Obsolescence Prediction based on Joint Feature Selection and Machine

Learning Techniques

Imen Trabelsi

1,2

, Besma Zeddini

, Marc Zolghadri

1,4

, Maher Barkallah

and Mohamed Haddar

Quartz Laboratory, SUPMECA, 3 rue Fernand Hainaut, 93407 Saint-Ouen, France

LA2MP Laboratory, ENIS, Route Soukra Km 3.5, 3038 Sfax, Tunisia

SATIE Laboratory CNRS, UMR 8029, CYTeh ENS Paris-Saclay, Cergy, France

LAAS - CNRS, 7 Avenue du Colonel Roche, 31400 Toulouse, France

Keywords:

Obsolescence Prediction, Artiﬁcial Intelligence, Machine Learning, Feature Selection.

Abstract:

Obsolescence is a serious phenomenon that affects all systems. To reduce its impacts, a well-structured man-

agement method is essential. In the ﬁeld of obsolescence management, there is a great need for a method

to predict the occurrence of obsolescence. This article reviews obsolescence forecasting methodologies and

presents an obsolescence prediction methodology based on machine learning. The model developed is based

on joint a machine learning (ML) technique and feature selection. A feature selection method is applied to

reduce the number of inputs used to train the ML technique. A comparative study of the different methods of

feature selection is established in order to ﬁnd the best in terms of precision. The proposed method is tested

by simulation on models of mobile phones. Consequently, the use of features selection method in conjunction

with ML algorithm surpasses the use of ML algorithm alone.

1 INTRODUCTION

Obsolescence is a problem that affects all sectors. It

is not a new phenomenon; since the early 1990s, the

rate of component obsolescence has increased rapidly.

This is particularly well illustrated by electronic com-

ponents, especially for smartphones. Smartphones

are the devices most subject to rapid renewal due

to their obsolescence, whether technical, software or

aesthetic. For example, when manufacturers still of-

fer new versions or updates that are incompatible with

previous models, we talk about software and aesthetic

obsolescence. But when battery usage is reduced due

to a small number of cycles or when repair becomes

increasingly complicated (with models that are almost

impossible to disassemble and spare parts unavail-

able), we speak of technical obsolescence.

Indeed, obsolescence is inevitable but anticipation

and careful planning can minimize its impact and po-

tentially high cost. The aim of obsolescence man-

agement is to ensure that obsolescence is managed

as an integral part of design, development, production

and maintenance in order to minimize costs and nega-

tive impact throughout the product life cycle (Group,

2016). Thus, the purpose of obsolescence manage-

ment is to determine: the optimal dates and quan-

tity of last time to buy (LTB), the optimal date for

redesign, the components that should be considered

for redesign or that should be replaced (Meng et al.,

2014).

Sandborn has deﬁned three terms for obsolescence

management as follows (Sandborn, 2013):

• Reactive management consists in taking actions

when the obsolescence has already occurred.

• Proactive management is implemented for critical

components that have a risk of going obsolete.

• Strategic management is done in addition to

proactive and reactive management and involves

the determination of the optimum mix of mitiga-

tion approaches and design refreshes.

The most common type of management used is reac-

tive management because it is easier to implement.

It is advisable to use it only if the cost associated

with the obsolescence of a component is low (Pingle,

2015). However, if the probability of obsolescence

and associated costs are high, it is recommended to

apply proactive management strategies to minimize

the risk of obsolescence and associated costs (Rojo

et al., 2010). In fact, forecasting the occurrence of

obsolescence is the key factor in proactive manage-

ment (Sandborn et al., 2011). For this reason, many

Trabelsi, I., Zeddini, B., Zolghadri, M., Barkallah, M. and Haddar, M.

Obsolescence Prediction based on Joint Feature Selection and Machine Learning Techniques.

DOI: 10.5220/0010241407870794

In Proceedings of the 13th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2021) - Volume 2, pages 787-794

ISBN: 978-989-758-484-8

787

researchers have focused on the development of meth-

ods based on the prediction of obsolescence (Solomon

et al., 2000; Ma and Kim, 2017; Sandborn, 2017; Tra-

belsi et al., 2020).

Recently, with the emergence of machine learning

techniques, some works have been published present-

ing methods based on machine learning for the pre-

diction of obsolescence (Jennings et al., 2016; Grichi

et al., 2017). In these researches, the authors mainly

used the random forest algorithm to predict the state

(obsolete or available) of the product. In the same

line, this work highlights the value of the ML tech-

niques to improve the quality of the obsolescence pre-

diction. In fact, the obsolescence depends on the fea-

tures of the product, but not all the features have the

same effect. For this, a feature selection method must

be used before processing the prediction approach in

aim to select the most important and signiﬁcant fea-

tures. In this case, the following challenges must be

overcome:

- What are the main algorithms that are best

adapted to select the most important features?

- How does the feature selection method improve

the obsolescence prediction model?

The remain of this paper is organized as follow: Sec-

tion 2 presents the related works on forecasting tech-

niques. In section 3, we present the proposed machine

learning-based methodology and the feature selection

procedure used to obsolescence prediction. Therefore

an explanation of the case study and results are pre-

sented in section 4. Finally, some conclusions and

future work are given in section 5.

2 RELATED WORKS

This section presents an overview of the obsolescence

concept in the ﬁrst part. Then, we present the prob-

lem of feature selection and different methods used

for supervised learning. Supervised learning is where

the computer is equipped with sample inputs that are

labeled with the desired outputs (Brownlee, 2016).

2.1 Obsolescence Concept

According to IEC 62402, obsolescence is ”the transi-

tion of an item from available to unavailable from the

manufacturer in accordance with the original speci-

ﬁcation”. Several factors are responsible for product

obsolescence, including technological evolution or in-

novation, government-imposed laws and regulations,

market demand, etc.(Bartels et al., 2012). Obsoles-

cence is considered as a change that may affect the

product. It can be considered as voluntary or involun-

tary. When the change is made by the manufacturer

itself to promote new products, increase market share

and sales (D

e and Marchand, 2015), or when the

customers decide to stop using the product for rea-

sons, such as economics (for example, when the cost

of maintenance is higher than the purchase of a new

one), or aesthetics, obsolescence is considered volun-

tary. While, when the change made to the product

is independent of the customer or manufacturer, such

as government imposed regulations, the obsolescence

is considered involuntary, (Bartels et al., 2012). In

the scientiﬁc literature (Bartels et al., 2012; Sandborn,

2007; Mellal, 2020), several obsolescence typologies

have been identiﬁed, including the following:

• Technological Obsolescence: occurs when there

is a new technology that can replace the old one.

• Functional Obsolescence: is related to a techni-

cal defect that makes the product unusable.

• Aesthetic Obsolescence: is related to fashion ef-

fects and consumer psychology.

• Logistical Obsolescence: means that the product

is no longer procurable due to diminishing manu-

facturing sources and material shortages.

• Economic Obsolescence: is related to the high

cost of using, repairing and/or maintaining the

product.

The obsolescence process ideally goes through dif-

ferent phases. The manufacturer announces the end

of life of the product (Product Discontinued Notiﬁ-

cation) and sets a date by which the ofﬁcially obso-

lete product will no longer be sold (Last Time Buy).

During the intermediate phase (Phase Out), customers

can still stock up and build up a stock of the obsolete

product. The increasing rate of obsolescence leads to

several risks. In (Mellal, 2020), the author discussed

the various risks causing by the obsolescence in dif-

ferent sectors.

To reduce the affects of obsolescence, many re-

searchers have already worked on forecasting obso-

lescence. The existing methods can be classiﬁed into

two categories: mathematical-based approaches and

machine learning-based approaches. In this paper, we

will focus on the ML-based approach developed by

(Jennings et al., 2016). The authors are presented

a ML-based method to predict the status (Available/

Discontinued) of the product. A comparative study of

ML techniques was also developed in this work. The

dataset used to illustrate this approach contains infor-

mation about the launch date, some technical features

(chosen by the authors), and the status of the smart-

phone. As discussed above, product obsolescence de-

pends on the most relevant features. Therefore, one of

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

788

the important points is to determine the most relevant

features for the classiﬁcation of these products.

2.2 Feature Selection

The problem of feature selection has become increas-

ingly important in the ﬁeld of data analysis and ma-

chine learning (Samb et al., 2012). The feature se-

lection is a technique that selects among original

features a subset of the most important and signif-

icant attributes. In the context of supervised learn-

ing, this subset should allow meeting the target pur-

pose, namely the accuracy of learning, the speed of

learning, or the applicability of the proposed model

(Khalid et al., 2014). According to (Babiker et al.,

2019), the feature selection methods can be divided

in three main categories: Filter, Wrapper, and Em-

bedded method:

1. Wrapper Method: evaluates a subset of features

by its classiﬁcation performance using a learning

algorithm (Kohavi et al., 1997). The learning al-

gorithm works on the totality of instances with

different subsets of features. It provides for each

of them the estimated precision of the classiﬁca-

tion of the new instances. The subset inducing the

most accurate classiﬁer is selected. The complex-

ity of the learning algorithm makes the wrapper

method very expensive in terms of time where it

becomes impractical when the number of initial

features is large.

2. Filter Method: is a pre-processing method that

is done before the learning process and indepen-

dently of any machine learning algorithms. In-

stead, the relevance of the feature to the outcome

variable is determined based on the feature’s score

in various statistical tests. Compared to wrapper

methods, ﬁlter methods are much faster because

they do not involve training the models.

3. Embedded Method: is performed by a speciﬁc

learning algorithm that performs feature selection

during the training process. It differs from other

methods in the way feature selection and learning

interact.

Figure 1 illustrates these three categories of feature

selection.

In this paper, a comparative study will be done to

choose the best algorithm for feature selection of Fil-

ter, Wrapper, and Embedded types. To make the com-

parison feasible, we propose to identify the subset of

features selected by automatic methods and then us-

ing it as the training data for the predictive model. In

section 3, we deﬁne precisely the steps of the used

method.

Figure 1: Comparison between (a) ﬁlter method, (b) wrap-

per method, and (c) embedded method for feature selection

(Lee, 2009).

3 PROPOSED APPROACH

More and more companies have large amounts of data

that are valuable resources for obsolescence manage-

ment. However, as these resources cannot be suf-

ﬁciently analyzed and evaluated, they are worthless

for the company. To overcome this problem, ma-

chine learning techniques are being developed. Some

researchers have focused on the application of ML

techniques for the prediction and detection of obso-

lescence (Jennings et al., 2016; Grichi et al., 2017).

To update the approach proposed by (Jennings et al.,

2016), in this work we propose an obsolescence pre-

diction based on joint feature selection and ML Tech-

nique.

As shown in Figure 2, the ﬁrst step is to collect the

data to identify the product’s obsolescence. The col-

lected data must contain information about technical

speciﬁcations, launch date, production end date, and

all other data may affect the product obsolescence. To

obtain more accurate forecast results, the strengths of

the feature selection method and the prediction model

will be uniﬁed. Therefore, a crucial step concerning

the feature selection will have to be done before start-

ing the prediction model. At this stage, a subset of the

most relevant features is selected. The third step is

to choose a predictive model among supervised learn-

ing techniques. The predictive model aims to predict

the status (available or obsolete) of the product in the

test dataset. The performance of predictive models is

evaluated using the confusion matrix (Zemouri et al.,

2018). This matrix contains a summary of the number

of correct and incorrect predictions allowing to quan-

tify errors made. In our case, the confusion matrix is

shown in Table 1.

Based on this information, three metrics are cal-

culated:

1. Accuracy: it generally indicates how ”right” pre-

dictions are. It is calculated as follow:

Accuracy =

T P + T N

T P + FP + T N + FN

(1)

Obsolescence Prediction based on Joint Feature Selection and Machine Learning Techniques

789

Figure 2: The proposed approach.

Table 1: The confusion matrix.

Predicted class

Obsolete Available

Actual

class

Obsolete

True Positive

(TP)

False Negative

(FN)

Available

False Positive

(FP)

True Negative

(TN)

1. Precision: it measures the class agreement of the

data labels with the positive labels given by the

classiﬁer.

Precision =

T P

T P + FP

(2)

2. Sensitivity or true positive rate (TPR): it mea-

sures how often the model chooses the positive

class when the observation is in fact in the posi-

tive class.

T PR =

T P

T P + FN

(3)

These metrics are calculated for different ML-

techniques to compare them and choose the best one.

4 CASE STUDY

The case study will demonstrate the utility of the fea-

ture selection method in the predictive model. The

smartphone market is evolving very fast, driven by

the regular introduction of new technologies and se-

rious competitors, and by fashion effects. Therefore,

smartphones are used as the illustrative case of the

proposed approach.

4.1 The Dataset

The database contains speciﬁc information on several

smartphone models and whether their status is avail-

able or discontinued. There are more than 59 features

such as digital mobile phone technology, battery ca-

pacity, smartphone dimensions, display characteris-

tics, operation system, etc. The data was collected on

one of the most popular mobile phone forums (GSM

Arena)

. Figure 3 shows an example of iPhone 6 Plus

speciﬁcations.

Figure 3: Example of technical speciﬁcation from GSM-

Arena.

According to estimates by Canalys

, the global

smartphone market was mainly divided between the

Apple, Samsung, Huawei, Xiaomi and Oppo brands.

Therefore, the number of instances is reduced to 1257

models of the 5 brands divided into 576 available

smartphones and 681 as discontinued (or obsolete).

As is well known in the machine learning commu-

nity, reducing input variables is a useful operation. It

has a great impact on the computer’s time and accu-

racy. Therefore, there is a great need to apply an al-

gorithm to select among these most relevant features.

The data present in the site can miss values and

even have erroneous information. These gaps reﬂect

the limitations imposed by some industries to have a

complete database. To remedy this problem, the data

must ﬁrst be prepared.

The dataset has been formatted in a machine learn-

ing compatible format. To handle missing data, the

mean imputation method is used. Once the data is

formatted, feature selection methods is applied. The

ﬁrst part of this case study is to compare the differ-

ent methods of feature selection. To this end, we have

used several feature selection algorithms for the three

methods.

https://www.gsmarena.com

https://www.canalys.com/newsroom/

canalys-global-smartphone-market-q4-2019

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

790

4.2 Used Tools

In this section, the used techniques for feature selec-

tion and for machine learning are presented. For su-

pervised learning the most used techniques are:

• Random Forest (RF): is a meta-estimator that

applies many decision tree models to different

subsamples of the data set and the end result is ob-

tained as an average of these models, see (Grichi

et al., 2017). The averaging process avoids over-

ﬁtting and therefore improves the accuracy of the

classiﬁcation.

• Artiﬁcial Neural Network (ANN): is a super-

vised machine learning tool that can learn a non-

linear function. It can be used for classiﬁcation

or regression which make it one of the most used

fault detection technique (Zemouri et al., 2019).

• Support Vector Machine (SVM): aims to ﬁnd a

separating hyperplane that separates the different

classes. To do this, the Optimal Separating Hyper-

plane (OSH) is deﬁned as the hyperplane situated

between the classes that maximize the margin be-

tween them (Wang et al., 2012).

• K-Nearest Neighbors (KNN): The nearest

neighbors approach is based on ﬁnding a ﬁxed

number k of samples from a training dataset,

which are the closest ones, in terms of distance, to

the new instance to predict its label, (Omri et al.,

2020).

• Naive Bayes (NB): The Naive Bayes (NB)

method is a simple, probabilistic, and supervised

classiﬁer, see (Rish et al., 2001). This technique

is based on coupling the Bayes theorem with the

Naive hypothesis of conditional independence be-

tween every pair of features given the value of the

class variable.

The used feature selection algorithms in this case

study are presented as follow. For ﬁlter method, we

used:

• Correlation-based Feature Subset Selection is an

algorithm that evaluates the importance of a sub-

set of attributes by considering the individual pre-

dictive ability of each feature along with the de-

gree of redundancy between them (Li et al., 2011).

• Variance Threshold is a simple baseline algorithm

to feature selection. According to (Zhao et al.,

2013), it aims to remove the features whose vari-

ance does not meet some threshold. It automati-

cally removes all features with zero variance, i.e.

features that have the same value in all samples.

For wrapper method we used:

• Wrapper attribute subset evaluator designed to

evaluate features subsets by using a learning tech-

nique. It uses cross validation for estimating the

precision of the learning technique used (Gutlein

et al., 2009).

• Backward Feature Elimination is a feature selec-

tion technique while building a machine learning

model that used to remove features without signif-

icant effect on the prediction of output (Kostrzewa

and Brzeski, 2017).

For embedded method we used:

• Ridge regularization is a process that consider all

the features into the model and try to regularize

the coefﬁcient estimates of a feature such that a

large number of coefﬁcient estimates shrink to-

wards zero (Cawley, 2008).

• Lasso allows to select a restricted subset of fea-

tures by shrinking some of the coefﬁcients to zero,

that means that a certain features will be multi-

plied by zero to estimate the target (Muthukrish-

nan and Rohini, 2016).

The different simulations of feature selection and

classiﬁcation with machine learning were done using

Python libraries.

4.3 Results and Discussion

Here in all the instances, we have the target which

takes 0 or 1 respectively when the smartphone is ob-

solete the output is 0, if it is available the output is

1. Table 2, represents the precision and the sensitiv-

ity of the classiﬁcation algorithms for different feature

selection techniques. A cross validation has been per-

formed with k=5. This allows to evaluate the impact

of the selected features on the model performance.

These results show that the Random Forest (RF)

technique has better accuracy than other algorithms

without the feature selection step. This accuracy is

improved by applying feature selection techniques

such as wrapper methods, as shown in Figure4.

Figure 4: The comparison of the Random Forest accuracy

with different feature selection techniques.

Obsolescence Prediction based on Joint Feature Selection and Machine Learning Techniques

791

Table 2: Simulation results.

Feature selection Performance RF ANN NB KNN SVM

Without feature selection

Accuracy 91.02±0.34 82.39±2.5 85.20 82.26 54.18

Precision 93.41±0.35 80.05±5.48 81.69 86.35 54.18

Sensitivity 89.62±0.35 86.55±2.46 93.69 79.88 100.00

Filter

Cfs Subset Eval

Accuracy 91.19±0.2 88.26±3.88 89.98 89.98 89.82

Precision 93.83±0.16 90.03±3.69 93.29 92.50 94.38

Sensitivity 89.62±0.35 88.34±7.13 87.81 88.69 86.34

Variance Threshold

Accuracy 90.77±0.3 80.78±2.96 80.35 82.18 55.93

Precision 93.24±0.24 84.67±3.07 75.23 84.05 55.37

Sensitivity 89.44±0.51 79.02±6.25 95.01 82.82 96.18

Wrapper

Backward Feature Elimination

Accuracy 91.36±0.23 88.88±0.77 89.02 82.74 91.09

Precision 93.38±0.33 90.35±1.07 88.95 85.91 94.52

Sensitivity 90.47±0.34 89±1.12 91.04 81.50 88.69

Wrapper subset eval

Accuracy 91.62±0.37 84.17±3.07 62.21 84.17 54.26

Precision 93.65±0.23 85.8±2.43 58.97 87.77 54.24

Sensitivity 90.69±0.6 84.94±6.5 99.41 82.23 99.56

Embedded

Ridge

Accuracy 91.43±0.35 81.63±5.76 80.90 91.65 91.49

Precision 93.91±0.14 85.94±7.59 76.53 94.44 94.29

Sensitivity 90.02±0.62 80.23±10.82 93.39 89.87 89.72

Lasso

Accuracy 88.1±0.22 82.42±2.68 86.48 82.26 54.18

Precision 88.86±0.29 87.31±2.43 81.43 86.35 54.18

Sensitivity 89.22±0.23 79.15±5.5 97.21 79.88 100.00

Moreover, for the predictive model SVM, the applica-

tion of the ridge technique improved its accuracy up to

91.49%. Whereas without feature selection technique

it has sensitivity 100% that means that the SVM only

can not identify the available smartphone, it consid-

ers that all the smartphones were obsolete. Therefore,

it is more important to use feature selection meth-

ods and ML algorithms than to use only ML algo-

rithms. As shown in Table 2, obsolescence predic-

tion based on backward feature elimination and ANN

Technique is more accurate than using ANN only. As

mentioned above, feature selection technique allows

to select only those features which are necessary.

Comparing all algorithms of feature selection, we

note that backward feature elimination seems to be

the best technique to increase the performance of the

predictive model. Figure 5 shows the improvement of

the accuracy of the different ML-techniques using the

backward feature elimination.

Figure 5: The improvement of the accuracy en applying

Backward Feature Elimination.

As shown in the Table 3, we can see that through

the application of the feature selection technique, the

number of features used to train the predictive model

has been reduced from 59 to 21 maximum. Therefore,

the feature selection technique allows to optimize the

training time and reduce data collection effort.

Table 3: The number of features selected by the different

feature selection techniques.

Feature selection techniques Features number

Filter

CfsSubsetEval 7

VarianceThreshold 16

Wrapper

Backward Feature Elimination + RF 10

Wrapper subset eval + RF 14

Backward Feature Elimination + ANN 10

Wrapper subset eval + ANN 20

Backward Feature Elimination + KNN 10

Wrapper subset eval +KNN 10

Backward Feature Elimination + NB 10

Wrapper subset eval + NB 7

Backward Feature Elimination + SVM 10

Wrapper subset eval + SVM 21

Embedded

Ridge 13

Lasso 16

The ﬁgure 6 illustrates the features selected by the dif-

ferent techniques. The most relevant feature, which is

selected by 11 techniques, corresponds to the launch

date of the smartphone.

From this study, it is proven that there is a small

set of features that impact the smartphone obsoles-

cence. Thus, an efﬁcient obsolescence prediction ap-

proach can be established by controlling these fea-

tures. However, the phenomenon of obsolescence is

a dynamic concept which evolves over time which

means that these features can be modiﬁed depend-

ing on the date of observation. Therefore, each obso-

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

792

Figure 6: The most important features selected by the different feature selection techniques.

lescence prediction must be presented with a validity

horizon. As a perspective for this work, we propose

to study the evolution of the features that impact ob-

solescence for a given system.

In this paper, obsolescence status is considered as

the case where the manufacturer announce the prod-

uct end-of-life. However, for the user, the problem

arises in other words. Indeed, he can decide for him-

self to abandon the use of a product that has become

useless due, for example, to its replacement by a new

product. In this context, an other limitation to our ap-

proach is the behavior of costumers. Obsolescence

status is then modiﬁed according to this issue. Hence,

it is necessary to integrate this customer-related infor-

mation into the study of obsolescence by proposing

an approach based on natural language processing.

5 CONCLUSIONS

In this paper, we have discussed a new joint technique

of feature selection and machine learning used for ob-

solescence prediction. Our approach is applied in a

real data set concerning the obsolescence of smart-

phones. The results show that the obsolescence of the

smartphone is linked to speciﬁc features such as the

launch date, the ability of the smartphone to insert a

memory card slot,etc. Thus, an effective obsolescence

prediction strategy lies in the prevention of obsoles-

cence of these features. However, this work is limited

in the obsolescence detection (binary class) without

taking into account its evolution over time and its de-

pendence on the system environment (market and en-

tities).

To go further, the authors believe that it is necessary

to evaluate obsolescence as a dynamic problem that

evolves over time. Thus, the features which have

an impact on the obsolescence of a system during a

given period can be changed during another period.

In this context, a data-driven obsolescence manage-

ment approach is needed. This approach should bring

together the different data that characterize the sys-

tem environment (technologies, laws, aesthetics, etc)

in order to control adapt the prediction model to these

environment change.

REFERENCES

Babiker, M., Karaarslan, E., and HOS¸CAN, Y. (2019). A

hybrid feature-selection approach for ﬁnding the digi-

tal evidence of web application attacks. Turkish Jour-

nal of Electrical Engineering & Computer Sciences,

27(6):4102–4117.

Bartels, B., Ermel, U., Sandborn, P., and Pecht, M. G.

(2012). Strategies to the prediction, mitigation and

management of product obsolescence, volume 87.

John Wiley & Sons.

Brownlee, J. (2016). Master Machine Learning Algorithms:

discover how they work and implement them from

scratch. Machine Learning Mastery.

Cawley, G. C. (2008). Causal & non-causal feature selec-

tion for ridge regression. In Causation and Prediction

Challenge, pages 107–128.

e, C. and Marchand, A. (2015). L’obsolescence des

produits

electroniques: des responsabilit

es partag

ees.

In Les ateliers de l’

ethique/The Ethics Forum, vol-

ume 10, pages 4–32. Centre de recherche en

ethique

de l’Universit

e de Montr

eal.

Grichi, Y., Beauregard, Y., and Dao, T. (2017). A random

forest method for obsolescence forecasting. In 2017

Obsolescence Prediction based on Joint Feature Selection and Machine Learning Techniques

793

IEEE International Conference on Industrial Engi-

neering and Engineering Management (IEEM), pages

1602–1606. IEEE.

Group, E. (2016). Expert group 21: Obsolescence manage-

ment ﬁnal report.

Gutlein, M., Frank, E., Hall, M., and Karwath, A. (2009).

Large-scale attribute selection using wrappers. In

2009 IEEE symposium on computational intelligence

and data mining, pages 332–339. IEEE.

Jennings, C., Wu, D., and Terpenny, J. (2016). Forecasting

obsolescence risk and product life cycle with machine

learning. IEEE Transactions on Components, Packag-

ing and Manufacturing Technology, 6(9):1428–1439.

Khalid, S., Khalil, T., and Nasreen, S. (2014). A survey of

feature selection and feature extraction techniques in

machine learning. In 2014 Science and Information

Conference, pages 372–378. IEEE.

Kohavi, R., John, G. H., et al. (1997). Wrappers for feature

subset selection. Artiﬁcial intelligence, 97(1-2):273–

324.

Kostrzewa, D. and Brzeski, R. (2017). The data di-

mensionality reduction in the classiﬁcation process

through greedy backward feature elimination. In

International Conference on Man–Machine Interac-

tions, pages 397–407. Springer.

Lee, M.-C. (2009). Using support vector machine with a hy-

brid feature selection method to the stock trend predic-

tion. Expert Systems with Applications, 36(8):10896–

10904.

Li, B., Wang, Q., and Hu, J. (2011). Feature subset selec-

tion: a correlation-based svm ﬁlter approach. IEEJ

Transactions on Electrical and Electronic Engineer-

ing, 6(2):173–179.

Ma, J. and Kim, N. (2017). Electronic part obsolescence

forecasting based on time series modeling. Interna-

tional Journal of Precision Engineering and Manu-

facturing, 18(5):771–777.

Mellal, M. A. (2020). Obsolescence–a review of the litera-

ture. Technology in Society, page 101347.

Meng, X., Th

ornberg, B., and Olsson, L. (2014). Strate-

gic proactive obsolescence management model. IEEE

Transactions on Components, Packaging and Manu-

facturing Technology, 4(6):1099–1108.

Muthukrishnan, R. and Rohini, R. (2016). Lasso: A fea-

ture selection technique in predictive modeling for

machine learning. In 2016 IEEE international confer-

ence on advances in computer applications (ICACA),

pages 18–20. IEEE.

Omri, N., Al Masry, Z., Mairot, N., Giampiccolo, S., and

Zerhouni, N. (2020). Industrial data management

strategy towards an sme-oriented phm. Journal of

Manufacturing Systems, 56:23–36.

Pingle, P. (2015). Selection of obsolescence resolution strat-

egy based on a multi criteria decision model.

Rish, I. et al. (2001). An empirical study of the naive bayes

classiﬁer. In IJCAI 2001 workshop on empirical meth-

ods in artiﬁcial intelligence, volume 3, pages 41–46.

Rojo, F. J. R., Roy, R., and Shehab, E. (2010). Obsoles-

cence management for long-life contracts: state of the

art and future trends. The International Journal of Ad-

vanced Manufacturing Technology, 49(9-12):1235–

1250.

Samb, M. L., Camara, F., Ndiaye, S., Slimani, Y., and

Esseghir, M. A. (2012). Approche de s

election

d’attributs pour la classiﬁcation bas

ee sur l’algorithme

rfe-svm. In 11

eme Colloque Africain sur la Recherche

en Informatique et Math

ematiques.

Sandborn, P. (2007). Software obsolescence-complicating

the part and technology obsolescence management

problem. IEEE Transactions on Components and

Packaging Technologies, 30(4):886–888.

Sandborn, P. (2013). Design for obsolescence risk manage-

ment. Procedia CIRP, 11:15–22.

Sandborn, P. (2017). Forecasting technology and part obso-

lescence. Proceedings of the Institution of Mechanical

Engineers, Part B: Journal of Engineering Manufac-

ture, 231(13):2251–2260.

Sandborn, P., Prabhakar, V., and Ahmad, O. (2011). Fore-

casting electronic part procurement lifetimes to enable

the management of dmsms obsolescence. Microelec-

tronics Reliability, 51(2):392–399.

Solomon, R., Sandborn, P. A., and Pecht, M. G. (2000).

Electronic part life cycle concepts and obsolescence

forecasting. IEEE Transactions on Components and

Packaging Technologies, 23(4):707–717.

Trabelsi, I., Zolghadri, M., Zeddini, B., Barkallah, M., and

Haddar, M. (2020). Fmeca-based risk assessment ap-

proach for proactive obsolescence management. In

IFIP International Conference on Product Lifecycle

Management, pages 215–226. Springer.

Wang, J., Liu, S., Gao, R. X., and Yan, R. (2012). Current

envelope analysis for defect identiﬁcation and diag-

nosis in induction motors. Journal of Manufacturing

Systems, 31(4):380–387.

Zemouri, R., Omri, N., Devalland, C., Arnould, L., Morello,

B., Zerhouni, N., and Fnaiech, F. (2018). Breast can-

cer diagnosis based on joint variable selection and

constructive deep neural network. In 2018 IEEE 4th

Middle East Conference on Biomedical Engineering

(MECBME), pages 159–164. IEEE.

Zemouri, R., Omri, N., Fnaiech, F., Zerhouni, N., and

Fnaiech, N. (2019). A new growing pruning deep

learning neural network algorithm (gp-dlnn). Neural

Computing and Applications, pages 1–17.

Zhao, Z., Zhang, R., Cox, J., Duling, D., and Sarle, W.

(2013). Massively parallel feature selection: an ap-

proach based on variance preservation. Machine

learning, 92(1):195–220.

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

794