Enhancing a User Matchmaking Algorithm using Personalized

PageRank
Santipong Thaiprayoon Herwig Unger
[email protected] [email protected]
Fernuniversität in Hagen Fernuniversität in Hagen
Hagen, Germany Hagen, Germany

ABSTRACT and a myriad of interests, leading to challenges in effectively iden-

With the increasing number of users in online communities and tifying and recommending potential friends or collaborators. One
social networking platforms, it is becoming more difficult for users of the most valuable tools is user matchmaking, which plays a sig-
to meet and connect with individuals who share similar opinions or nificant role in facilitating users by suggesting potential friends or
interests. The paper proposes a user matchmaking algorithm based matches based on various factors associated with a particular user,
on personalized PageRank to provide potential friends to individ- such as personal information, interests, demographics, preferences,
ual users. A set of user profiles is transformed into a graph model social connections, and behavioral data [20, 21].
for efficiently discovering meaningful connections and influential Matchmaking algorithms are crucial for numerous online plat-
users. The semantic relationship between two user profiles is then forms, including dating applications, social networks, e-commerce
estimated using word and sentence embeddings. By incorporating sites, and online games [5]. These matchmaking algorithms utilize
both embedding models and personalized graph analytics, the pro- various methods and data sources to analyze user profiles, behav-
posed algorithm can capture complex semantic information and iors, and feedback to generate personalized recommendations for
high-order user relationships, making the matchmaking process potential matches. For example, in the context of online dating
more accurate. Experiments conducted on a simulated user profile sites, it can assist users in expanding their social circle and meeting
dataset show that the proposed algorithm consistently outperforms romantic partners who share their values, goals, and personality
existing state-of-the-art methods in terms of the F1 score and mean traits. Similarly, social media platforms apply user matchmaking
average precision metrics. algorithms to match users based on their interests, preferences,
and behaviors. Additionally, gaming platforms use matchmaking
CCS CONCEPTS algorithms to match players based on their skill level. This leads to
increased user engagement and satisfaction, which benefits both
• Information systems → Information retrieval; Retrieval
users and platforms.
tasks and goals; Recommender systems;
Of course, the current state of research in the fields of user match-
making and friend recommendation relies on a variety of techniques
KEYWORDS [2, 3, 10], including Collaborative Filtering (CF), Graph-Based Ap-
Social Networks, Matchmaking Algorithm, Friend Recommenda- proaches (GBA), Natural Language Processing (NLP), and Machine
tion, Personalized PageRank, Graph Analytics, Semantic Textual Learning (ML). Most studies have attempted to propose a hybrid
Similarity model that incorporates several techniques, such as content-based
ACM Reference Format: filtering, distance similarity, and user-based collaborative filtering,
Santipong Thaiprayoon and Herwig Unger. 2023. Enhancing a User Match- with semantic and social recommendations. The semantic dimen-
making Algorithm using Personalized PageRank. In 2023 7th International sion suggests semantically close friends based on calculating the
Conference on Natural Language Processing and Information Retrieval (NLPIR similarity between members by leveraging interest and preference
2023), December 15–17, 2023, Seoul, Republic of Korea. ACM, New York, NY, data. The social dimension is based on some social-behavior metrics,
USA, 9 pages. https://doi.org/10.1145/3639233.3639346 such as friendship and credibility degree [11]. The remaining study
proposed a friend recommendation method based on the social
1 INTRODUCTION structures and behaviors of users [22]. The degree of interaction
between users is computed to recommend candidate friends using
The ubiquity of online communities and social networking plat-
random walk algorithms with a restart model. However, previous
forms has fundamentally transformed the way individuals connect,
studies still lack research on the semantic-based similarity of users
interact, communicate, and build meaningful relationships [14, 17].
in textual data and graph-based personalized matchmaking. This fa-
These platforms are a collection of people with diverse backgrounds
cilitates improving the matchmaking process, resulting in enhanced
accuracy and personalized matching.
To bridge the gap, the paper proposes a user matchmaking algo-
This work is licensed under a Creative Commons Attribution International rithm using Personalized PageRank (PPR) and semantic analysis.
4.0 License. The proposed algorithm aims to discover potential users with per-
sonalized and relevant recommendations based on user profiles,
NLPIR 2023, December 15–17, 2023, Seoul, Republic of Korea
© 2023 Copyright held by the owner/author(s). enhancing the social experience and satisfaction in online commu-
ACM ISBN 979-8-4007-0922-7/23/12 nities and social networking platforms. This algorithm leverages
https://doi.org/10.1145/3639233.3639346

319
NLPIR 2023, December 15–17, 2023, Seoul, Republic of Korea Santipong and Herwig

textual profiles of users consisting of various attributes, includ- system. The system considers the personality traits and harmony
ing interests, hobbies, occupations, and biographies, to construct ratings of users instead of common physical or social features. Anju
a weighted graph for efficiently finding hidden connections be- et al. [23] proposed a buddy recommendation model that employs
tween users and influential users. The relationship between two collaborative filtering to compare and contrast similar and dissimi-
user profiles is calculated using word and sentence embedding lar user data. The model generates user recommendations based
techniques to capture their semantic textual similarity rather than on their comparable choices, activities, and preferences. Finally, Ali
lexical similarity. This algorithm can also determine complex se- et al. [4] proposed a framework for a friend recommender system
mantic information, individual priorities with a particular user, and that leverages hashtags to enhance the content and quality of user
high-order user relationships by utilizing the weighted graph. The profiles. The framework initiates the construction of a user profile
PPR method then ranks the importance of each user in the graph by leveraging shared hashtags. The matching approach is utilized
and provides a list of matching users or friends to individual users to calculate the degree of similarity between profiles and group
according to PageRank score values. users with similar interests using advanced clustering methods.
The main contributions of this paper are summarized as follows: Others have explored graph-based approaches, which lever-
• This paper proposes a new perspective on user matchmaking age the social connections between users to recommend potential
algorithms by utilizing the PPR algorithm. The goal is to friends. For instance, Runa et al. [8] developed an integrated frame-
automatically identify users within a virtual community who work that learns user attributes, associated interactions, network
share common interests, preferences, and traits. structure, and timeline history from an online social network using
• This paper introduces a method for measuring the seman- a graph-theoretic approach to generate friend recommendations.
tic similarity between textual user profiles based on word Bu-Xiao et al. [27] developed a user similarity graph based on the
and sentence embeddings. This method could help to un- interests of individual users. The LDA algorithm is used to deter-
derstand the meanings between texts and achieve superior mine topics of interest to users. The graph is constructed using
performance in text matching. the multi-view similarity method, which calculates the degree to
• Extensive experiments on a simulated user profile dataset which users share similar topics of interest. The system then makes
are conducted to evaluate the performance and effectiveness recommendations between users by analyzing the graph. Addi-
of the proposed algorithm in comparison to existing state- tionally, there have been research efforts to incorporate machine
of-the-art methods. learning and deep learning models to improve the accuracy of user
matchmaking and recommendation systems [3].
The remainder of this paper is organized as follows: Section 2 Despite the advancements of previous approaches, a few weak-
provides a concise overview of the related work on user matching nesses still need to be addressed. One weakness is that they do
and friend recommendation. In Section 3, the proposed algorithm not fully consider the semantic relationship between users in tex-
is described. The experimental details are presented in Section 4. tual data. Additionally, there is a lack of research on analyzing
Section 5 discusses experimental results. The limitations of evalua- the connections between users in a graph model with personal-
tion and impact research are explained in Section 6. Section 7 gives ization. To address these shortcomings, this paper introduces a
opportunities and challenges. Section 8 concludes the paper and user matchmaking algorithm from a new perspective by combining
outlines some possible directions for future work. semantic analysis and the PPR method. This approach can improve
the accuracy and personalization of the matchmaking algorithm.
2 RELATED WORK
The field of user matchmaking and recommendation algorithms has 3 METHODOLOGY
gained significant attention in recent years. This section provides This section describes the methodology for discovering potential
a review of research studies on the recommendation and match- users using the PPR method and semantic analysis, considering tex-
making of users, also known as friends, members, or individuals, tual user profiles instead of mutual friends or social graph criteria.
which aim to recommend friends to users based on their preferences, The goal of this algorithm is to automatically match users within
interests, and social connections. The review of research studies a virtual community who have similar interests, preferences, and
discusses the different types of user matchmaking algorithms, their traits. The processing pipeline of the user matchmaking algorithm
advantages and disadvantages, and their effectiveness in different is illustrated in Figure 1.
settings.
Some of the early research studies focused on content-based fil-
tering, which utilizes user preferences and interests to recommend
friends or people. For example, a study by Jingda et al. [15] proposed
an algorithm based on Latent Dirichlet Allocation (LDA) for rec-
ommending friends to learners in online education. The algorithm
groups together learners with similar learning interests to find the
top-𝑘 friend recommendation sequences. It does this by making
learner document datasets, figuring out how similar learners are, Figure 1: The Pipeline of the User Matchmaking Algorithm
and modeling the friend topic. Huansheng et al. [19] proposed a
friend recommendation system that utilizes the Big-Five personal- The proposed algorithm starts by preprocessing each user pro-
ity traits model and hybrid filtering to improve the accuracy of the file. This involves cleaning the textual data, removing stop words,

320
Enhancing a User Matchmaking Algorithm using Personalized PageRank NLPIR 2023, December 15–17, 2023, Seoul, Republic of Korea

and stemming words. Each text attribute of a user profile is then Table 1: Basic Notations and Symbols Used
converted into a dense vector representation using embedding mod-
els. The cosine similarity score is calculated for each attribute pair Notations Description
of user profiles. These scores are then averaged to get an aggre- 𝑈 A set of all users
gated similarity score. The score is inserted into a similarity matrix 𝑃 A set of all user profiles
to construct a weighted graph, where nodes represent users and 𝑁 The element number of user profile 𝑃
edges represent their relationships. The weight of an edge can be 𝐴𝑖1 The interest attribute of a user profile 𝑃𝑖
an aggregated similarity score, which indicates the strength of the 𝐴𝑖2 The hobby attribute of a user profile 𝑃𝑖
relationship between two users. The PPR method is then used to 𝐴𝑖3 The occupation attribute of a user profile 𝑃𝑖
identify influential users based on a target user within the weighted 𝐴𝑖4 The biography attribute of a user profile 𝑃𝑖
graph and to generate a list of matching users who have similar 𝑊 A set of all words or phrases in an attribute
profiles. The pseudocode of the user matchmaking algorithm is


ˆ 𝐴𝑖 𝑗
𝑣𝑒𝑐 A single vector representation of each attribute in 𝑃𝑖
illustrated formally in Algorithm 1.

Algorithm 1: User Matchmaking Algorithm The user matchmaking algorithm consists of three main parts:
Data: List of User Profiles 𝑃, Target User 𝑇 (1) text representation; (2) graph representation; and (3) user match-
Result: Ranked List of Users According to PageRank Score making. The following subsection explains each part in detail.
to 𝑇
Function Embed(𝑡𝑒𝑥𝑡)
3.1 Text Representation
return Vector representation of 𝑡𝑒𝑥𝑡 using embedding This part preprocesses user profiles and converts them to high-
models dimensional vector representations, called embeddings. These vec-
Function CosineSimilarity(𝐴, 𝐵) tor representations enable efficient comparison and matching of
user profiles based on their textual information, which captures
return ∥𝐴∥𝐴·𝐵
× ∥𝐵 ∥
2 2 the semantic meaning of individual words and texts. The funda-
Function PersonalizedPageRank(𝐺, 𝑡𝑎𝑟𝑔𝑒𝑡_𝑢𝑠𝑒𝑟 ) mental concept is that words or texts with similar meanings will
Initialize rank vector 𝑟 with target user set to 1 and have similar vectors. The text representation consists of two major
others set to 0 components, which are described in the following sections.
Initialize transition matrix 𝑀 from 𝐺
while not converged do 3.1.1 Text Preprocessing. The text preprocessing component plays
𝑟 ′ = (1 − 𝑑) × 𝑀 × 𝑟 + 𝑑 × 𝑣 a significant role in boosting the performance of the proposed
𝑟 = 𝑟′ algorithm. The main idea behind this component is to clean up and
transform unstructured textual profiles into a suitable format for
end
further analysis. The following modules are involved in the text
return sorted list of users based on 𝑟 , excluding target
preprocessing component:
user
for each user profile 𝑝𝑖 in 𝑃 do • Tokenization: This module breaks down the textual profiles
for each user profile 𝑝 𝑗 in 𝑃 do into individual words, or tokens.
𝑣𝑒𝑐ˆ 𝑖 ← Concatenated Embed(of all attributes of 𝑃𝑖 ) • Lowercasing: All the words in the profiles are converted to
𝑣𝑒𝑐ˆ 𝑗 ← Concatenated Embed(of all attributes of 𝑃 𝑗 ) lowercase. This helps in standardizing the text and avoiding
𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦𝑀𝑎𝑡𝑟𝑖𝑥 [𝑖] [ 𝑗] ← duplication of words due to case differences.
CosineSimilarity(𝑣𝑒𝑐 ˆ 𝑖 , 𝑣𝑒𝑐
ˆ 𝑗) • Removing Stop Words: Stop words are common words such
as “the”, “is”, and “and”. They are removed from the profiles
end
to reduce noise.
end
• Removing Special Characters: This module removes any
for each 𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦𝑀𝑎𝑡𝑟𝑖𝑥 [𝑖] [ 𝑗] do special characters such as punctuation marks, symbols, and
Construct a weighted edge between user 𝑖 and user 𝑗 in
emojis from the profiles. This helps eliminate any unnec-
graph 𝐺 with weight 𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦𝑀𝑎𝑡𝑟𝑖𝑥 [𝑖] [ 𝑗]
essary noise and ensure that only meaningful words are
end included.
𝑟𝑎𝑛𝑘𝑒𝑑𝑈 𝑠𝑒𝑟𝑠 ← PersonalizedPageRank(𝐺,𝑇 ) • Stemming and Lemmatization: These modules convert
return 𝑟𝑎𝑛𝑘𝑒𝑑𝑈 𝑠𝑒𝑟𝑠 words to their base or root form. This helps improve the
accuracy of the proposed algorithm by grouping together
Before discussing the proposed algorithm, essential notations words that have similar meanings.
are formally defined in Table 1. Let 𝑈 = {𝑈 1, 𝑈 2, ..., 𝑈 𝑁 } be a set The preprocessed user profiles are sent to the text embedding
n users, 𝑃 = {𝑃o1, 𝑃2, ..., 𝑃 𝑁 } be a set of user profiles, and 𝑊 =
of component, which generates high-dimensional vector representa-
𝑤𝑖1𝑗 , 𝑤𝑖2𝑗 , ..., 𝑤𝑖𝑚𝑗 be a set of words. Given a set of 𝑁 user profiles tions.
𝑃𝑖 , and 𝑃𝑖 = {𝐴𝑖1, 𝐴𝑖2, 𝐴𝑖3, 𝐴𝑖4 }. Therefore, 𝑤𝑖𝑚𝑗 𝜖𝐴𝑖 𝑗 𝜖𝑃𝑖 is the word 3.1.2 Text Embedding. This component aims to convert each text
𝑚𝑡ℎ of 𝐴𝑖 𝑗 , where 𝑖 = {1, 2, ..., 𝑁 }, and 𝑗 = {1, 2, 3, 4}. attribute of a user profile into a dense vector representation by

321
NLPIR 2023, December 15–17, 2023, Seoul, Republic of Korea Santipong and Herwig

utilizing two embedding models. The purpose of these dense vector 3.2.1 Pairwise Similarity Matrix Building. This component calcu-
representations is to understand the semantic relationships between lates the cosine similarity between every pair of user profiles based
users. on their vector representations and inserts them into a pairwise
• Term-based Embedding: This embedding model considers similarity matrix. This matrix construction consists of the following
the text attributes of a user profile, including interests, hob- three steps:
bies, and occupations. Each text attribute is mapped into the • Cosine Similarity Calculation: The cosine similarity
word vectors of all words in the attribute using Word2Vec, calculation is used to compute the semantic similarity score
an existing pre-trained word embedding model with 300- between two single vectors for every attribute pair. This
dimensional word vectors [18]. These word vectors are then calculation is performed for each attribute pair in the user
averaged as the mean vector to get a single vector, represent- profiles, resulting in a similarity score for each pair. The
ing the attribute. For the Out-Of-Vocabulary (OOV) issue, calculation of the semantic similarity between any two sin-
words that do not exist in the vocabulary of the model are gle vectors 𝑣𝑒𝑐 1 and 𝑣𝑒𝑐 2 , representing attributes from user
ignored. In the case of multi-word attributes, such as “data profiles, is defined by Equation 3:
analysis”, the average of the vectors of individual words is
ˆ 1, 𝑣𝑒𝑐
𝑐𝑜𝑠𝑖𝑛𝑒_𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦_𝑠𝑐𝑜𝑟𝑒 (𝑣𝑒𝑐 ˆ 2) =
computed to represent the entire attribute. In the case of
word order alternation, the meaning between two terms is 𝑣𝑒𝑐ˆ 1 · 𝑣𝑒𝑐ˆ2
(3)
considered to be the same. For instance, the terms “infor- ˆ 1 || 2 × ||𝑣𝑒𝑐
||𝑣𝑒𝑐 ˆ 2 || 2
mation system” and “system information” have the same where 𝑣𝑒𝑐 ˆ 1 and 𝑣𝑒𝑐 ˆ 2 are the vector representations of two
meaning. Therefore, both terms have the same meaning in attributes, 𝑣𝑒𝑐 ˆ 1 ·𝑣𝑒𝑐 ˆ 1 || 2
ˆ 2 is the dot product of the vectors, ||𝑣𝑒𝑐
context. The text-based embedding is defined by Equation 1: denotes the L2 norm of the vector representation of 𝑣𝑒𝑐 ˆ 1 , and
𝑛 ˆ 2 || 2 denotes the L2 norm of the vector representation of
||𝑣𝑒𝑐

1 ∑︁  
ˆ 2 . The cosine similarity value ranges from -1 to 1, where
ˆ 𝐴𝑖 𝑗 =
𝑣𝑒𝑐 𝑣𝑒𝑐 𝑤𝑖𝑚𝑗 for 𝑗𝜖 {1, 2, 3} (1) 𝑣𝑒𝑐
𝑛 a value of 1 indicates perfect similarity and a value of -1
𝑚=1,𝑤𝑖𝑚𝑗 𝜖𝐴𝑖 𝑗
indicates complete dissimilarity.


where 𝑣𝑒𝑐ˆ 𝐴𝑖 𝑗 is a single vector representation • Similarity Score Aggregation: To combine different at-
  of each
𝑚
attribute, 𝑛 is the number of words, and 𝑣𝑒𝑐 𝑤𝑖 𝑗 is a vector tributes of a user profile, this step calculates an aggregated
similarity score for two user profiles by taking a simple aver-
representation of a word.
age of all the cosine similarity scores obtained in the previous
• Context-based Embedding: This embedding model focuses
step across their multiple attributes. The simple average is
on the biography attribute in a user profile. The biography
calculated by summing up all the similarity scores and then
attribute is directly converted into a single vector repre-
dividing by the number of scores. This gives equal impor-
sentation using a pre-trained Bidirectional Encoder Repre-
tance to every score. The score aggregation is calculated
sentations from Transformers (BERT), a transformer-based
using Equation 4:
model [6]. This model is capable of understanding the con-
text and semantic meanings of texts in an effective way. 𝑎𝑔𝑔𝑟𝑒𝑔𝑎𝑡𝑒𝑑_𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦_𝑠𝑐𝑜𝑟𝑒 (𝑈 1, 𝑈 2 ) =
Instead of aggregating word vectors, this model considers Í𝑞
the entire text as input and generates comprehensive embed- 𝑗=1 𝑐𝑜𝑠𝑖𝑛𝑒_𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦_𝑠𝑐𝑜𝑟𝑒 𝑗
(4)
dings that capture the overall semantic content. This ensures 𝑞
that the vector representation captures not only individual where 𝑐𝑜𝑠𝑖𝑛𝑒_𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦_𝑠𝑐𝑜𝑟𝑒 𝑗 is the cosine similarity of
word meanings but also the overall context and theme of the the 𝑗 𝑡ℎ attribute, and 𝑞 is the total number of attributes.
biography attribute. The context-based embedding is defined • Matrix Insertion: The aggregated similarity score of two
by Equation 2: users is inserted into a pairwise similarity matrix. This is
a square matrix that stores the similarity scores between
ˆ (𝐴𝑖4 ) = 𝐵𝐸𝑅𝑇𝑤𝑖4𝑚 𝜖𝐴𝑖4 (𝐴𝑖4 )
𝑣𝑒𝑐 (2)
pairs of users. The matrix is of size 𝑁 × 𝑁 . The 𝑁 is the total
ˆ (𝐴𝑖4 ) is a single vector representation of the biog-
where 𝑣𝑒𝑐 number of users. Each row and column represent a user, and
raphy attribute. each entry represents the similarity score between two users.
Finally, the embedded vector representations are utilized to com- The pairwise similarity matrix serves as a foundation for building
pute the semantic similarity scores between different user profiles a weighted graph model, enabling further analysis and visualization
in the next part. of the relationships between different users.
3.2.2 Graph Building. A weighted graph, 𝐺 = (𝑉 , 𝐸, 𝑆), where
3.2 Graph Representation 𝑉 is the set of nodes, 𝐸 is the set of edges, and 𝑊 is the set of
This part uses a pairwise similarity matrix to construct a weighted weights, is built from the pairwise similarity matrix. Each node 𝑉ℎ
graph model that represents the semantic relationships between in the graph represents a user. If the aggregated similarity score
different users. The edges between nodes indicate their scores for between two nodes exceeds a certain threshold set to 0.70, they
semantic similarity. The graph representation consists of two main are connected with an edge. The weight 𝑆ℎ of the edge 𝐸ℎ is then
components, which are described in the following sections. set by an aggregated similarity score, which indicates the strength

322
Enhancing a User Matchmaking Algorithm using Personalized PageRank NLPIR 2023, December 15–17, 2023, Seoul, Republic of Korea

of the relationship between them. A high score refers to a strong 3.3.2 Ranking. After convergence, the component sorts the user
relationship, and a weak relationship has a low score. The edge indices based on their PageRank scores for the target user in de-
weighting is defined by Equation 5: scending order. The top-𝑘 users are the users who are most similar
to the target user, where 𝑘 is a threshold or specific number.

By incorporating the semantic similarity of users into textual
𝑎𝑔𝑔𝑟𝑒𝑔𝑎𝑡𝑒𝑑_𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡 𝑦_𝑠𝑐𝑜𝑟𝑒 (𝑈 1 , 𝑈 2 ) , if 𝑎𝑔𝑔𝑟𝑒𝑔𝑎𝑡𝑒𝑑_𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡 𝑦_𝑠𝑐𝑜𝑟𝑒 > 𝜏 data, it is possible to obtain an in-depth understanding of user
𝑆ℎ =
No 𝑎𝑔𝑔𝑟𝑒𝑔𝑎𝑡𝑒𝑑_𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡 𝑦_𝑠𝑐𝑜𝑟𝑒, Otherwise
(5)
preferences, interests, and traits. This approach can enhance the
where 𝜏 is a predefined threshold. accuracy and effectiveness of personalized user matchmaking by
considering not only social structures and behaviors but also the
content shared by users. Additionally, integrating personalized
3.3 User Matchmaking graph-based approaches into matchmaking algorithms can further
This part enables the PPR algorithm to take into account the pref- improve the quality of personalized user matchmaking by leverag-
erences and interests of users to generate a list of matching users. ing the semantic relationships between users within a graph.
Each user is assigned a ranking based on their connections and
similarity scores. The user matchmaking algorithm consists of two
main components, which are described in the following sections. 4 EXPERIMENTS
This section describes a series of experiments conducted on a simu-
3.3.1 Personalized PageRank. The PPR algorithm is a graph-based lated user profile dataset to compare the performance of the pro-
algorithm that assigns a score to each node in a graph based on posed algorithm to existing state-of-the-art methods. The primary
their connections and importance. This score reflects the influence objective of the experimental design is to address the following
of each node in the graph. In the context of user matchmaking, Research Questions (RQs):
the PPR algorithm can be used to identify the most influential
users based on their similarity to a target user. The PPR algorithm • RQ1: Can the proposed algorithm outperform state-of-the-art
takes into account the relevance of each connection and similarity graph methods?
score, ensuring that users with more meaningful connections are • RQ2: Which graph-based method has better performance,
ranked higher. The PPR algorithm modifies the original PageRank and why is that?
algorithm slightly to produce personalized matchmaking for users.
Instead of jumping to any random node, the algorithm biases the The following subsections detail the experimental procedures, in-
random walk towards a specific starting node or a set of nodes cluding dataset descriptions, evaluation metrics, baseline methods,
based on a personalization vector. This provides a set of nodes and experimental settings.
that are reachable from the starting node. The steps of the PPR
algorithm for user matchmaking are explained below.
4.1 Dataset Description
• Step 1: Start with a vector 𝑟 where the entry corresponding
User profiles typically contain private and sensitive information
to the target user of interest is set to 1, and all others are set
about individuals, making it difficult to collect and access this in-
to 0.
formation without their consent. To address this issue, AI text
• Step 2: The equation 6 for updating the PageRank vector
generation systems, such as ChatGPT [12] and Google Bard [16],
𝑟 ′ in each iteration can be represented as:
are used to generate a user profile dataset that simulates human
language autonomously. The simulated user profile dataset includes
𝑟 ′ = (1 − 𝑑) × 𝑀 × 𝑟 + 𝑑 × 𝑣 (6) 150 user profiles with various attributes, including names, interests,
hobbies, occupations, and biographies. The attributes of interests,
where 𝑑 is the damping factor, typically set to 0.85, 𝑀 is the hobbies, occupations, and skills are typically described in the form
transition matrix derived from the weighted graph, and 𝑣 is of controlled vocabulary. On the other hand, a biography attribute
a personalization vector which represents the probability of is a short text written in natural language format that provides
jumping to the node corresponding to the target user. a background summary of a user. Lastly, this dataset serves as a
• Step 3: This process must be repeated until it converges. benchmark test for user matchmaking algorithms. By using this
• Step 4: After convergence, the resulting 𝑟 gives a ranking dataset, researchers can evaluate the performance of their algo-
of higher scores to users that are reachable from the target rithms without having to collect and access personal information.
user. The simulated user profile dataset is randomly divided into two
sets: a training set and a test set. The training set is used to con-
The PPR algorithm is an efficient method for matching users struct a graph structure and learn features of the proposed algo-
based on their similarity to a target user. The algorithm starts by rithm, while the test set is used to evaluate the performance of the
setting the entry corresponding to the target user to 1 in the starting proposed algorithm. The training set contains 100% of the users
vector 𝑟 . It then updates the vector iteratively until convergence. in the whole dataset, while the test set contains 20% of the users
The algorithm assigns higher scores to users who are reachable in the whole dataset selected using a random sampling technique.
from the target user. This ranking helps identify potential matches Meanwhile, the test set is prepared as a ground-truth dataset, where
based on their connectivity and relevance to the target user. human judgment is used to assign matches manually.

323
NLPIR 2023, December 15–17, 2023, Seoul, Republic of Korea Santipong and Herwig

4.2 Evaluation Metrics • Random Walk with Weighted Edges [1]: This method gen-
The accuracy of the proposed algorithm and state-of-the-art meth- erates random paths in a graph, where the probability of mov-
ods is measured using common evaluation metrics for recommender ing from one node to another depends on the edge weights.
systems [25]. This includes precision at 𝑘 denoted as 𝑃@𝑘, recall The basic idea is that it starts at a node in the graph and then
at 𝑘 denoted as 𝑅@𝑘, F1 score at 𝑘 denoted as 𝐹 1@𝑘, and mean av- randomly chooses one of its neighbors to move to, according
erage precision at 𝑘 denoted as 𝑀𝐴𝑃@𝑘. The following evaluation to the weights of the edges.
metrics are described below: • SimRank [28]: This is a graph-based method for calculating
the similarity between two nodes. The similarity of two
• Precision@k: This metric measures the accuracy of the top-
nodes is determined by the similarity of the nodes that they
𝑘 users by computing the fraction of relevant users that are
are connected to. In other words, two nodes are considered
found in the top-𝑘 users. A high precision indicates that
similar if they are connected to similar nodes. The nodes
most of the top-𝑘 users are matched, while a low precision
with the highest SimRank scores are considered to be the
indicates that many irrelevant users are matched among the
most similar to the starting node.
top-𝑘 users. The precision at 𝑘 is defined by Equation 7 as
• Girvan-Newman Algorithm [24]: This method detects com-
follows:
munities or clusters, which are groups of nodes that are more
Number of Relevant Users in Top-𝑘 Matches
𝑃@𝑘 = (7) densely connected to each other than to the rest of the graph.
𝑘 The method works by iteratively removing the edges with
• Recall@k: This metric measures the fraction of relevant the highest betweenness centrality, which measures how
users that are found in the top-𝑘 users. A high recall indicates frequently an edge lies on the shortest path between any
that most of the relevant users are found among the top-𝑘 two nodes.
users. The recall at 𝑘 is defined by Equation 8 as follows: • Louvain [9]: This method detects communities in a graph.
Number of Relevant Users in Top-𝑘 Matches It maximizes a modularity score for each community, where
𝑅@𝑘 = (8)
Total Number of Relevant Users modularity measures the quality of assigning nodes to com-
• F1 Score@k: This is a weighted average of precision and munities.
recall at 𝑘 that balances both metrics into a single number. • Node2Vec [26]: This method embeds nodes in a graph into
It ranges from 0 to 1, where higher values indicate better low-dimensional vector representations that preserve neigh-
performance. Equation 9 is defined as follows: borhood relationships, such that nodes that are similar in
the graph are also similar in the embedding space. It uses
2 × 𝑃@𝑘 × 𝑅@𝑘
𝐹 1@𝑘 = (9) skip-gram models to learn node embeddings based on the
𝑃@𝑘 + 𝑅@𝑘
context of the random walks and a random walk strategy to
• Mean Average Precision@k: This metric provides an overview find structural similarities between nodes in the graph.
of the quality of the ranking algorithm, which is the mean
Girvan-Newman and Louvain are used to partition groups of sim-
of the Average Precisions (APs) across all users. The mean
ilar users. Consequently, users from the same cluster are matched,
of the AP score for each user and the MAP are defined by
with the underlying assumption that users in the same group as
Equations 10, 11, and 12, respectively, as follows:
the target user are likely to be similar.
Í |𝑈 |
𝐴𝑃@𝑘𝑢
𝑀𝐴𝑃@𝑘 = 𝑢=1 (10) 4.4 Experimental Settings
|𝑈 |
Í𝐾
𝑃@𝑘 × 𝑟𝑒𝑙𝑘 Extensive experiments were conducted on the simulated user profile
𝑘=1 dataset, which imitates real-world user preferences and characteris-
𝐴𝑃@𝑘 = (11)
Total Number of Relevant Users tics, ensuring the validity of the experimental results. The proposed

1, If a user at 𝑘 rank is relevant algorithm was evaluated against existing state-of-the-art methods
𝑟𝑒𝑙𝑘 = (12)
0, Otherwise using both the F1 and MAP metrics for the top 5, 10, 15, and 20
where 𝐴𝑃@𝐾 is the average precision at 𝑘, 𝑟𝑒𝑙𝑘 is a binary candidate users.
indicator that is 1 if the user at rank 𝑘 is relevant and 0 All experiments were written in the Python programming lan-
otherwise, 𝑃@𝑘 is the precision at rank 𝑘, |𝐾 | is the number guage and executed on a Personal Computer (PC) with an Intel (R)
of relevant users, and |𝑈 | is the number of users. Core (TM) i5-4570 CPU at 3.20 GHz and 8 GB of DDR3 RAM, which
was running on an Ubuntu 22.04 LTS Linux server. Each method
4.3 Baseline Methods was implemented and tested under the same experimental condi-
To demonstrate the effectiveness of the proposed algorithm, exist- tions for a fair comparison. The hyperparameters of each method
ing state-of-the-art methods, considering graph-based approaches, were safely set to default values according to the NetworkX library
were compared. These methods are briefly described below: [13], aiming to maintain consistency and eliminate any potential
• Random Walk [7]: This method simulates random paths in a bias in the comparison of the different methods.
graph by following some rules. The basic idea is to start at a
node in the graph and then choose one of its neighbors to 5 EXPERIMENTAL RESULTS AND DISCUSSION
move to. This process is repeated until a certain number of The experimental results on the simulated user profile dataset are
steps are taken or a certain condition is met. reported in Table 2 to address RQ1 and RQ2, which are concerned

324
Enhancing a User Matchmaking Algorithm using Personalized PageRank NLPIR 2023, December 15–17, 2023, Seoul, Republic of Korea

with the effectiveness of the proposed algorithm in predicting user means that the findings of this study may need to be more general-
matches. The best results are in boldface, and the second-best results izable to real-world scenarios. Additionally, the study considered a
are underlined. relatively small dataset, which limits the scalability of the methods.
Table 1 shows that the proposed algorithm significantly outper- Future work could involve scalability testing on larger datasets
forms all state-of-the-art methods for user matchmaking in terms for comprehensive evaluations. Additionally, deeper dives into the
of F1 and MAP metrics across the top 5, 10, 15, and 20 candidate hyperparameters of personalized PageRank, such as the number
users. The proposed algorithm achieves an average improvement of iterations, could provide further improvements. It is essential to
of more than 8% over the best baseline method. The second-best tune these learning parameters carefully to achieve the best results.
method is Node2Vec, which uses random walks to embed nodes
in a low-dimensional vector space and a skip-gram model to learn
their representations. This method is able to capture the structural 7 OPPORTUNITIES AND CHALLENGES
relationships between nodes in a graph. The third-best method is User matchmaking systems have become key components of many
SimRank, which measures the similarity between two nodes based social media platforms and online services. Their primary goal is to
on their structural context in the graph. This method does not con- connect users based on shared interests, habits, or other contextual
sider the weights of the edges between nodes. The worst methods information, offering matching friends or users that meet individual
are Random Walk, Girvan-Newman, and Louvain. These methods needs. However, these systems also face some opportunities and
do not consider edge weights in their computations. Random Walk challenges.
is a simple method that selects one adjacent node at random and User matchmaking systems have the potential to significantly
assigns equal weights to each edge. Girvan-Newman and Louvain enhance effectiveness and accuracy by integrating contextual infor-
focus primarily on detecting communities or clusters within graphs. mation about users, such as geographical location, social connec-
Their goal is to partition a graph into clusters, which is not directly tions, and behavioral data. By leveraging this information, these
optimized for finding similar users. systems significantly impact the effectiveness of user matchmak-
There are several reasons for the higher performance of the pro- ing, leading to increased user satisfaction and retention. Intelligent
posed algorithm. Firstly, the proposed algorithm based on the PPR matchmaking systems should consider various factors and utilize
method assigns weights to each edge, representing semantic rela- machine learning and data analysis techniques to provide more
tionships between users in the graph. This allows it to capture the accurate and personalized matches by learning and adapting user
relationships between users based on their profiles more accurately. data over time. However, integrating these multiple data points
For example, if two users have similar interests, the PPR method into a user matchmaking algorithm requires advanced techniques
assigns a higher weight to the edge between them. Secondly, PPR and a deep understanding of human behavior.
introduces a personalized aspect to the ranking mechanism. This Another challenge is that user matchmaking systems often rely
means it is designed to prioritize nodes based on a particular source on sensitive and behavioral information that users do not want
node, ensuring the results are more tailored and relevant. For exam- to reveal. This raises concerns about user privacy and security.
ple, if a user is looking for friends who are interested in hiking, the To protect against information leakage, these systems should im-
PPR method will prioritize nodes connected to the user node with a plement privacy and security measures to ensure that user data
high weight for hiking interests. Lastly, this personalization factor is protected and anonymized. This includes processing raw data
ensures that a traversal of the graph does not travel too far from the locally instead of uploading it to servers and having clear and easy-
starting node. This characteristic is especially valuable in contexts to-understand privacy policies that explain what data is collected,
where the graph is large and the goal is to find nodes similar to a how it is used, and who it is shared with. Additionally, providing
particular starting node. For example, if a user is looking for friends users with control over the information they share can help build
in a large city, the PPR method will prioritize nodes that are close trust and encourage participation in user matchmaking systems. For
to the user. example, users should be able to choose who can see their profile
In summary, the experimental results indicate that the proposed or what information is visible to others. Finally, these systems must
algorithm is significantly effective in predicting users who are follow data protection laws and regulations, such as the General
similar to the target user in terms of preferences, interests, and traits. Data Protection Regulation (GDPR) in the European Union or the
This is because the algorithm uses both edge weights calculated by California Consumer Privacy Act (CCPA) in the United States. They
text embedding techniques and personalized nodes in the random must also take steps to prevent the spread of misinformation and
walk process. Text embedding techniques allow the algorithm to moderate content to guarantee user safety. Implementing strong
capture the semantic similarities between nodes, while personalized security measures, such as encryption and secure authentication
nodes allow the algorithm to capture the preferences and interests protocols, is also crucial to safeguarding user data and ensuring
of the target user. Therefore, the proposed algorithm is a suitable their privacy. Additionally, regularly updating and auditing these
and effective solution for user matchmaking on a variety of real- systems can help identify and address any potential vulnerabilities
world online platforms. that could compromise user information.
Last but not least, new users may also face challenges in building
their profiles and establishing their preferences within the platform.
6 LIMITATIONS Without sufficient data, matchmaking algorithms may struggle to
One major limitation of this evaluation is the lack of public or accurately suggest compatible matches for these users. However,
real-world datasets suitable for evaluating user matchmaking. This implementing strategies such as employing hybrid recommender

325
NLPIR 2023, December 15–17, 2023, Seoul, Republic of Korea Santipong and Herwig

Table 2: Experimental Results with F1 and MAP Metrics Comparing with Classical Methods

F1 Score MAP
Methods
F1@5 F1@10 F1@15 F1@20 MAP@5 MAP@10 MAP@15 MAP@20
Random walk 0.204 0.243 0.184 0.232 0.146 0.230 0.183 0.242
Random walk w/ weighted edges 0.252 0.296 0.271 0.265 0.188 0.270 0.282 0.273
SimRank 0.373 0.493 0.533 0.493 0.245 0.374 0.474 0.512
Girvan-newman algorithm 0.045 0.047 0.051 0.100 0.020 0.023 0.013 0.028
Louvian 0.406 0.480 0.479 0.429 0.269 0.399 0.444 0.462
Node2Vec 0.542 0.580 0.557 0.513 0.431 0.582 0.668 0.696
The proposed algorithm 0.543 0.664 0.626 0.543 0.432 0.654 0.742 0.764
Improvement (%) 0.184 14.482 12.387 5.847 0.232 12.371 11.077 9.770

systems to make initial recommendations based on similar user [3] Liang Chen, Yuanzhen Xie, Zibin Zheng, Huayou Zheng, and Jingdun Xie. 2020.
profiles and item characteristics or providing incentives for new Friend Recommendation Based on Multi-Social Graph Convolutional Network.
IEEE Access 8 (2020), 43618–43629. https://doi.org/10.1109/ACCESS.2020.2977407
users to actively engage with the platform can help overcome these [4] Ali Choumane and Zein Al Abidin Ibrahim. 2020. Friend Recommendation
challenges and improve the overall user experience. based on Hashtags Analysis. CoRR abs/2003.03531 (2020). arXiv:2003.03531
https://arxiv.org/abs/2003.03531
[5] Indrakant Dana, Udit Agarwal, Akshat Ajay, Saurabh Rastogi, and Ahmed
Alkhayyat. 2023. Recommendation Mechanism to Forge Connections Between
8 CONCLUSION Users with Similar Interests. In International Conference on Innovative Computing
and Communications, Aboul Ella Hassanien, Oscar Castillo, Sameer Anand, and
This paper introduces a new perspective on user matchmaking Ajay Jaiswal (Eds.). Springer Nature Singapore, Singapore, 455–470.
strategies by proposing a user-based matchmaking algorithm that [6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert:
uses personalized PageRank. This algorithm aims to identify po- Pre-training of deep bidirectional transformers for language understanding. arXiv
preprint arXiv:1810.04805 (2018).
tential users with similar interests, preferences, and traits based on [7] Magdalini Eirinaki, Jerry Gao, Iraklis Varlamis, and Konstantinos Tserpes. 2018.
semantic analysis and graph analytics. The personalization feature Recommender systems for large-scale social networks: A review of challenges
of the graph allows the algorithm to specify the number of neigh- and solutions. , 413–418 pages.
[8] Runa Ganguli, Akash Mehta, Narayan Debnath, Sultan Aljahdali, and Soumya
boring nodes surrounding a source node. The algorithm also uses Sen. 2020. An Integrated Framework for Friend Recommender System based on
advanced natural language processing techniques to gain a deeper Graph Theoretic Approach. In Proceedings of 35th International Conference on
Computers and Their Applications (EPiC Series in Computing, Vol. 69), Gordon Lee
understanding of textual user profiles. This helps improve the ac- and Ying Jin (Eds.). EasyChair, 242–255. https://doi.org/10.29007/4bwn
curacy of the matchmaking process. Extensive experiments on a [9] Fabio Gasparetti, Giuseppe Sansonetti, and Alessandro Micarelli. 2021. Commu-
simulated user profile dataset demonstrate that the proposed algo- nity detection in social recommender systems: a survey. Applied Intelligence 51
(2021), 3975–3995.
rithm consistently outperforms existing state-of-the-art methods. [10] Way-Siang Goh, Chian-Wen Too, Meei-Hao Hoo, and Kok-Chin Khor. 2023.
This suggests that the proposed algorithm is a promising approach Evaluation of Recommendation Models for Matchmaking. In Intelligent Com-
for user matchmaking. Finally, the proposed user matchmaking puting & Optimization, Pandian Vasant, Gerhard-Wilhelm Weber, José Antonio
Marmolejo-Saucedo, Elias Munapo, and J. Joshua Thomas (Eds.). Springer Inter-
algorithm can be implemented in several domains, including social national Publishing, Cham, 843–852.
networking sites, online dating platforms, and online multiplayer [11] Jibing Gong, Xiaoxia Gao, Yanqing Song, Hong Cheng, and Jingjing Xu. 2016.
Individual Friends Recommendation Based on Random Walk with Restart in
games. In these domains, the algorithm can match users with simi-
Social Networks. In Social Media Processing, Yuming Li, Guoxiong Xiang, Hongfei
lar interests, preferences, and characteristics, improving the overall Lin, and Mingwen Wang (Eds.). Springer Singapore, Singapore, 123–133.
user experience. [12] Roberto Gozalo-Brizuela and Eduardo C. Garrido-Merchan. 2023. ChatGPT
is not all you need. A State of the Art Review of large Generative AI models.
To enhance the effectiveness and accuracy of the proposed algo- arXiv:2301.04655 [cs.LG]
rithms, future research should incorporate additional contextual [13] Aric Hagberg and Drew Conway. 2020. Networkx: Network analysis with python.
information about users. This could include various factors, such Retrieved September 10, 2023 from https://networkx.github.io
[14] Richard Harrison and Michael Thomas. 2009. Identity in online communities:
as time, location, weather, demographic information, behavioral Social networking sites and language learning. International Journal of Emerging
data, historical interactions, and activity. In addition, reinforcement Technologies and Society 7, 2 (2009), 109–124.
[15] Jingda Kang, Juntao Zhang, Wei Song, and Xiandi Yang. 2021. Friend Relationships
learning techniques are used to continuously improve and refine Recommendation Algorithm in Online Education Platform. In Web Information
the process of user matchmaking by collecting both explicit and Systems and Applications, Chunxiao Xing, Xiaoming Fu, Yong Zhang, Guigang
implicit user feedback derived from the interactions and character- Zhang, and Chaolemen Borjigin (Eds.). Springer International Publishing, Cham,
592–604.
istics of users over time. [16] James Manyika. 2022. An overview of Bard: an early experiment with generative
AI. https://ai.google/static/documents/google-about-bard.pdf
[17] Lynn A McFarland and Robert E Ployhart. 2015. Social media: A contextual
REFERENCES framework to guide research and practice. Journal of applied psychology 100, 6
(2015), 1653.
[1] Hakan Bagci and Pinar Karagoz. 2016. Context-aware location recommendation
[18] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient
by using a random walk-based approach. Knowledge and Information Systems 47
Estimation of Word Representations in Vector Space. arXiv:1301.3781 [cs.CL]
(2016), 241–260.
[19] Huansheng Ning, Sahraoui Dhelim, and Nyothiri Aung. 2019. PersoNet: Friend
[2] Lamia Berkani. 2020. A semantic and social-based collaborative rec-
Recommendation System Based on Big-Five Personality Traits and Hybrid Fil-
ommendation of friends in social networks. Software: Practice and
tering. IEEE Transactions on Computational Social Systems 6, 3 (2019), 394–402.
Experience 50, 8 (2020), 1498–1519. https://doi.org/10.1002/spe.2828
https://doi.org/10.1109/TCSS.2019.2903857
arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1002/spe.2828

326
Enhancing a User Matchmaking Algorithm using Personalized PageRank NLPIR 2023, December 15–17, 2023, Seoul, Republic of Korea

[20] Iván Palomares, Carlos Porcel, Luiz Pizzato, Ido Guy, and Enrique Herrera-Viedma. Conference on Ubiquitous Information Technologies and Applications. Springer,
2021. Reciprocal Recommender Systems: Analysis of state-of-art literature, chal- 599–605.
lenges and opportunities towards social recommendation. Information Fusion 69 [25] Daniel Valcarce, Alejandro Bellogín, Javier Parapar, and Pablo Castells. 2020.
(2021), 103–127. Assessing ranking metrics in top-N recommendation. Information Retrieval
[21] Francesco Ricci, Lior Rokach, and Bracha Shapira. 2015. Recommender systems: Journal 23, 4 (June 2020), 411–448. https://doi.org/10.1007/s10791-020-09377-x
introduction and challenges. Recommender systems handbook (2015), 1–34. [26] Janu Verma, Srishti Gupta, Debdoot Mukherjee, and Tanmoy Chakraborty. 2019.
[22] Animesh Chandra Roy and A. S. M. Mofakh Kharul Islam. 2023. Friend Rec- Heterogeneous edge embedding for friend recommendation. In Advances in
ommendation System Based on Heterogeneous Data from Social Network. In Information Retrieval: 41st European Conference on IR Research, ECIR 2019, Cologne,
Proceedings of International Joint Conference on Advances in Computational In- Germany, April 14–18, 2019, Proceedings, Part II 41. Springer, 172–179.
telligence, Mohammad Shorif Uddin and Jagdish Chand Bansal (Eds.). Springer [27] Bu-Xiao Wu, Jing Xiao, and Jie-Min Chen. 2015. Friend Recommendation by User
Nature Singapore, Singapore, 565–580. Similarity Graph Based on Interest in Social Tagging Systems. In Advanced Intel-
[23] Anju Taiwade, Nitish Gupta, Rakesh Tiwari, Shashi Kumar, and Upendra Singh. ligent Computing Theories and Applications, De-Shuang Huang and Kyungsook
2022. Hierarchical K-Means Clustering Method for Friend Recommendation Han (Eds.). Springer International Publishing, Cham, 375–386.
System. In 2022 International Conference on Inventive Computation Technologies [28] Lu Yang, Tao Hong, and Anilkmar Kothalil Gopalakrishnan. 2018. A Framework
(ICICT). 89–95. https://doi.org/10.1109/ICICT54344.2022.9850852 for Recommender System Based on Game Theory in Social Networks. In 2018
[24] Sadriddinov Ilkhomjon Rovshan Ugli, Doo-Soon Park, Daeyoung Kim, Yixuan 10th International Conference on Knowledge and Smart Technology (KST). IEEE,
Yang, Sony Peng, and Sophort Siet. 2022. Movie Recommendation System Using 95–100.
Community Detection Based on the Girvan–Newman Algorithm. In Interna-
tional Conference on Computer Science and its Applications and the International

327