Toward a Deep Contextual Product Recommendation for SO-DSPL

Framework

Najla Maalaoui

, Raoudha Beltaifa

and Lamia Labed Jilani

RIADI Lab., National School of Computer Sciences, Manouba University, Tunisia

Keywords:

Service Oriented Dynamic Software Product Lines, Recommender System, User Requirements, Ontology,

Deep Neural Network.

Abstract:

Today’s demand for customized service-based systems requires that industry understands the context and

the particular needs of their customers. Service Oriented Dynamic Software Product Line practices enable

companies to create individual products for every customer by providing an interdependent set of features

presenting web services that are automatically activated and deactivated depending on the running situation.

Such product lines are designed to support their self-adaptation to new contexts and requirements. Users

conﬁgure personalized products by selecting desired features based on their needs. However, with large feature

models, users must understand the functionalities of features and the impact of their gradual selections and

their current context in order to make appropriate decisions. Thus, users need to be guided in conﬁguring

their product. To tackle this challenge, users can express their product requirements by textual language and a

recommended product will be generated with respect to the described requirements. In this paper, we propose

a deep neural network based recommendation approach that provides personalized recommendations to users

which ease the conﬁguration process. In detail, our proposed recommender system is based on a deep neural

network that predicts to the user relevant features of the recommended product with the consideration of their

requirements, contextual data and previous recommended products. In order to demonstrate the performance

of our approach, we compared six different recommendation algorithms in a smart home case study.

1 INTRODUCTION

On-time and On-demand service-based systems are

constantly increasing from day to day. Thereby, soft-

ware application providers have to take drastic mea-

sures to speed up their development process in order

to survive in the ever-demanding competitive soft-

ware market and deliver software systems accord-

ing to speciﬁc customer needs . Faced with such a

challenge, Service Oriented Architecture (SOA) pro-

vides a promising means for supporting continuously

changing of customers’ requests, needs, requirements

and contexts. To address this issue, the reuse ap-

proach has been suggested as one of the pioneer so-

lutions. The Service Oriented Dynamic Software

Product Line (SO-DSPL) technique has already been

adopted as one of the most promising techniques for

reuse. This framework is addressed by combining

SOA with Dynamic Software Product Lines Engi-

https://orcid.org/0000-0002-3896-920X

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

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

neering (DSPLE). A SO-DSPL (Capilla et al., 2014)

is known as a family of service-oriented systems shar-

ing a set of common features and managing a set of

variable features, which are managed according to

the needs of a speciﬁc market segment and/or envi-

ronment by its automatically activation and deactiva-

tion according to the running situation. SO-DSPLs

support their self-adaptation to new contexts and re-

quirements. In SO-DSPL framework, deriving a cus-

tomized software refers to the dynamic conﬁguration

process that is deﬁned by the decision-making pro-

cess of dynamically selecting/activating/deactivating

a set of features from the product line that complies

with the feature model (FM) constraints and fulﬁll the

product’s requirements and the running context.

To enable such a dynamic conﬁguration, sev-

eral approaches have been proposed (Jonathan et al.,

2005). Nevertheless, the applicability of these ap-

proaches is still limited. In particular, for SO-DSPLs

with a huge exponential conﬁguration space, they

have shown to be infeasible. The diversity of options

provided by the product line and their sensitivity to

138

Maalaoui, N., Beltaifa, R. and Jilani, L.

Toward a Deep Contextual Product Recommendation for SO-DSPL Framework.

DOI: 10.5220/0011855800003464

In Proceedings of the 18th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2023), pages 138-148

ISBN: 978-989-758-647-7; ISSN: 2184-4895

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

different factors, such as the user’s context and the

context of the DSPL, make conﬁguration more difﬁ-

cult. With the large number of features, the user may

have difﬁculty understanding the features, the rela-

tionships between them and the constraints. To tackle

these challenges, expressing user needs in textual for-

mat became a necessity to help them, furthermore rec-

ommendation techniques have become also essential

to efﬁciently ﬁlter the huge amount of SPL variants

and suggest to the user suitable products. In recom-

mender systems, user requirements may be inferred

from consumption patterns and contextual informa-

tion can be used to ensure relevant product recom-

mendations (Pereira et al., 2016).

In this paper, we propose a deep-learning based

approach encompasses a requirement-based recom-

mender system that suggest to the user a product

based on its textual product requirements and its con-

textual information. The main contribution of this pa-

per is to ease the conﬁguration process by effectively

predicting which combination of feature is the best

suited to users, based on their requirements that are

collected in several ways. In addition, domain experts

and product developers can also take advantage of this

approach as it offers a simple conﬁguration process.

In summary, we provide these main contributions:

• We propose a requirement-based personalized

recommender system that suggest to users the

product conﬁguration based on their given tex-

tual requirements and their contextual informa-

tion. The system is based on the products chosen

by previous users to generate personalized recom-

mendations for a current user.

• We reuse a set of swrl rules proposed by DSPL

sub-ontology (Maalaoui et al., 2021) in order to

verify the validation of the product conﬁguration

with the DSPL constraint and context.

• As a pre-treatment phase, we use the approach

presented in (Maalaoui et al., 2022) to extract the

core requirements from the user requirements in

order to have data that follows the same pattern

and consequently have a reliable recommenda-

tion.

• We empirically evaluate the performance of our

proposed recommender system on “smarthome”

dataset (Murukannaiah et al., 2016) that contain

crowd user requirements with different additional

information. We extend the given data by other

contextual data based on a proposed SO-DSPL

ontology (Maalaoui et al., 2021) and we extract

manually the associated products (the set of se-

lected features) in order to evaluate our proposed

approach.

2 BACKGROUND

2.1 Software Product-Line Engineering

Software Product-line engineering is a paradigm

within software engineering, used to deﬁne and derive

sets of similar products from reusable assets (Capilla

et al., 2014). The development life cycle of a prod-

uct line encompasses two main processes: domain en-

gineering and application engineering (Capilla et al.,

2014). While domain engineering focuses on estab-

lishing a reuse platform, application engineering is

concerned with the effective reuse of assets across

multiple products. Domain engineering is responsi-

ble of deﬁning all common and variable assets of a

product line and their respective interdependencies.

The aim of the application engineering process is to

derive speciﬁc applications by exploiting the variabil-

ity of the software product line. Feature modeling

is the main activity to represent and manage PL re-

quirements as reusable assets by allowing users to de-

rive customized product conﬁgurations (Capilla et al.,

2014). Product conﬁguration refers to the decision-

making process of selecting an optimal set of features

from the product line that comply with the feature

model constraints and fulﬁll the product’s require-

ments (Bashari et al., 2017). A common visual rep-

resentation for a feature model is a feature diagram

(Kang et al., 1990). The feature diagram deﬁnes com-

mon features found in all products of the product line,

known as mandatory features, and variable features

that introduce variability within the product line, re-

ferred to as optional and alternative features. In ad-

dition, feature diagrams often contain cross-tree con-

straints: a feature can include or exclude other fea-

tures by using requires or excludes constraints, re-

spectively. Figure 1 shows an example of a feature

diagram associated to the Smart Home SO-DSPL.

2.2 Service Oriented Dynamic Software

Product-Line Engineering

SPL approaches moved to recent development ap-

proaches like Dynamic Software Product Lines as an

emerging paradigm to handle variability at runtime

and at any time. Dynamic software product lines are

able to dynamically switch an executing system from

one variant to another, without stopping its execu-

tion, without violating its feature model’s constraints

and without degrading its functional and nonfunc-

tional behavior. Consequently, the feature model it-

self must become a runtime entity that can dynami-

cally evolve to satisfy new variability dimensions in

the system. Dynamic Software Product Lines (SPLs)

Toward a Deep Contextual Product Recommendation for SO-DSPL Framework

139

Figure 1: Smart home SO-DSPL.

provide conﬁguration options to adjust a software sys-

tem at runtime to deal with changes in the users’ con-

text (Capilla et al., 2014). Service oriented dynamic

software product lines (SO-DSPL) represent a class of

DSPLs that are built on services and service-oriented

architectures (SOAs) (Capilla et al., 2014).

3 RELATED WORKS

Due to the importance and the complexity of SPL

conﬁguration process, a large body of literature has

been dedicated to facilitate conﬁguration process

by recommending features. Several approaches are

based on feature models. In this scenario, some stud-

ies (Martinez et al., 2015a) aim to predict the util-

ity of an entire set of features that compose a prod-

uct, others (Bagheri et al., 2010a),(Bagheri et al.,

2010b),(Czarnecki et al., 2012) aim to predict the util-

ity of each feature for the user. Mazo et al. (Mazo

et al., 2014) present a collection of recommendation

heuristics to prioritize choices and recommend a col-

lection of features to be conﬁgured with the objec-

tive of reducing the number of conﬁguration. How-

ever, conﬁguration process still difﬁcult while the au-

thors do not propose any mechanism to help users and

contextual information are not considered by the pro-

posed recommendation heuristics.

Bagheri et al. (Bagheri et al., 2010a) propose a

feature ranking approach based on soft and hard con-

straints in order to satisfy the soft stakeholder con-

straints. Then, the proposed decision maker select the

recommended feature interactively until the full con-

ﬁguration of the feature model.

In (Pereira et al., 2016), Pereira et al. propose

a recommendation approach that provides the deci-

sion maker with a smaller set of relevant features,

such as a top-10 classiﬁcation. The authors adapt

in their work six state-of-the-art recommender algo-

rithms to the product line conﬁguration, which are

neighbourhood-based collaborative ﬁltering, collab-

orative ﬁltering-signiﬁcance weighting, collaborative

ﬁltering-shrinkage, collaborative ﬁltering-Hoeffding,

average similarity, and matrix factorization. Mar-

tinez et al. (Martinez et al., 2015b) propose an ap-

proach that predict conﬁgurations’ relevance. The au-

thors approach uses tailored data mining interpolation

techniques to predict conﬁguration likability based on

people’s vote feedback for a dataset of conﬁgurations.

Their approach is developed in two phases. First, a

dataset of conﬁgurations is created using a genetic al-

gorithm. Second, based on people’s vote feedback for

the dataset of conﬁgurations, tailored data mining in-

terpolation techniques are used to predict conﬁgura-

tion likability. Thus, a ranking is created among all

possible conﬁgurations.

Few works are interested in product conﬁguration

based on user requirements. Shamim Ripon et al.

(Ripon et al., 2020)propose an approach that extracts

automatically requirements, derive conﬁguration and

verify product conﬁguration. The proposed approach

is divided into two phases: requirement extraction and

product conﬁguration veriﬁcation. In requirement ex-

traction phase, requirements are collected from the

user as a textual description, then features are ex-

tracted, and a technical product conﬁguration is de-

rived. In the second phase, the derived conﬁguration

is checked for validation before sending a feedback

to user. In (Maalaoui et al., 2022), the authors pro-

posed an approach to better understand the user re-

quirements, to predict other information about the re-

quirements and to derive an appropriate service (soft-

ware application as a combination of several services)

in the SO-DSPL application engineering phase. In the

ﬁrst step, the core requirements are extracted from

user requirements, then features and knowledge are

extracted to populate in an ontology which infer rele-

vant knowledge and execute predeﬁned rules to derive

or adapt a product with the consideration of contex-

tual information.

Although there are many interesting studies that

aim to make the conﬁguration process easier by rec-

ommending features and products, the process is still

far from trivial. Even features recommendation and

ENASE 2023 - 18th International Conference on Evaluation of Novel Approaches to Software Engineering

140

ranking algorithms are proposed, the user remains

very involved in the conﬁguration process while FM

still complicated. However, contextual information is

not taken into account during the conﬁguration pro-

cess. Also, the satisfaction of the recommended fea-

tures with respect to the DSPL context is not yet ver-

iﬁed, while a combination of features may be valid

with respect to the constraints of the DSPL but invalid

with respect to the DSPL context.

To address this issue, we propose a requirement-

based recommender system to suggest to the user a

product conﬁguration (set of features) that satisﬁes its

requirements. In contrast to the current literature, our

approach predicts features that are not mentioned in

the given requirements. The proposed recommenda-

tion approach takes into consideration the user con-

text and the derived product conﬁguration is veriﬁed

with respect to the DSPL context. As well, our pro-

posed approach beneﬁts from the products conﬁgu-

rations space, their associated requirements and the

contextual information of their users and uses this

data to learn intelligently by the use of deep leaning

models.

4 CONTEXTUAL PRODUCT

RECOMMENDATION

FRAMEWORK

According to the limitations presented in related

works (see Section 3), recommending products in

DSPL framework is still challenging. To ease

this process, we propose a product recommendation

framework that aims to attend these main objectives:

1. To recommend a set of features which compose a

product to an active user

2. To consider contextual information in the recom-

mendation process.

3. To recommend a product adaptation in the case of

requirements or context changes.

As shown in the Figure 3, our recommendation frame-

work takes as input a set of requirements given by the

user, which are associated to the desired product, and

a set of contextual information related to the active

user. Given a product requirements and user context,

our recommender system is responsible to predict a

set of features that satisfy users’ requirements, its pro-

ﬁle, context and its future context changes with the

respect to the DSPL properties in one hand, and pre-

dict an adaptation of the product in the case of context

change. To recommend relevant products and ensure

the integration of an intelligent learning process, our

proposed recommendation framework exploits the ad-

vantages of machine learning algorithms and the pro-

posed ontology in (Maalaoui et al., 2021). Thus, we

use:

• The NLP approach presented in (Maalaoui et al.,

2022) to extract relevant information from the

given product requirements. The relevant infor-

mation extracted from the product requirements

are named core requirements and this step is

called “Core requirement Recognition”.

• A neural network model to predict relevant feature

based on the extracted core requirements with the

attention of contextual data. The proposed model

supports the dynamic change of requirements and

context by recommending the adapted product.

Our proposed recommendation process starts by

extracting core requirements from product require-

ments in order to keep only relevant information.

Then, based on the user context and its given core

product requirements, the recommended features are

predicted based on a deep learning model which is

trained on previous products requirements, their as-

sociated products, and the contextual information of

their users. Finally, the combination of the prediction

features is veriﬁed based on the proposed DSPL sub-

ontology presented in (Maalaoui et al., 2021).

4.1 Core Requirement Recognition

The core requirement Recognition phase is the data

pre-preparation step. In this step, the objective is to

remove non-essential and noisy data from the given

user requirements, by the interpretation of the user

requirements through his/her textual request and its

transformation to a formal representation (require-

ments structure deﬁned as ”Core Requirement”) that

will be later used by the recommendation process. In

this step we use the method presented in (Maalaoui

et al., 2022), where the input is a textual product re-

quirements given by a user, and the output is the prod-

uct core requirements built in accordance with the

core requirement structure.

A core requirements has the following structure:

<feature>+ <obligationdegree>? <goal>+ <item

>* <condition>*

Where :

– * : denotes a zero repetition to an inﬁnite number

of times repetitions.

– + : denotes repetition once or more number of

times.

– ? :denotes a repetition zero or one time.

– — :denotes a disjunction (it signiﬁes OR).

– ! : denotes a negation

Toward a Deep Contextual Product Recommendation for SO-DSPL Framework

141

Figure 2: Contextual product recommendation framework overview.

As an example, the input is deﬁned by this prod-

uct requirement:

My sensor should detect movement. if I get up late

night, lights in my house should turn on to enhance

convenience and safety”.

As a result, this product core requirements are gener-

ated:

Sensor should detect movement. Lights should turn

on if I get up nigh. The used method is based on a

set of linguistic rules and the support of uncertainty.

These rules facilitate the building of core requirement

structure which is then loaded as an instance of an on-

tology that infer new relevant knowledge.

4.2 Product Recommendation Based on

Deep Learning Architecture

Understanding users’ requirements is very crucial for

the success of software development process, espe-

cially in dynamic software product line engineering.

As an example, in the DSPL framework, the success

of product derivation process is based on the good

matching between the expressed user requirements

and the features that satisfy them. However, match-

ing users requirements that are expressed in a natu-

ral language and its corresponding features with re-

gards to additional information (i.e. the user context)

is a challenge. In this section, we present the product

derivation approach, which is responsible for recom-

mending a set of features that constitutes a valid prod-

uct with the goal of the satisfaction of the expressed

users’ core requirements and its contextual informa-

tion.

Our proposed recommendation process exploits

previous conﬁgurations (i.e. selected features ), their

correspondence with the expressed requirements and

their contextual information in order to predict hid-

den features that the user has not the attention to note

them in his requirements. To achieve our objective,

we propose a framework based on the combination of

deep neural networks, where convolutional and recur-

rent neural networks are combined together to predict

the corresponding product’s features.

The choice of this combination is based on the

success of Deep neural networks text learning and

processing and in recommender system (Katarya and

Arora, 2020). Recurrent neural networks (RNN) and

short-term memory (LSTM) have been widely ap-

plied to text, they have good performance for learn-

ing and processing text representation (Sengar et al.,

2021). Convolutional Neural Networks (CNN) has

also been applied to text classiﬁcation problems. Es-

pecially when integrating with the recurrent model,

recurrent-CNN can achieve better performance. As

examples, C-LSTM utilizes CNN to extract a se-

quence of higher-level phrase representations and are

fed into a long short-term memory recurrent neural

network (LSTM) to obtain the sentence representa-

tion. The 2-D CNN can extract not only local fea-

tures between adjacent words but also the small re-

gions (word stem) inside of words than 1-D CNN. Bi-

LSTM can extract sequential features from both past

and future.

ENASE 2023 - 18th International Conference on Evaluation of Novel Approaches to Software Engineering

142

Thus, in our proposed approach, the stacked 2-D

CNN and Bi-LSTM are combined together to rec-

ommend product’s features. This combination can

achieve the best accuracy in our problem, because

they can extract not only local features between adja-

cent words through convolution layers but also global

features (long-term dependencies) of sentences by re-

current layers. As well, our proposed framework

adopts Roberta (Bidirectional Encoder Representa-

tions from Transformers)(Liu et al., 2019) to embed

both products requirements and user contextual in-

formation with a context-dependent word embedding

method. Roberta is a method of pre-training language

representations, which can generate different word

embedding for a word based on its context(the other

words in the sentence).

As consequence, our proposed model can not only

automatically abstract the features from core require-

ments, but also the features from contextual infor-

mation such as user roles, user interest, user prefer-

ence and other contextual information chosen by the

product line expert such as user region (to deduce the

weather). In our approach, we used a smart home

DSPL (FM in Figure1 as a running example, based on

an analysis of data that inﬂuence the recommended

result, we have chosen the user roles (worker, home

occupant, student, pets owner..) and the user region

as contextual information.

4.2.1 Product Recommendation Approach

To recommend products based on user’s core require-

ments with context awareness, we propose an ap-

proach using deep neural networks. It mainly con-

sists of three components, including Features Extrac-

tion of the Core requirements and the contextual data,

feature merge and product’s features generation. Fig-

ure 4 shows the overall framework of product recom-

mendation. In the ﬁrst stage, product’s core require-

ments and contextual information are embedded us-

ing Roberta model. In the second stage, the stacked

2-D CNN and Bi-LSTM are then used to extract core

requirements Features for the embedded ones. After

getting the vectors embedding of the user roles and its

region from Roberta output, we use a fully connected

neural network for feature extraction. After retrieving

the feature vectors of product’s core requirements h1,

user role h2 and user region h3, we merge them into

a uniﬁed feature. Finally, the task layers do the ﬁ-

nal features recommendation based a fully connected

feed-forward neural network, which inputs the high-

level representation from the feature extraction layers

(the merged features) and outputs a set of probabilities

assigned to each feature of the product conﬁguration

that determines whether it is selected or not.

4.2.2 Core Requirements Embedding and

Feature Extraction

In order to embed core requirements, Roberta Model

is applied. The embedding layer takes a product’s

core requirements extracted from a product’s require-

ments given by a user as an input, where each word

is transformed into a n-dimensional vector according

to the contexts of words. This matrix keeps the or-

der and representation of each word, further feature

extraction can be done based on this matrix. The em-

bedding layer of product’s core requirements outputs

a nLen (nLen is the max length of inputs) by n de-

scription matrix e for the next layer. We assume that

x1 is a set of product’s core requirements. The em-

bedding of product’s core requirements froberta can

be deﬁned as :

e = f roberta(x1) (1)

To extract feature of a product’s core requirements,

the stacked 2-D CNN and Bi-LSTM are adopted, with

f2-D CNN is a deep convolutional model, which uses

the trained ﬁlters to extract local features between

adjacent words and f-Bi-lstm is a deep sequence

model, which extract sequential features through

time-Therefore, the output feature h1 of product’s

core requirements is:

h1 = f Bi −lstm ∗ f 2 − DCNN(e) (2)

To prevent the model from over-ﬁtting on the train-

ing, dropout layer is used. Furthermore, for faster

convergence, an activation function ReLU is used to

improve the model learning efﬁciency.

4.2.3 Contextual Information Embedding and

Feature Extraction

In this step, contextual information chosen by the ex-

pert are embedding and its features are extracted. In

our running example (smart home DSPL), user roles

and user region are chosen as contextual informa-

tion since they inﬂuence the choice of the recom-

mended features and consequently the user’s satis-

faction. Firstly, the user roles are embedded using

Roberta Model. Since the user role usually contains

a few words, the further complex feature extraction

is not necessary. After getting the embedding of user

role from Roberta output, we use a fully connected

neural network ffull with activation function tanh for

feature extraction. We assume that x2 is the user roles

associated with a given core requirements. The em-

bedding and feature extraction of a user roles can be

deﬁned as:

h2 = f f ull ∗ f roberta(x2) (3)

Toward a Deep Contextual Product Recommendation for SO-DSPL Framework

143

Figure 3: Product recommendation approach.

The same process is applied to the user region. In the

ﬁrst step, Roberta Model is used to embed user region.

After that, a fully connected neural network ffull with

activation function tanh for is applied to the Roberta

output for feature extraction. We assume that x3 is the

user region associated with a given core requirements.

The embedding and feature extraction of a user region

can be deﬁned as:

h3 = f f ull ∗ f roberta(x3) (4)

4.2.4 Feature Merge and Features

Recommendation

After retrieving the feature vectors of product’s core

requirements h1, user roles h2 and user region h3, we

merge them into a uniﬁed feature while the lengths of

those features are the same, we adopt simple vector

addition for feature merging. The output of feature

merging is:

h = h1 + h2 + h3 (5)

After feature merging, the output result is used by the

task layers to do the ﬁnal product’s features recom-

mendation. It contains a fully connected feed-forward

neural network ffc, which inputs high-level represen-

tation h from feature extraction layers and outputs a

recommended list of product’s features l:

l = f f c(h) (6)

ffc contains three fully connected layers with ac-

tivation function σ1,σ2 and σ3. Each layer computes

ai+1 by the weight Wi, bias bi, and the output ai from

the previous layer:

ai + 1 = σ(Wi ai + bi) (7)

where σ1 and σ2 are the tanh function and σ3 is

the sigmoid function that computes the probability of

each feature.

In short, our proposed model takes product’s core

requirements x1, user roles x2 and user region x3 as

input, and outputs the vector l that denotes the proba-

bility of the selection of each feature of the DSPL to

be recommended:

l = f f c ∗ ( f Bi − lstm ∗ f 2 − DCNN ∗ f roberta(x1)+

f f ull ∗ f roberta(x2) + f f ull ∗ f roberta(x3))

(8)

4.2.5 Hyper-Parameters

The hyper-parameters of our proposed model contains

the network conﬁguration and training setting.

1. Network hyper-parameters: We choose the

pre-trained ROBERTA model (Roberta-base)

from Transformers (rob, ), which transforms each

word of the input into a 768-dimension vector.

Our proposed model contains two convolution

layers and one bidirectional LSTM layer. The ﬁrst

convolution layer has 64 ﬁlters with kernel size

9 by 9, the second convolution layer has 32 ﬁl-

ter with kernel size 3 by 3. The hidden state of

LSTM is a 1024-dimension vector. The task layer

contains 34 nodes with an activation function sig-

moid to compute the probabilities of each DSPL

feature. To avoid over-ﬁtting, we add a dropout

layer between every two layers of our proposed

model.

2. Training hyper-parameters: our proposed recom-

mendation model adopts the binary cross-entropy

as the loss function. The Adam optimization algo-

rithm is used for training with the learning rate 5e-

5, beta1 0.9, beta2 0.999, epsilon=1e-6, and learn-

ing decay 0.01.

Xavier normal initializer is used to initialize the

kernel parameters.The total epoch number is 40

with batch size 32. The hidden state of Bi-LSTM

and all bias are initialized to zero.

ENASE 2023 - 18th International Conference on Evaluation of Novel Approaches to Software Engineering

144

4.3 Recommended Product Consistency

Checking

In order to ensure the validity of the recommended

product, we check its consistency with respect to SO-

DSPL constraints, we use SWRL rules presented in

(Maalaoui et al., 2021). In the ﬁrst step, the rec-

ommended product populate the ontology conceptual-

ized for SO-DSPL framework (Maalaoui et al., 2021),

then swrl rules are executed. If the product is valid (no

violation constraint is detected), it will be taken as the

ﬁnal recommended product, else if the recommended

product is not valid (violation constraint is detected),

the appropriate swrl rules are triggered to correct the

product by activating the valid features and removing

the invalid ones. We present in Table1 an extract of

the proposed SWRL rules.

5 EVALUATION

In this section, we introduce metrics, and the evalu-

ation results of our proposed approach for product’s

features recommendation, which includes the com-

parison results with other machine learning methods.

To prepare dataset, we have extended the smart home

requirements data set (dat, )on a larger volume of

data, we have augmented the existed dataset using

data augmentation algorithms (Murukannaiah et al.,

2016), which consists in altering an existing data to

create a new one. The objective is to augment the

dataset by generating new product requirements with

the same meaning as the existing requirements but

written in another form. Thus, we have used “nlpaug”

libraries using the Substitution by contextual word

embeddings RoBERTA technique. This data collec-

tion consists of textual product requirements. We

extend the data with user context information based

context elements presented in the “User context sub-

ontology” and the “DSPL sub-ontology” (Maalaoui

et al., 2021). Indeed, we associated to each prod-

uct requirements the contextual information of its user

that are basically: user interests, user roles, user skills,

user preferences and personal information such as its

region.

5.1 Competitive Methods

In order to show the feasibility and effectiveness of

our approach, we carried out experiments and com-

pared with different ML approaches and using various

embedding methods.

We compare 6 combination of machine learning

methods from conventional machine learning meth-

ods to deep learning models for product’s features

recommendation on the product’s core requirements

dataset, which include CNN , LSTM, BI-LSTM and

using three embedding methods that include Bert,

Roberta and FASTtext. All the conventional mod-

els are implemented by scikit-learn library4 and all

the deep learning models are implemented by Tensor-

Flow. All the models are trained and tested on Google

Colab (col, ).

5.2 Evaluation Metrics

In order to measure the accuracy of our proposed

product’s features recommendation approach among

different approaches, Recall and MAP (Mean Av-

erage Precision) are used as the evaluation metrics.

Each evaluation indicator is described as below.

(1) Recall. It refers to the ratio of the correct number

of the recommended product’s features in the Prod-

uct’s recommendation list to the total number of the

relevant product’s features which is deﬁned as:

Recall =

{

p f eatures

}

| ∩

{

recomended f eature

}

{

p f eatures

}

Where pfeatures is the real set of features of

a product itself, and recomendedfeature is a set of

predicted features recommended by a method. For

example, given a bloc of core requirements crs asso-

ciated to a product p and its corresponding features

are f1; f2; f3, if the recommended features are f1; f3;

f5, then the recall of features recommendation is:

{

f 1, f 3

}

{

f 1, f 2, f 3

}

=0,667

Recall@n refers to the rate when the total number

of features of a product itself |

{

p f eatures

}

|= n.

That is, the ﬁrst n features are used to calculate the

recall rate. Here, recall measures the degree of the

recommendation results that cover all of the Product

features.

(2) MAP. It is the expected average of the precision

that is the ratio of the correct number of predicted

features in the recommendation list, which is deﬁned

as:

MAP@n =

∑

u=1

∑

K=1

P(K) ∗ Rel(K) (9)

where where P(k) is the precision value at the kth

recommendation, rel(k), is just an indicator that

says whether that kth recommendation was relevant

(rel(k)=1) or not (rel(k)=0) and U is the number of

Toward a Deep Contextual Product Recommendation for SO-DSPL Framework

145

Table 1: Extract of the SO-DSPL SWRL rules.

ID SWRL Rule Description

dspl:conﬁguration(?cf)∧ composed-of(?cf,f1)∧

recommended-with(?f1,?f2) -> composed-of(?cf,?f2)

if a feature is selected in a conﬁg-

uration and it recommends another

feature (with the semantic relation

”recommended-with” then the rec-

ommended feature is with be se-

lected in the running conﬁguration.

swrlx:makeOWLThing(?S, ?y)∧ dspl:Feature(?y)) ->

implements(?y, ?S))∧ ws:Service(?S)

For each dspl feature an individ-

ual service is created and related by

the relationship “implements” to ex-

ecute its functionalities.

swrlx:makeOWLThing(?f, ?c)∧ dspl:Feature(?F) ->

composed-of(?f, ?c)∧ Conﬁguration(?c)

each conﬁguration must be com-

posed by more than one feature.

uc:CoreRequirement(?CR1)∧ uc:CoreRequirement(?CR2)

∧ uc:ProductRequirement(?PR)

∧ uc:composed-of(?PR,?CR1) ∧

uc:composed-of(?PR,?CR2) ∧ has-priority(?CR1, Desirable) ∧

has-priority(?CR2, Essential) ∧ dspl:Feature(?F1)

∧ dspl:Feature(?F2) ∧ dspl:alternative-with(?F2,?F1)

∧ dspl:satisfy(?F1,?CR1)

∧ dspl:satisfy(?F2,?CR2) ∧ dspl:Conﬁguration(?CF)

∧ satisfy(?CF,?PR) ∧ dspl:composed-of(?CF,?F1)

-> dspl:composed-of(?CF,?F2)∧ dspl:selected-for(?CF,?F2)

∧ dspl:eliminated-for(?CF,?F1)

The features associated to a prod-

uct’s core requirements and their

are related with the relationship

”alternative-with” then an adaptation

is triggered to the running conﬁgura-

tion by eliminating the optional fea-

ture and selecting the essential one.

uc:CoreRequirement(?CR1)∧ uc:CoreRequirement(?CR2)

∧ uc:ProductRequirement(?PR)

∧ uc:composed-of(?PR,?CR1) ∧

uc:composed-of(?PR,?CR2) ∧ has-priority(?CR1, Essential) ∧

has-priority(?CR2, Essential) ∧ dspl:Feature(?F1)

∧ dspl:Feature(?F2) ∧ dspl:alternative-with(?F2,?F1)

∧ dspl:satisfy(?F1,?CR1)

∧ dspl:satisfy(?F2,?CR2) ∧ dspl:Conﬁguration(?CF)

∧ satisfy(?CF,?PR) ∧ dspl:composed-of(?CF,?F1)

-> uc:alternative-conﬂict(?CR1,?CR2)

The features associated to a prod-

uct’s core requirements and their

are related with the relationship

”alternative-with” and the two core

requirements are essential then a

conﬂict of alternative constraint vi-

olation is detected.

relevant recommendation, with P(K) is deﬁned as:

P(K)=

{

p f eatures

}

|∩

{

recommended f eature(k)

}

{

}

—

and precision P is deﬁned as:

P =

{

p f eatures

}

|∩

{

recomended f eature

}

{

recomended f eature

}

5.2.1 Experimental Results and Analyses

To test the performance of our proposed approach,

the extensive experiments on product recommenda-

tion are conducted among the mentioned competitive

approaches. Table 2 shows the experimental results

on the different evaluation metrics.

Table 2: Comparison result of machine learning methods.

ML Methods MAP@10 RECALL@10

Roberta+CNN 0,57 0,41

Roberta +LSTM 0,72 0,587

Roberta +BI-LSTM 0,78 0,594

Bert+cnn+lstm 0.857 0,65

FASTTEXT+cnn+lstm 0.753 0,61

Roberta+cnn+lstm 0,88 0,692

Roberta+2-D

cnn+BI-lstm

0,91 0,78

ENASE 2023 - 18th International Conference on Evaluation of Novel Approaches to Software Engineering

146

It can be seen from the experimental results that

our proposed product’s features recommendation ap-

proach is superior to the existing ones among the

evaluation metrics. In the experiments, the CNN

has the lowest performance on MAP@10 and Re-

call@10, that means only considering local feature

between adjacent words is not working for SPL fea-

tures recommendation problem. For sequence mod-

els, they can abstract learn global features such as

long-term dependencies from the past time steps,

which makes LSTM reach 72,34% MAP@10 and

58,79% Recall@10. When stacking sequence model

with 1-D CNN for learning the local features, LSTM

can get higher MAP@10 and Recall@10 88,21% and

69,32%. BI-LSTM can learn the long-term depen-

dencies from the past as well as the future informa-

tion. That makes BI-LSTM reach 78.70% MAP@10

and 59.43% Recall@10. Finally, our proposed model

uses 2-D CNNs with BI-LSTM, it can learn both

local and global features as well as the features in

small regions inside of words. Moreover, it uses

both information from core requirements, user roles

and user region with context-dependent word embed-

ding. Applying three context-dependent word embed-

ding methods with our proposed model, it has a lower

performance with FASTTEXT embedding than with

BERT embedding. While, using Roberta embedding,

our proposed model has the highest performance.

In order to further test the parameter inﬂuence on

the product’s features recommendation, a set of ex-

periments are carried out among four competitive ap-

proaches. The experimental results of parameter tun-

ing are illustrated in Figure4 and Figure5.

Figure 4: The experimental results of MAP@n affected by

the parameter n.

Along with the changes of parameter for the num-

ber of recommended features, Figs. 5 and 6 illus-

trate the experiments results on MAP@n and Re-

Figure 5: The experimental results of Recall@n affected by

the parameter n.

call@n among four competitive approaches. In the

experiments, the parameter of MAP@n and Recall@n

ranges from 1 to 10 and the results are calculated with

different values.

Compared to the existing approaches, Fig. 4

shows that the experimental results of our proposed

approach has better MAP@n along with the changes

of n. Speciﬁcally, as n of MAP@n changes from 1

to 10, the experimental results of each approach have

a certain decrease, and reach the best features rec-

ommendation at MAP@1. Therefore, as the number

of features to be evaluated increases, the predictive

power of each recommendation approaches declines.

However, our self-developed approach has the lowest

decline compared with the other approaches.

6 CONCLUSION AND FUTURE

WORKS

In this paper, we targeted an open research question in

the SO-DSPL conﬁguration domain: How to predict

a suitable set of features based on explicit information

from users? We answer this question by providing an

advanced a contextual product recommender system.

Our system helps user in the product conﬁguration

process by suggesting relevant features according to

its described textual requirements. Our experimental

results show that the proposed approach is very use-

ful as it provides feature predictions that are in accor-

dance with the user context and requirements. Since

there are no other publicly available datasets with real

user requirements and associated conﬁgurations, we

could not test our work on further datasets. Moreover,

we will extend this work to take into account non-

functional properties, adapt our approach to be more

Toward a Deep Contextual Product Recommendation for SO-DSPL Framework

147

efﬁcient with a large number of features and training

our proposed deep learning model on more dynamic

adaptation cases that can be triggered at runtime.

REFERENCES

Google colab. https://colab.research.google.com/. Ac-

cessed: 2022-05-15.

nlpaug. https://github.com/makcedward/nlpaug. Accessed:

2022-05-15.

Roberta. https://huggingface.co/roberta-base. Accessed:

2022-05-15.

Bagheri, E., Asadi, M., Gasevic, D., and Soltani, S. (2010a).

Stratiﬁed analytic hierarchy process: Prioritization

and selection of software features. In Bosch, J. and

Lee, J., editors, Software Product Lines: Going Be-

yond, pages 300–315, Berlin, Heidelberg. Springer

Berlin Heidelberg.

Bagheri, E., Di Noia, T., Ragone, A., and Gasevic, D.

(2010b). Conﬁguring software product line feature

models based on stakeholders’ soft and hard require-

ments. In Proceedings of the 14th International Con-

ference on Software Product Lines: Going Beyond,

SPLC’10, page 16–31, Berlin, Heidelberg. Springer-

Verlag.

Bashari, M., Bagheri, E., and W.Du (2017). Dynamic soft-

ware product line engineering: A reference frame-

work. International Journal of Software Engineering

and Knowledge Engineering, pages 191–234.

Capilla, R., Bosch, J., Trinidad, P., Ruiz-Cortes, A., and

Hinchey, M. (2014). Overview of dynamic software

product line architectures and techniques:observations

from research and industry. The Journal of Systems

and Software, pages 3–23.

Czarnecki, K., Gr

unbacher, P., Schmid, R. R. K., and

Wasowski, A. (2012). Cool features and tough

decisions:a comparison of variability modeling ap-

proaches. In Proceedings of the 6th Interna-

tional Workshop onVariability Modelling of Software-

Intensive Systems (VaMoS’12), pages 173–182. ACM.

Jonathan, Y., James, T., and Audrey, T. (2005). Evaluat-

ing ontology criteria for requirements in a geographic

travel domain. In Meersman, Robert, Tari, and Za-

hir, editors, On the Move to Meaningful Internet Sys-

tems 2005: CoopIS, DOA, and ODBASE, pages 1517–

1534, Berlin, Heidelberg. Springer Berlin Heidelberg.

Kang, K., Cohen, S., Hess, J., Novak, W., and Peterson,

A. (1990). Feature-oriented domain analysis (foda)

feasibility study. Rep. CMU/SEI-, 90.

Katarya, R. and Arora, Y. (2020). Capsmf: a novel product

recommender system using deep learning based text

analysis model. Multimedia Tools and Applications,

pages 1–22.

Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D.,

Levy, O., Lewis, M., Zettlemoyer, L., and Stoyanov,

V. (2019). Roberta: A robustly optimized BERT pre-

training approach. CoRR, abs/1907.11692.

Maalaoui, N., Beltaifa, R., and Labed Jilani, L. (2022). De-

riving service-oriented dynamic product lines knowl-

edge from informal user-requirements: Ai based ap-

proach. pages 58–68. ICSEA 2022.

Maalaoui, N., Beltaifa, R., Labed Jilani, L., and Mazo, R.

(2021). An ontology for service-oriented dynamic

software product lines knowledge management. In

ENASE.

Martinez, J., Rossi, G., Ziadi, T., Bissyand

e, T. F. D. A.,

Klein, J., and Le Traon, Y. (2015a). Estimating and

predicting average likability on computer-generated

artwork variants. In Proceedings of the Companion

Publication of the 2015 Annual Conference on Ge-

netic and Evolutionary Computation, GECCO Com-

panion ’15, page 1431–1432, New York, NY, USA.

Association for Computing Machinery.

Martinez, J., Rossi, G., Ziadi, T., Bissyand

e, T. F. D. A.,

Klein, J., and Le Traon, Y. (2015b). Estimating and

predicting average likability on computer-generated

artwork variants. In Proceedings of the Companion

Publication of the 2015 Annual Conference on Ge-

netic and Evolutionary Computation, GECCO Com-

panion ’15, page 1431–1432, New York, NY, USA.

Association for Computing Machinery.

Mazo, R., Dumitrescu, C., Salinesi, C., and Diaz, D. (2014).

Recommendation Heuristics for Improving Product

Line Conﬁguration Processes. In M, R., W., M., R.,

W., and T., Z., editors, Recommendation Systems in

Software Engineering, page 100. Springer.

Murukannaiah, P. K., Ajmeri, N., and Singh, M. P. (2016).

Acquiring creative requirements from the crowd: Un-

derstanding the inﬂuences of personality and creative

potential in crowd re. 2016 IEEE 24th International

Requirements Engineering Conference (RE), pages

176–185.

Pereira, J. A., Matuszyk, P., Krieter, S., Spiliopoulou, M.,

and Saake, G. (2016). A feature-based personalized

recommender system for product-line conﬁguration.

SIGPLAN Not., 52(3):120–131.

Ripon, S., Rasel, F. S., Howlader, R. K., and Islam,

M. (2020). Automated Requirements Extraction and

Product Conﬁguration Veriﬁcation for Software Prod-

uct Line, pages 27–51. Springer Singapore, Singa-

pore.

Sengar, N., Singh, A., and Yadav, V. (2021). Classiﬁca-

tion of documents using bidirectional long short-term

memory recurrent neural network. In Reddy, V. S.,

Prasad, V. K., Wang, J., and Reddy, K. T. V., editors,

Soft Computing and Signal Processing, pages 149–

156, Singapore. Springer Singapore.

ENASE 2023 - 18th International Conference on Evaluation of Novel Approaches to Software Engineering

148