the mandatory relationship between the

three features: MobilePhone (F), Hardware (F1), and

Software (F2). Since the relationship between these

features is mandatory, we have set the support value

and minsup to 1.

So, itemset1 consists of these three features: item-

set1 = MobilePhone (F), Hardware (F1), Software

(F2)with cardinality = 1. Subsequently, from item-

set1, the combination of three features yields itemset2

with cardinality = 1. The association rules are gener-

ated with a conﬁdence of 100%, as shown in Figure

12.

Association Rule Learning Based Approach to Automatic Generation of Feature Model Conﬁgurations

269

Figure 9: Frequent itemsets extraction.

Figure 10: Example of Generation of feature item set and frequent item-set.

Figure 11: Example of Generation of feature item set and

frequent item-set.

To extract the conﬁguration of the SPL Mobile

Phone, we can use the association rules generated by

the advanced Apriori algorithm. Each rule represents

a possible conﬁguration of features. We can ﬁlter out

the rules that include the Mobile Phone feature and

Figure 12: Result of implementing the example of Genera-

tion of feature item set and frequent item-set.

extract the corresponding conﬁguration. The Figure

13 represents an excerpt of code from SPLs conﬁgu-

ration

ENASE 2024 - 19th International Conference on Evaluation of Novel Approaches to Software Engineering

270

Figure 13: Code snippet assumes that the conﬁgurations.

The Figure 14 represents the SPL conﬁguration

using the advanced Apriori algorithm. If a feature is

present in a conﬁguration, its value is set to 1; other-

wise, it is set to zero.

Figure 14: Extraction of the SPL Mobile Phone conﬁgura-

tion.

8 CONCLUSIONS

Feature model conﬁgurations in Software Product

(SPL) Lines still remain problematic and speciﬁcally

for Extended Feature models. This paper proposes the

use of an unsupervised learning approach for conﬁg-

uration generation. Hence, we use an advanced apri-

ori algorithm for frequent item set minig and asso-

ciation rule. These latters will be able to let feature

model generation in order to derive speciﬁc applica-

tions (SPL products). We have exeprimented all the

proposed approach on an extended mobile phone fea-

ture model. This latter is enriched by non functional

requirements, so complicating the conﬁguration. The

obtained results are promising and show that learning

is really an efﬁcient way for product derivations in the

context of Software Product Line. As future work, we

will experiment our approach on more SPL cases and

try to use other learning methods and compare their

results in the context of conﬁguration generations.

REFERENCES

Alsanad, A. and Altuwaijri, S. (2022). Advanced persistent

threat attack detection using clustering algorithms. In-

ternational Journal of Advanced Computer Science

and Applications, 13(9).

Bashroush, R., Garba, M., Rabiser, R., Groher, I., and Bot-

terweck, G. (2017). Case tool support for variability

management in software product lines. ACM Comput-

ing Surveys (CSUR), 50(1):1–45.

Batory, D. (2004). A tutorial on feature oriented program-

ming and the ahead tool suite (ats). Revised: Sept,

8:2004.

Benavides, D., Felfernig, A., Galindo, J. A., and Reinfrank,

F. (2013). Automated analysis in feature modelling

and product conﬁguration. In Safe and Secure Soft-

ware Reuse: 13th International Conference on Soft-

ware Reuse, ICSR 2013, Pisa, June 18-20. Proceed-

ings 13, pages 160–175. Springer.

Benavides, D., Segura, S., Trinidad, P., and Ruiz-Cortés, A.

(2005). Using java csp solvers in the automated analy-

ses of feature models. In International Summer School

on Generative and Transformational Techniques in

Software Engineering, pages 399–408. Springer.

Berger, T., Rublack, R., Nair, D., Atlee, J. M., Becker, M.,

Czarnecki, K., and W ˛asowski, A. (2013). A survey of

variability modeling in industrial practice. In Proceed-

ings of the 7th International Workshop on Variability

Modelling of Software-intensive Systems, pages 1–8.

Clements, P., N. L. (2001). Software product lines practices

and patterns, addison-wesley professional.

El-Sharkawy, S., Yamagishi-Eichler, N., and Schmid, K.

(2019). Metrics for analyzing variability and its im-

plementation in software product lines: A systematic

literature review. Information and Software Technol-

ogy, 106:1–30.

Elsner, C., Botterweck, G., Lohmann, D., and Schröder-

Preikschat, W. (2010). Variability in time-product line

variability and evolution revisited. VaMoS, 10:131–

137.

Faltings, B. and Freuder, E. C. (1998). Guest editors’ in-

troduction: Conﬁguration. IEEE Intelligent Systems,

13(04):32–33.

Ferchichi, O., Beltaifa, R., and Jilani, L. L. (2020). An onto-

logical rule-based approach for software product lines

evolution. In 2020 International Multi-Conference

on:“Organization of Knowledge and Advanced Tech-

nologies”(OCTA), pages 1–9. IEEE.

Ferchichi, O., Beltaifa, R., Jilani, L. L., and Mazo, R.

(2021). Towards a smart feature model evolution. IC-

SEA 2021, page 159.

Gérard, S., Espinoza, H., Terrier, F., and Selic, B. (2007). 6

modeling languages for real-time and embedded sys-

tems: Requirements and standards-based solutions.

Association Rule Learning Based Approach to Automatic Generation of Feature Model Conﬁgurations

271

In Dagstuhl Workshop on Model-Based Engineer-

ing of Embedded Real-Time Systems, pages 129–154.

Springer.

Heradio, R., Fernandez-Amoros, D., Galindo, J. A., Bena-

vides, D., and Batory, D. (2022). Uniform and scal-

able sampling of highly conﬁgurable systems. Empir-

ical Software Engineering, 27(2):44.

Hubaux, A., Classen, A., and Heymans, P. (2009). For-

mal modelling of feature conﬁguration workﬂows. In

Proceedings of the 13th International Software Prod-

uct Lines Conference (SPLC’09), San Francisco, CA,

USA, pages 221–230. SEI, Carnegie Mellon Univer-

sity.

Hubaux, A. et al. (2012). Feature-based Conﬁguration:

Collaborative, Dependable, and Controlled. PhD the-

sis, Citeseer.

Kaur, M. and Kumar, P. (2014). Mobile media spl creation

by feature ide using foda. Global Journal of Computer

Science and Technology, 14(3).

Kumar, S. L. (2017). State of the art-intense review on arti-

ﬁcial intelligence systems application in process plan-

ning and manufacturing. Engineering Applications of

Artiﬁcial Intelligence, 65:294–329.

Lee, K., Kang, K. C., and Lee, J. (2002). Concepts and

guidelines of feature modeling for product line soft-

ware engineering. In International Conference on

Software Reuse, pages 62–77. Springer.

Machado, L., Pereira, J., Garcia, L., and Figueiredo, E.

(2014). Splconﬁg: Product conﬁguration in software

product line. In Brazilian Congress on Software (CB-

Soft), Tools Session, pages 1–8. Citeseer.

Mendonca, M., W ˛asowski, A., and Czarnecki, K. (2009).

Sat-based analysis of feature models is easy. In Pro-

ceedings of the 13th International Software Product

Line Conference, pages 231–240.

Metzger, A. and Pohl, K. (2014). Software product line en-

gineering and variability management: achievements

and challenges. Future of software engineering pro-

ceedings, pages 70–84.

Ochoa, L., González-Rojas, O., Cardozo, N., González,

A., Chavarriaga, J., Casallas, R., and Díaz, J. F.

(2019). Constraint programming heuristics for con-

ﬁguring optimal products in multi product lines. In-

formation Sciences, 474:33–47.

Pohl, K., Böckle, G., and Van Der Linden, F. (2005). Soft-

ware product line engineering: foundations, princi-

ples, and techniques, volume 1. Springer.

Saxena, A. and Rajpoot, V. (2021). A comparative analysis

of association rule mining algorithms. In IOP Con-

ference Series: Materials Science and Engineering,

volume 1099, page 012032. IOP Publishing.

Sreekumar, A. (2023). Guided requirements engineering

using feature oriented software modeling.

Suneetha, K. and Krishnamoorti, R. (2010). Advanced ver-

sion of a priori algorithm. In 2010 First Interna-

tional Conference on Integrated Intelligent Comput-

ing, pages 238–245. IEEE.

White, J., Galindo, J. A., Saxena, T., Dougherty, B., Bena-

vides, D., and Schmidt, D. C. (2014). Evolving feature

model conﬁgurations in software product lines. Jour-

nal of Systems and Software, 87:119–136.

ENASE 2024 - 19th International Conference on Evaluation of Novel Approaches to Software Engineering

272