Knowledge graph-based method for solutions detection and evaluation in an online problem-solving community

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 12, No. 6, December 2022, pp. 6350~6362
ISSN: 2088-8708, DOI: 10.11591/ijece.v12i6.pp6350-6362  6350
Journal homepage: http://ijece.iaescore.com
Knowledge graph-based method for solutions detection and
evaluation in an online problem-solving community
Houda Sekkal1
, Naïla Amrous2
, Samir Bennani1
1
Laboratory of Research in Computer Science and Education Mohammadia School of Engineers, Mohammed V University Rabat,
Rabat, Morocco
2
Laboratoire Mediation-Information-Knowledge-Society, School of Information Sciences Rabat, Rabat, Morocco
Article Info ABSTRACT
Article history:
Received Jul 9, 2021
Revised Jul 14, 2022
Accepted Aug 10, 2022
Online communities are a real medium for human experiences sharing. They
contain rich knowledge of lived situations and experiences that can be used
to support decision-making process and problem-solving. This work presents
an approach for extracting, representing, and evaluating components of
problem-solving knowledge shared in online communities. Few studies have
tackled the issue of knowledge extraction and its usefulness evaluation in
online communities. In this study, we propose a new approach to detect and
evaluate best solutions to problems discussed by members of online
communities. Our approach is based on knowledge graph technology and
graphs theory enabling the representation of knowledge shared by the
community and facilitating its reuse. Our process of problem-solving
knowledge extraction in online communities (PSKEOC) consists of three
phases: problems and solutions detection and classification, knowledge
graph constitution and finally best solutions evaluation. The experimental
results are compared to the World Health Organization (WHO) model
chapter about Infant and young child feeding and show that our approach
succeed to extract and reveal important problem-solving knowledge
contained in online community’s conversations. Our proposed approach
leads to the construction of an experiential knowledge graph as a
representation of the constructed knowledge base in the community studied
in this paper.
Keywords:
Community mining
Information usefulness
Knowledge extraction
Knowledge graph
Knowledge management
Knowledge representation
Online communities
This is an open access article under the CC BY-SA license.
Corresponding Author:
Houda Sekkal
Laboratory of Research in Computer Science and Education in Mohammadia School of Engineers,
Mohammed V University Rabat
Avenue Ibn Sina B.P 765, Agdal Rabat 10090, Morocco
Email: houda.sekkal@gmail.com
1. INTRODUCTION
Online communities have become one of the principal sources of practical knowledge among users
seeking for problem solving knowledge in the web [1]–[3]. Those communities are a gathering of members
discussing and solving common interest problems and issues in different domains. This makes online
communities room for the creation and gathering of special type of knowledge called problem solving
knowledge. Indeed some researchers argue that online social interactions in online communities facilitate
learning, expertise sharing, problem solving and innovation [4]. Many authors have defined problem solving
knowledge as a way of knowing and understanding through direct engagement and being in the situation [5]
[6] and as a truth learned from personal experience with a phenomenon [7]. Problem solving is a primary
vehicle for better understanding of our environment, learning, and discovery of new opportunities [8]. This

Int J Elec & Comp Eng ISSN: 2088-8708 
Knowledge graph-based method for solutions detection and evaluation in an online … (Houda Sekkal)
6351
type of knowledge is what distinguishes online communities as they contain discussions of users solving
problems of other members through mutual efforts and the presence and testimony of persons who have
solved similar problems [7]. The process of solving issues through online discussions leads to the
accumulation of big amounts of valuable knowledge in the online community. Users need to seek, reuse and
exploit this problem-solving knowledge to benefit from it in different contexts. In [9], Tanis focused on
information seeking in online health communities and found that there are several reasons behind seeking
information in online communities as to look for emotional support, inclusion, support others or pass time.
According to [10], seeking information from peers in online health communities is a new way of pursuing
health by banding together and sharing knowledge. Despite all those information seeking needs, actual
platforms are limited in terms of extracting knowledge from the online communities. According to [11],
existing methods for knowledge and information seeking in online communities lack of relevance and
reliability as they are based only on full-text search or topics search which the process is shown in Figure 1.
This is due to the limited matching terms in the query to the indexed post without consideration for the
context introduced by these terms [11].
Figure 1. Existing methods of information retrieval from OSGs [11]
The full-text search does not enable seekers to rapidly get the answer for their questions or even to
evaluate the usefulness and accuracy of the results. It only returns raw textual data as a result and push the
seeker to spend more time and effort to look for the right answer to his question among all the online
community conversations. Information usefulness as a topic is addressed in the literature from different
perspectives. Generally, most studies discussed information usefulness in the context of business intelligence
[12] and more specifically in business internet services [13]. Evaluating information usefulness shared in
online communities is a critical issue specially in this area of big data and knowledge sharing in social media.
In [14], the author proposed an approach for classifying the usefulness of solutions shared in knowledge
communities. In his method, the author used a text analytic framework to extract important features from
online forums in order to evaluate the information usefulness. In another work, Daradkeh [12] examines the
usefulness of user generated content with large amounts of redundant information in order to analyze the
digital voice of users in online open innovation communities. Generally, used approaches for information
usefulness evaluation are based on extracting many features from online discussions and classify the
usefulness of either threads or posts [14]. Such features concerns for example relevancy, objectivity,
timelessness an completeness of each post [12], [14]. Despite the critical need for evaluating information
usefulness in online communities it is noticeable that few studies have tackled this issue. In our work, we
propose a new approach for evaluating information usefulness based on knowledge graph technology.
According to [15] a knowledge graph is a structured representation of facts, consisting of entities,
relationships, and semantic descriptions. Knowledge graphs enable the representation of knowledge and
reasoning inspired by human problem solving [16]. Knowledge graphs are generally used in many real-world
applications such as recommendation systems and question answering tasks with the ability of commonsense
understanding and reasoning [15]. In our case, we built a process for extracting and evaluating best solutions
to problems in an online community using knowledge graph technology and more specifically link prediction
method used for knowledge graph completion. The goal is to reveal hidden semantic links between entities
enabling to evaluate them. To the best of our knowledge, our work is the first to propose such an approach
for information usefulness evaluation.

 ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 12, No. 6, December 2022: 6350-6362
6352
Finding useful solutions proposed and discussed in online communities is the key goal behind our
research. In our study context, we are interested on extracting useful solutions to issues discussed in an online
health community about breastfeeding and to evaluate the best one according to the member’s experiences.
According to [14] there is a strong motivation to identify useful solutions automatically in online
communities to better serve users’ knowledge needs and bolster the success of these communities. The main
goal of our study is to conceive a process capable of capturing, distilling and evaluating automatically
problem-solving knowledge. We attempt to apply our process to an online community discussing
breastfeeding issues. The main goal is to extract the problems encountered by new mothers and the proposed
“working” solutions for each problem. The contribution this study is to demonstrate how a textual content
analysis-based process and a graph analysis-based technique can be applied to textual online community’s
conversations to extract problem-solving knowledge and to evaluate it. The scope of this article is limited to
self-help-group concerning breastfeeding issues. The purpose is to build a knowledge graph to link problem’s
type entities to solution’s type entities and then to infer solutions of each problem in the graph. Finally, we
deduce the best solution for each problem. Existing studies about information usefulness evaluation are based
on complexes and time-consuming computational methods that do not take into consideration graph theory
techniques. Our approach provides simple and accurate method for evaluating best solutions to problems
with measurable score. To the best of our knowledge, our study is the first to provide processes to evaluate
breastfeeding problem-solving information shared in online communities.
The reminder of this paper unfolds: we firstly present the main theoretical basis related to
information and knowledge usefulness evaluation in online communities. We then introduce our proposed
research method to extract and evaluate best solutions for specific issues related to breastfeeding from online
community’s discussions. We then describe the experimental results and the discussions. Finally, we present
the conclusion and future work.
2. METHOD
2.1. Modelization of the problem-solving knowledge creation process in online communities
Modeling the process of problem-solving knowledge creation (PSKC) in an online community
facilitates its understanding for extraction and capitalization. As we are focusing on problem solving
knowledge, it is important first to understand the process of the creation of such content. As online support
groups offer an environment of connection, friendship, information sharing, and an increase in confidence
among their members [17], members expose unresolved problems and seek for solutions from experienced
members. This leads to the creation of problem-solving knowledge based on problem solving process which
can be represented as shown in Figure 2.
Figure 2. Modelization of the problem-solving knowledge creation (PSKC) process in online communities
According to Figure 2, PSKC process can be created starting by the definition of its steps: stated
problem by the user, proposed solutions by the community members and the asker feedback of the best
solution. For example, a member asks a question to describe a specific encountered problem. The community
members start proposing solutions which are based on personnel experiences or heard from others. The asker
is then in front of one or many proposed solutions to be tested. At this level, the asker can test the solution
and replay to the community by giving feedback of the best tested solution. In many cases and situations, the
asker may not give his feedback about the best tested solution that worked for him. In this case, the process
of knowledge creation will be composed basically of the first two steps: stated problem and proposed
solutions. In order to constitute a representation of problem-solving knowledge in online communities, we
propose in the next chapter the extraction process of problem-solving knowledge components in online
communities (OC).
2.2. Problem-solving knowledge extraction (visualization and evaluation) in online communities
Our goal is to automate the extraction of problem-solving knowledge. As shown in Figure 3, we
conceived a process of problem-solving knowledge extraction composed of three phases: keywords problem
Stated
problem by
the user
Proposed
solutions by
the
community
Validated
solution by
the asker

6353
and solutions detection, keywords problem and solution mapping and finally best solution evaluation phase.
Each step of this process will be detailed in the following sub-sections.
Figure 3. The process of problem-solving knowledge extraction in online communities (PSKEOC)
2.2.1. Keywords problem and solutions detection phase
Keywords are technical terms agreed and understood by the community that describe special issues,
situations or solutions given by the members. Those terms are used and mentioned in online discussions by
the members. The goal behind this phase is to automate the extraction and the classification of keywords
representing problems and solutions.
a. Keyword’s extraction
The objective is to extract those technical terms from user-generated content in online discussions.
Our approach for keywords extraction is based on the occurrence frequency of the candidate terms that may
represent keywords. Those keywords may be unigram terms or noun phrases. Keywords should be
semantically rich and encapsulate the maximum of information in order to be representative. A
domain-independent automatic term recognizer (ATR) is used for this step. Automated term extraction or
recognition (ATE) from textual documents is a sub-field in text mining. It can be used for different important
purposes, for example as inputs in ontology learning [18]. ATE aims to identify terminological units in
specific domain corpora [19]. Section 3 present the details of the conducted experiment.
b. Keyword’s classification
After keyword’s extraction step comes keyword’s classification. As our objective is the extraction of
problem-solving knowledge composed of the stated problems and proposed solution, the extracted keywords
are of two types: keywords describing encountered problems and keywords describing proposed solutions.
Keyword’s classification step aims to classify the extracted keywords into these two categories. We choose to
build a machine learning classifier to classify our keywords based on a supervised learning approach as
detailed in section 3.
2.2.2. Keywords problem and solution mapping phase
In this phase, we aim to visualize all the extracted keywords in order to infer the relationship
between them. The relation between a keywords of type “problem” and a keywords of type “solution” that
we want to infer is: “is a solution for”. In the following, we describe our approach for keywords mapping
phase.
a. Unique identifier attribution (UIA)
As shown in Figure 3, the first step of this phase is to assign to each post and comment created in the
online community a unique identifier. This step will allow to our process the mapping of all the content
shared in the online community and to detected the exact position (post or comment) where a keyword is
mentioned in the online discussions. Figure 4 describe our UIA approach. In Figure 4, Pi represents a content
of type “Post” whereas AjPi represents a content of type “Comment”. For each post Pi there is many
comments AjPi assigned which means that for a published post, there are many comments assigned.

 ISSN: 2088-8708
6354
Figure 4. Unique identifier attribution to posts and answers of the online community
b. Keywords co-occurrence detection
This step consists of assigning to each keyword all the positions (unique identifier UI) of posts or
comments where it is cited in the online community discussions. The goal is to link keywords with the UI
previously created. To this end we developed an algorithm as shown in the following code (algorithm 1:
mentions creation). According to the algorithm, the idea is to fetch in all posts and answers in our database
each keyword found in step “Keyword’s extraction” and then assign the cell address of the post or answer
where it is found to the keyword. To do so, the algorithm is composed of three loops: the first one (line 1)
browse every keyword to look if it is mentioned in posts (the second loop, line 2) and in answers (the third
loop, line 3). If the first keyword is mentioned in a post or an answer (line 4) then the algorithm returns the
cell address of the post or the answer (line 5). The algorithm generates a double column table where the first
one represents all the keywords and the second one the UI (cell address) where each keyword was mentioned.
Algorithm 1. Mentions creation
For all keywords
For all posts (Pi) in the first column
For all answers (AjPi) from the second column to the last column
if a keyword “Ki” is mentioned in a Post “Pi” or in an answer “AjPi”0
Then return the cell address where the keyword "Ki" is found and assign it to the
keyword "Ki"
End if
c. Keywords co-occurrence mapping
Keyword’s co-occurrence mapping step consists of building a keyword’s graph to infer the relation
between keywords of type problems with keywords of type solutions. The keyword’s graph we are willing to
construct is described: i) G (V, E) is an undirected graph where V is the set of nodes and E the set of edges;
ii) the set of nodes V is composed of two types of nodes: a set of keyword nodes {Ki} and the set of posts
and comments nodes {AjPi}; and iii) E is the set of edges between nodes that represents the occurrence of a
keyword in a specific content (post or comment). As illustrated in Figure 5, i) Kp is a keyword node of type P
(Problem) mentioned in three answers: AiPi and AsPi and AnPi. So, Kp has edges with nodes of type AtPi
related to the post Pi and ii) Ks is a keyword node of class S (solution) mentioned in three answers: AiPi and
AsPi and AnPi. So, Ks has edges with nodes of type AtPi related to the post Pi. The resulted graph is presented
in the result section.
Figure 5. Illustration example of our keyword’s graph
2.2.3. Best solution evaluation phase
In the previous phase, we detected the problems discussed and solutions proposed by the online
community. As cited previously, for each post describing a particular problem, many solutions may be

6355
proposed by the community. We need then to recommend to the community users the best solution for each
problem. In this last phase, we present our method for detecting and evaluating the best solution for each
problem mapped in the graph. This phase is composed of link prediction and best solution evaluation steps.
Each step is detailed bellow.
d. Link prediction
After mapping keywords co-occurrences through the construction of a graph of keywords and posts
and comments, the present step consists of inferring the relationship between keywords. This type of relation
is not represented directly in the graph. This relationship between keywords we are aiming to infer from the
graph will enable to detect the solutions proposed for a specific problem discussed in the online community.
This approach will help users and members to find relevant solutions to encountered problems related to the
subject discussed by the community as shown in Figure 6. For link prediction between keywords nodes, we
use common neighbor’s method from graph theory.
Figure 6. Illustration example of keyword’s graph, dashed red line depicts possible link between
keyword’s nodes
Considering an undirected simple network G (V, E), where V is the set of nodes and E is the set of
links. For a pair of nodes x, y∈V, let Γ(x) denote the set of neighbors of x. According to common neighbors
theory [20], [21], in common sense, two nodes, x and y, are more likely to have a link if they have many
common neighbors. The simplest measure of this neighborhood overlap is the directed count, namely:
CN(x, y) = |Γ(x) ∩ Γ(y)| (1)
In Figure 6, nodes Kp and Ks has many common neighbors which are AjPi, AsPi and AnPi, then
according to link prediction approach based on common neighbors, there is an inferred link between Ks and
Kp. The inferred link between the two nodes of different type namely “solution” and “problem” indicate that
the keyword Ks is a solution for keyword Kp. Concretely speaking, if many members recommend the same
solution many times (keyword node Ks) for the same problem (keyword node Kp) in a community, then
necessary that solution is working. Our approach supports the fact that if the same problem keyword has
many common neighbors with the same solution keyword, then it is highly the right solution for that
problem. Now, as many solutions might be recommended for the same problem in the community, one may
ask which is the best and working one? In the next sub-section, we present an approach for solutions
evaluation based on a confidence score.
e. Best solution evaluation
In order to rank the best solutions proposed by the online community to a given problem, we
propose to process a score for each pair of nodes of different type (problem/solution) based on common
neighbors’ theory. We say that the keyword Ks is a solution to keyword Kp to a certain degree of confidence
called confidence degree (Cd). We propose (1) to calculate the confidence degree of Ks being a solution to
Kp:
Cd ((Ks|Kp)) =
CN(Ks,Kp)
D(Kp)
(2)
where: Cd ((Ks|Kp)) the confidence degree of Ks being a solution for Kp, with Kp is a keyword node of type
problem and Ks a keyword node of type solution; CN(Ks, Kp) is the number of common neighbors between
nodes Ks and Kp (1); D(Kp) is the degree of node Kp meaning the number of all connections Kp has with all

 ISSN: 2088-8708
6356
nodes in the graph. In the case where a node of type problem Kp is not linked to any common nodes with a
node of type solution Ks, then Cd=0. In the following section we expose the conducted experiment and test of
our approach.
3. RESULTS AND DISCUSSION
In this section, we study the effectiveness of our proposed approach on a real online community
forum of breastfeeding mothers. The goal behind this section is to test if our method succeeded at extraction
valuable knowledge from raw instructed forum text. For this purpose, we compare the obtained results with
the World Health Organization Publication about breastfeeding knowledge.
3.1. Dataset
Data was extracted from an online discussion’s forum about breastfeeding publicly available using
Octoparse tool. Octoparse is a powerful web scraping tool that can help extracting open data from almost all
the websites and turn it to a well-structured file [22], [23]. In our case, we used Octoparse software to extract
all posts and comments created in the online forum until the date of the data collection. The tool generated a
cell surface vimentin (CSV) file containing two types of columns about online discussions: posts and answers
(each column contains a comment to a specific post). In the generated CSV file, each line represents a post.
As mentioned previously, the constituted dataset contains posts and answers generated by mothers sharing
their problems and issues about breastfeeding. The generated CSV file contains 2,557 posts and
20,456 answers. All the data’s characteristics are detailed in Table 1. In the following we give results about
the conducted experiment.
Table 1. Dataset characteristics
Characteristic Description
Text Language English
Type of text User-generated text in an online forum community
Number of posts 2,557 posts
Number of answers 20,456 answers
3.2. Results
3.2.1. Keywords problem and solutions extraction
After constructing our dataset composed of posts and their related answers, we start the keyword
extraction step. Our goal is to extract meaningful keywords characterizing each problem and its solutions. In
our context, we choose TerMine [24], a multi-word automatic term recognizer (ATR) that annotates the input
text with candidate multiword terms recognized by the C-value method and acronyms. C-value is a
domain-independent method for multiword ATR which aims to improve the extraction of nested terms as
detailed in [24]. Nested terms are those that appear within other longer terms, and may or may not appear by
themselves in the corpus [24]. This method takes as input a corpus and produces a list of candidates
multi-word terms. The C-value approach combines linguistic and statistical information. The linguistic
information consists of the part-of speech tagging of the corpus, the linguistic filter constraining the type of
terms extracted, and the stop-list [24]. The part-of-speech tagger used for our dataset is the TreeTagger [25].
The statistical information consists of measuring the above characteristics: the total frequency of occurrence
of the candidate string in the corpus, the frequency of the candidate string as part of other longer candidate
terms, the number of these longer candidate terms and the length of the candidate string (in number of
words). The resulting keywords database is composed of 69 keywords extracted from posts (questions) and
135 keywords extracted from answers as presented in Table 2.
Table 2. Keyword’s extraction results
Post’s/question’s text Answer’s text
Number of extracted keywords 69 135
The extracted keywords are bigram composed of two terms. Bi-gram keywords play a very
important role in our knowledge extraction process because they allow us to capture more meaning and
semantics than uni-gram terms. In order to obtain meaningful results, we decide to filter the extracted
keywords and only keep those who’s scores are greater than or equal to 2 [26]. The following step in our

6357
approach concerns keywords classification. To this end, we conceived an implementation process based on
machine learning classifiers presented in Figure 7 using Weka [27], an open-source machine learning
software.
The first step of the process is about the dataset constitution. Our task consists of classifying
keywords into two categories as mentioned previously. As we adopt a supervised machine learning approach,
our test dataset of keywords should be annotated. The annotation process consisting of manually labelling the
test data set with right classes was done by a breastfeeding experimented mother who manually classified the
keywords according to solutions or problems. The second step consist of applying a StringToWordVector
filter to convert text attributes into numerical attributes representing word occurrences.
Figure 7. Keywords classification process
The final step consists of applying the machine learning algorithms to train the classifiers. In our
case we choose to apply and compare three machine learning classifiers: support vector machine (SVM)
algorithm, multilayer perceptron (MLP) algorithm and DL4jMlp deep learning algorithm in WEKA platform.
The results of this step are detailed in subsection 3.3.1. The performance of the applied classifiers was
estimated based on a k-fold cross-validation. In our case, we tried many values for k and choose the value 10
because it generated the best result.
3.2.2. Keyword mapping, graph construction, and solution evaluation
After assigning a unique identifier (UI) to each post and comment in our dataset, we run the
algorithm 1 presented in sub-section 2.2.2. The result is an excel file that we imported in Gephi [28], an open
software for graph visualization. The resulted graph contains nodes representing keywords, posts and
answers and edges representing mentions of co-occurrences. Figure 8 shows part of our resulted knowledge
graph. The next and final step of the graph construction process aims to answer the following question: given
a problem, what is, according to the community the solution for it? Best solution evaluation step is performed
after applying (2) to each keyword pairs (problem/solution) in the generated graph. In order to give example
of some results generated by the knowledge graph, we apply (2) to some keywords represented in the
constructed graph as shown in Figure 8. The result is shown in Table 3.
As exposed in Table 3, the problem “clogged duct” has three possible solutions according to the
community: “warm compress”, “hot shower” and “hot water”. This means that a breastfeeding mother may
suffer from clogged duct which is an area of the breast where milk is blocked. According to the community
shared experiences, this issue can be resolved by taking a hot shower or by putting a warm compress on the
blocked area of the breast. To decide which of the solutions is effective, we applied our proposed confidence
degree measure in (2). According to Table 3, “hot shower” is more likely to be a good solution for clogged
duct issue with a confidence degree of 22.7% compared to “warm compress” where its confidence degree is
4.5%. In the same Table 3, solutions “nipple cream” and “Lactation consultant” are proposed by the
community to the problem tongue tie. Tongue tie can cause serious nipple pain for the mother. According to
calculated Cd, the best solution for this issue is to consult a lactation consultant which the Cd is 27.2%. For
the problem “Nipple confusion” the only given solution by the community members is to use “Nipple Shield”
which is a flexible silicone nipple that is worn over the mom’s nipple during a feeding. In the next
sub-section, we present the evaluation method to assess the effectiveness of our approach in extracting and
evaluating solutions to problems discussed in online communities about breastfeeding.
3.3. Evaluation and discussion
In order to evaluate our process of problem-solving knowledge extraction in online communities
(PSKEOC) applied to breastfeeding knowledge, we choose to compare the obtained results with the Infant
and young child feeding: model chapter [29] by the World Health Organization (WHO). It describes essential
knowledge and basic skills that every health professional who works with new mothers and young children
health care should master [29]. Most importantly, it contains in his 7th
session details and solutions about how
to manage breastfeeding difficulties. Our evaluation process aims to compare solutions to problems provided
by this chapter to those extracted by our system and test the credibility of the obtained results. Our evaluation

 ISSN: 2088-8708
6358
process is composed of two parts: the first one concerns evaluating the extraction and classification step and
the second concerns evaluating the attribution of correct solutions to extracted problems.
Figure 8. Problem solving knowledge graph
Table 3. Comparison of confidence degree value between solutions according to
the obtained graph in Figure 8
Keywords of type Problem Common neighbor nodes of
type keywords (Solutions)
Cd*100 (confidence
score*100)
Best deduced solution
(with higher Cd)
“Clogged duct” “Warm compress” =1/22*100=4,5% “Hot Shower”
“Hot Shower” =5/22*100
=22.7%
“Hot water” =1/22*100
=4.5%
“Tongue tie” “Nipple cream” =2/11*100
=18.2%
“Lactation consultant”
“Lactation consultant” =3/11*100
=27.2%
“Nipple confusion”
“Nipple shield”
=3/18*100
=16.6%
“Nipple shield”

6359
3.3.1. Evaluation of keywords problem and solutions extraction and classification
To evaluate our classification task, we use three metrics: precision, recall and F-measure. Precision
measures the percentage of items that the system detected as in fact positive (according to human labels).
Recall measures the percentage of items (posts) present in the input that were correctly identified by the
system. F-measure incorporates aspects of both precision and recall [30]. As detailed in as detailed in
sub-section 3.2.1, we trained three machine learning classifiers which are: SVM, MLP and deep learning-
based classifier. The results are presented in Tables 4 to 6 which expose the tree evaluation metrics of the
classifiers for each category.
Table 4. Classification evaluation metrics using
SVM classifier
Precision Recall F-measure
Category “Problem” 0.745 0.897 0.814
Category “Solution” 0.877 0.704 0.781
Table 5. Classification evaluation metrics
using MLP
Table 6. Classification evaluation metrics using deep learning classifier
To compare the three classifiers’ results, we processed the weighted average of each classifier as
presented in Figure 9. According to Figure 9, SVM achieved the best results with 79.874% of correctly
classified instances. This means that SVM classifier can reach good results when it comes to classify natural
language text related to breastfeeding difficulties.
Figure 9. Evaluation metrics weighted average of the three classifiers
3.3.2. Evaluation of best solution evaluation phase
Table 7 presents an overview of the difficulties mentioned in the WHO reports about breastfeeding
with the proposed solutions for managing them. Table 7 presents also extracted problems by our system with
their extracted solutions as well. The goal behind this table is to compare both difficulties and solutions to
conclude if our system succeeded in extracting solutions to each breastfeeding problem.
According to the Table 7, there are many difficulties extracted by our system that are discussed in
the WHO chapter. For instance, our system succeeds to extract keyword difficulties like: irritated nipples,
clogged duct, breast abscess, and slow flow. However, the same difficulties are not managed the same way in
the community as by the WHO chapter. For example, according to the WHO model chapter, the issue “large
nipples” should be resolved by “expressing breast milk” or by “taking a different position for breastfeeding”.
However, with the same issue, the online community recommend to apply a “nipple shield” to breastfeed the
baby as shown in Table 7. Additionally, there are some breastfeeding difficulties discussed in the community
that are not in the WHO Chapter like “growth spurt” issue.

 ISSN: 2088-8708
6360
Breastfeeding is a domain where new mothers are highly concerned as they need a lot of support in
order to succeed breastfeeding their babies. Lacking of support while breastfeeding is considered as a serious
problem that can interrupt breastfeeding. We believe that our process can help mothers and breastfeeding
professionals as well. The obtained knowledge graph can be considered as a real knowledge base for the
community as it can help both professionals and members of the online community to understand
encountered problems and find most working solutions to specific issues related to breastfeeding domain.
Our approach can also be helpful in the decision-making process as it provides a confidence score that
characterizes the confidence degree of each possible solution based on lived experiential knowledge of
members.
Table 7. comparison between difficulties and solutions mentioned in the WHO report and
our PSKEOC process
Some difficulties
mentioned in the WHO
report
Proposed solutions by the WHO report Some problem
keywords extracted by
our PSKEOC process
Proposed solutions to
problems extracted by our
PSKEOC process
Full breasts Breastfeed the baby frequently to remove
the milk
Huge breasts -
Breast engorgement remove the breast milk or remove the
breast milk
- -
Blocked duct: localized
lump in one breast
− Feed from the affected breast frequently
and gently massage the breast over the
lump while her baby is suckling.
− Apply warm com-presses, and to vary
the position of the baby.
− Gentle massage over the lump
Clogged duct − Hot Shower
− Hot compress
Mastitis − Improve the removal of milk.
− Apply warm compresses
− Vary the position of the baby
Milk blister Lactation nurse
Breast abscess − Penicillinase-resistant antibiotics
− Express milk
Breast abscess -
Sore or fissured nipple − Improve baby’s position and
attachment.
Irritated nipple Coconut oil
Mastitis, abscess and nipple
fissure in an HIV- infected
woman
− Avoid breastfeeding on the affected side
− Remove the milk from the affected
breast by expression
− Give antibiotics for 10–14 days
− Candida infection (thrush)
in mother and baby
− White spots inside the
cheeks or over the tongue
− Gentian violet or nystatin Supply issue − Pump session
− Lactation consultant
− Milk tea
− Brewers yeast
Inverted, flat, large and
long nipples
− The mother takes a different position
for breastfeeding.
− Express breast milk and feed it by cup
Flat nipple Nipple shield
- - Growth spurt − Hot shower
− Licithine capsule
− Birth control pill
− Lactation consultant
− Warm compress
− Football hold and
− Ice Pack.
4. CONCLUSION
Extracting, evaluating and facilitating the reuse of knowledge encapsulated in user generated content
is an important and challenging task. Online communities are a medium for sharing knowledge components
between users through asking and answering questions and resolving problems. Our work proposes an
approach for solutions extracting and evaluation in an online community of breastfeeding mothers. Few
studies have tacked this important question of capitalizing shared knowledge in online communities. Our
presented method is based on knowledge graphs technology and graph theory approaches which enable
representing extracted knowledge and reasoning on it. We developed a process for problem-solving
knowledge extraction in online communities (PSKEOC) composed of three steps. The first one enables the
extraction and classification of knowledge components from the online community conversations based on
machines learning algorithms. The second step concerns the representation of knowledge components in a
knowledge graph. The third step aims to evaluate and return best solutions to problems discussed in the

6361
community using link prediction method. In order to evaluate the obtained results, we compared them to the
WHO chapter on managing breastfeeding issues. The results show that our approach succeeded at extracting
breastfeeding difficulties and their right evaluated solutions. Our process even extracted some difficulties that
are not discussed in the WHO report but shared by the online community members with their corresponding
solutions. The resulted knowledge graph about breastfeeding difficulties management can be considered as a
real way for improving problem solving process in online communities and capitalizing its results. We aim in
our future work to apply our process to many other datasets in different domain in order to evaluate the
performance of our process in extracting problem solving-knowledge components in online communities. We
aim also to enrich the obtained knowledge graph with more semantics in order to maximize the knowledge
extraction, representation and capitalization.
REFERENCES
[1] Y. Tausczik and X. Huang, “Knowledge generation and sharing in online communities: current trends and future directions,”
Current Opinion in Psychology, vol. 36, pp. 60–64, Dec. 2020, doi: 10.1016/j.copsyc.2020.04.009.
[2] A. Zagalsky, D. M. German, M.-A. Storey, C. G. Teshima, and G. Poo-Caamaño, “How the R community creates and curates
knowledge: an extended study of stack overflow and mailing lists,” Empirical Software Engineering, vol. 23, no. 2, pp. 953–986,
Apr. 2018, doi: 10.1007/s10664-017-9536-y.
[3] J. Cole, C. Watkins, and D. Kleine, “Health advice from internet discussion forums: how bad is dangerous?,” Journal of Medical
Internet Research, vol. 18, no. 1, Jan. 2016, doi: 10.2196/jmir.5051.
[4] I. Buunk, C. F. Smith, and H. Hall, “Tacit knowledge sharing in online environments: Locating ‘Ba’ within a platform for public
sector professionals,” Journal of Librarianship and Information Science, vol. 51, no. 4, pp. 1134–1145, Dec. 2019, doi:
10.1177/0961000618769982.
[5] N. Nimkulrat, C. Groth, O. Tomico, and J. Valle-Noronha, “Knowing together-experiential knowledge and collaboration,”
CoDesign, vol. 16, no. 4, pp. 267–273, Oct. 2020, doi: 10.1080/15710882.2020.1823995.
[6] R. Julian, B. Bliesemann de Guevara, and R. Redhead, “From expert to experiential knowledge: exploring the inclusion of local
experiences in understanding violence in conflict,” Peacebuilding, vol. 7, no. 2, pp. 210–225, May 2019, doi:
10.1080/21647259.2019.1594572.
[7] T. Borkman, “Experiential knowledge: a new concept for the analysis of self-help groups,” Social Service Review, vol. 50, no. 3,
1976
[8] P. H. Gray, “A problem-solving perspective on knowledge management practices,” Decision Support Systems, vol. 31, no. 1, pp.
87–102, 2001, doi: 10.1016/S0167-9236(00)00121-4.
[9] M. Tanis, “Health-related on-line forums: what’s the big attraction?,” Journal of Health Communication, vol. 13, no. 7, pp. 698–
714, Oct. 2008, doi: 10.1080/10810730802415316.
[10] Y. Zhao and J. Zhang, “Consumer health information seeking in social media: a literature review,” Health Information and
Libraries Journal, vol. 34, no. 4, pp. 268–283, Dec. 2017, doi: 10.1111/hir.12192.
[11] T. R. Bandaragoda, D. De Silva, D. Alahakoon, W. Ranasinghe, and D. Bolton, “Text mining for personalized knowledge
extraction from online support groups,” Journal of the Association for Information Science and Technology, vol. 69, no. 12, pp.
1446–1459, Dec. 2018, doi: 10.1002/asi.24063.
[12] M. K. Daradkeh, “Exploring the usefulness of user-generated content for business intelligence in innovation,” International
Journal of Enterprise Information Systems, vol. 17, no. 2, pp. 44–70, Apr. 2021, doi: 10.4018/IJEIS.2021040103.
[13] M. Nowakowski, “Comparative analysis of information usefulness evaluation methods on business internet services,” European
Research Studies Journal, vol. 23, no. 2, pp. 292–306, Nov. 2020, doi: 10.35808/ersj/1825.
[14] X. Liu, G. A. Wang, W. Fan, and Z. Zhang, “Finding useful solutions in online knowledge communities: a theory-driven design
and multilevel analysis,” Information Systems Research, vol. 31, no. 3, pp. 731–752, Sep. 2020, doi: 10.1287/isre.2019.0911.
[15] S. Ji, S. Pan, E. Cambria, P. Marttinen, and P. S. Yu, “A survey on knowledge graphs: representation, acquisition, and
applications,” IEEE Transactions on Neural Networks and Learning Systems, vol. 33, no. 2, pp. 494–514, Feb. 2022, doi:
10.1109/TNNLS.2021.3070843.
[16] A. Newell, J. C. Shaw, and H. A. Simon, “Report on a general problem-solving program,” 1969.
[17] T. Noorani, M. Karlsson, and T. Borkman, “Deep experiential knowledge: reflections from mutual aid groups for evidence-based
practice,” Evidence and Policy, vol. 15, no. 2, pp. 217–234, May 2019, doi: 10.1332/174426419X15468575283765.
[18] V. Kosa et al., “Cross-evaluation of automated term extraction tools,” 2017.
[19] M. da Silva Conrado, T. A. Salgueiro Pardo, and S. O. Rezende, “A machine learning approach to automatic term extraction using
a rich feature set,” in Proceedings of the 2013 Annual Conference of the North American Chapter of the Association for
Computational Linguistics: Human Language Technologies, NAACL-HLT 2013-Student Research Workshop, 2013, pp. 16–23.
[20] T. Zhou, L. Lü, and Y.-C. Zhang, “Predicting missing links via local information,” The European Physical Journal B, vol. 71, no.
4, pp. 623–630, Oct. 2009, doi: 10.1140/epjb/e2009-00335-8.
[21] X. Fan et al., “Local core members aided community structure detection,” Mobile Networks and Applications, vol. 24, no. 4, pp.
1373–1381, Aug. 2019, doi: 10.1007/s11036-018-0994-2.
[22] A. J. Arumugham, D. Ahamad, A. Mahmoud, M. Mahmoud, and M. Akhtar, “Strategy and implementation of web mining tools,”
International Journal of Innovative Research in Advanced Engineering, vol. 12, no. 12, pp. 1–7, 2017
[23] M. F. Shadiqin Thirafi and F. Rahutomo, “Implementation of naïve bayes classifier algorithm to categorize indonesian song lyrics
based on age,” in 2018 International Conference on Sustainable Information Engineering and Technology (SIET), Nov. 2018, pp.
106–109., doi: 10.1109/SIET.2018.8693201.
[24] K. Frantzi, S. Ananiadou, and H. Mima, “Automatic recognition of multi-word terms: the C-value/NC-value method,”
International Journal on Digital Libraries, vol. 3, no. 2, pp. 115–130, Aug. 2000, doi: 10.1007/s007999900023.
[25] H. Schmid, “Probabilistic part-of-speech tagging using decision trees,” 1994.
[26] H. Sekkal, N. Amrous, and S. Bennani, “Knowledge components detection in user-generated content,” in 2020 International
Conference on Intelligent Systems and Computer Vision (ISCV), Jun. 2020, pp. 1–6., doi: 10.1109/ISCV49265.2020.9204188.
[27] I. H. Witten, E. Frank, M. A. Hall, and C. J. Pal, Data mining, fourth edition: practical machine learning tools and techniques,
4th ed. Morgan Kaufmann, 2016.

 ISSN: 2088-8708
6362
[28] M. Bastian, S. Heymann, and M. Jacomy, “Gephi: An open source software for exploring and manipulating networks,” in
International AAAI Conference on Weblogs and Social Media, 2009, vol. 3, no. 1, pp. 361–362.
[29] W. health organization, “Infant and young child feeding: Model Chapter,” World Health Organization. 2009.
[30] S. Houda, A. Naila, and B. Samir, “Machine learning based knowledge organization in online communities,” in Proceedings of
the 4th International Conference on Big Data and Internet of Things, Oct. 2019, pp. 1–8., doi: 10.1145/3372938.3372965.
BIOGRAPHIES OF AUTHORS
Houda Sekkal is Data and Knowledge Engineer and a Ph.D. Candidate at the
Laboratory of Research in Computer Science and Education, Mohammadia School of
Engineers, Mohammed V University of Rabat, Morocco. Her research project is about the use
of knowledge management and data mining methods to improve the process of learning in
online learning communities. She can be contacted at email: houda.sekkal@gmail.com.
Naïla Amrous is a full professor at School of Information Sciences. She has a
PhD in Information and Communication Sciences. Ongoing research: information literacy
(IL), knowledge management (KM); innovative teaching practices. She can be contacted at
email: amrousnaila@yahoo.fr.
Samir Bennani is a Full Professor and Deputy Director of students and academic
affairs at Mohammadia School of Engineers. Engineer degree in Computer Science in 1982;
PhD in Computer Science in 2005; Professor at the Computer Science Department-EMI; 34
recent publications papers between 2014 and 2017; Ongoing research interests: SI, modeling
in software engineering, information system, e-learning content engineering, tutoring,
assessment and tracking. He can be contacted at email: sbennani07@gmail.com.

Knowledge graph-based method for solutions detection and evaluation in an online problem-solving community

More Related Content

Similar to Knowledge graph-based method for solutions detection and evaluation in an online problem-solving community (20)

More from IJECEIAES (20)

Recently uploaded (20)

Knowledge graph-based method for solutions detection and evaluation in an online problem-solving community