Leveraging recommendations using a multiplex graph database

• Date: 25 October 2024
• Pages: 537-582
• DOI: https://doi.org/10.1108/IJWIS-05-2024-0137
• Published date: 25 October 2024
• Subject Matter: Information & knowledge management,Information & communications technology,Information systems,Library & information science,Information behaviour & retrieval,Metadata,Internet
• Author: Foutse Yuehgoh,Sonia Djebali,Nicolas Travers

Leveraging recommendations using a

multiplex graph database

Foutse Yuehgoh

Research Center, Léonard de Vinci Pôle Universitaire, Paris La Défense, France;

CEDRIC Laboratory, Conservatoire National des Arts et Métiers,

Paris, France and Coexel, Toulon, France

Sonia Djebali

Research Center, Léonard de Vinci Pôle Universitaire, ParisLa Défense, France, and

Nicolas Travers

Research Center, Léonard de Vinci Pôle Universitaire, Paris La Défense, France,

and CEDRIC Laboratory, Conservatoire National des Arts et Métiers,

Paris, France

Abstract

Purpose –By applyingtargeted graph algorithms,the method used by the authors enables effectiveprediction

of user interactionsand thus fulf‌ils the complex requirementsof modern recommender systems.This study sets

a new benchmark for multidimensional recommendation strategies andoffers a path towards more advanced

and user-centri c models.

Design/methodology/approach –Toimprove multidimensional data recommendation systems, multiplex graph

structures are useful to capture various typesof user interactions. This paper presents a novel framework that uses a

graph database to compute and manipulate multiplex graphs. The approach enables f‌lexible dimension management

and increases expressive power through a specialised algebra designed for multiplex graph manipulation.

Findings –The authors comparethe multiplex graph approach with traditionalmatrix methods, in particular

random walk withrestart, and show that the method not only providesdeeper insights into user preferencesby

integratingscores from different layersof the multiplex graph, but also outperformingmatrix-basedapproaches

in most conf‌igurations.The results highlight thepotential of multiplex graphs for developingsophisticated and

customisedrecommender systemsthat signif‌icantly improve bothperformance and explainability.

Originality/value –The study provides a formal specif‌ication of a multiplex graph construction based on

interaction and content-based information; and the study also developed an algebra dedicated to multiplex

graphs, enablingrobust and precise graph manipulations necessary for effective recommendationqueries. The

authors implementthese algebraic operations within the Neo4j graph databasesystem with a thorough analysis

and experimentationwith three different data sets, benchmarkedagainst traditional matrix-basedmethods.

Keywords Multiplex graph, Graph database, Neo4j, Recommendation systems,

Centrality algorithms, Hybrid recommendation

Paper type Research paper

1. Introduction

In today’s digital landscape, users are bombarded with a vast array of information, including

academic papers, news articles, e-books, blog posts and movies. This often leads to information

This work has been supported by COEXEL, with the chair “Artif‌icial Intelligence for Business

Intelligence”at ESILV.

International

Journal of Web

Information

Systems

537

Received14 May 2024

Revised 31 July 2024

29 August2024

Accepted29 August 2024

InternationalJournal of Web

InformationSystems

Vol.20 No. 5, 2024

pp. 537-582

© Emerald Publishing Limited

1744-0084

DOI 10.1108/IJWIS-05-2024-0137

The current issue and full text archive of this journal is available on Emerald Insight at:

https://www.emerald.com/insight/1744-0084.htm

overload (Klingberg, 2009), where f‌inding relevant and engaging content becomes a daunting

task. To overcome this, recommendation systems (RS) (Roy and Dutta, 2022;Xu et al.,2023;

Ma et al.,2024) play a key role in f‌iltering and delivering targeted content that is customised to

user preferences. These systems achieve this by using techniques such as collaborative (Gao

et al., 2023a;Heckel and Ramchandran, 2017) and content-based (Lops et al.,2011,2019)

f‌iltering, or hybrid techniques (Kulkarni et al.,2023;Çano and Morisio, 2017), to enhance user

experience by making content discovery more personalised and eff‌icient. Despite advances in

RS, the dynamics of human interests and the ever-increasing complexity of data pose a major

challenge. Traditional RS, optimised for more static contexts, often struggle to adapt to the ever-

evolving patterns of user interaction (Adomavicius and Tuzhilin, 2005;Liu, 2015;Zhu et al.,

2020). Furthermore, as data volumes grow, many organisations opt for scheduled batch

processing of their recommendation engines (Vin iski et al.,2021), which is effective for

comprehensive data analysis but less suitable for real-time recommendations (Kwieciński et al.,

2023).

RS intersect with information theory in their joint focus on eff‌iciently managing and

extracting meaningful patterns from large data sets for predictive purposes. Core to both

f‌ields are principles like entropy, which measures information content; information gain,

which assesses increase in informational value; and data compression, which reduces the

volume of data while maintaining its utility. These principles are crucial in differentiating

useful “signal”(relevant information) from “noise”(irrelevant content). Therefore, when

developing an RS, it is essential to effectively rely on these principles to ensure the system’s

accuracy and eff‌iciency, as these principles are essential for managing and extracting

meaningful patterns from large data sets, which is crucial for the predictive accuracy of the

RS (Isinkaye et al.,2015;Roy and Dutta, 2022;Xu et al., 2024;Liet al., 2024).

To further enhance the capabilities of RS, researchers have turned to graph-based

methods. Graphs offer a powerful way to represent complex relationships and interactions

between users and items. By modelling these relationships as graphs,RS can better capture

the nuances of user preferences and item similarities (Wang et al., 2021a). Graph-based RS

have shown promising results in improving recommendation accuracy and diversity by

leveraging techniques such as graph neural networks (GNNs) and node embeddings (Jang

et al.,2017;He et al., 2024). Advancing from simple graphs, the concept of multi-layer

graphs, has received signif‌icant attention in recent years. These graphs represent multiple

types of relationships or interactions as separate layers, allowing for a more comprehensive

analysis of data (Boccaletti et al., 2014). This approach enables RS to integrate diverse

sources of information, suchas social connections, item categories and user behaviours, into

a unif‌ied framework.

Traditionalgraph-based systems oftenstruggle with computationalcomplexity, scalability

and the need for accuratedata entry, which can limit theireffectiveness in dynamic and large-

scale environments(Coimbra et al.,2021;Fan, 2022;Spitalas and Tsichlas, 2023). It can also

be challenging for these systems to process large amounts of interconnected data with the

real-time proce ssing requiremen ts needed to make a ccurate recomme ndations. In con trast,

multiplex graphs which extend the concept of multi-layer graphs, offer a more insightful

approach by capturing different types of interactions between the same sets of nodes across

multiple layers, enriching the representation of user preferences and item similarities (Cozzo

et al.,2018). This leads to moreaccurate and diverse recommendations and improvesoverall

quality by ref‌lectingin a better way the complexityof user behaviour.

In addition, graph databases enhance these capabilities by providing a robust infrastructure

for eff‌iciently managing and querying highly interconnected data [1]. Unlike traditional

relational databases, graph databases are specif‌ically designed to handle complex relationships,

IJWIS

20,5

538

such as those requiring high computational cost for similarity calculations in large-scale data (Li

et al.,2024;Khan et al.,2022) and to effectively manage large amount of data. They enable fast

and f‌lexible data operations (Bonifati et al.,2018). Hen ce, the integration of multiplex graphs

and graph databases will signif‌icantly increase the performance of graph-based RS, offering

better scalability, lower computational complexity and a more expressive data model.

This paper introduces a novel approach that leverages Multiplex graphs, a homogeneous

multi-layer, mono-partite graph where different layers represent various types of

relationships among the sameentities. This methodology enables the integration of multiple

dimensions of information and user interactions, offering a more dynamic, f‌lexible and

comprehensive solution compared to traditional single-layer analyses [2]. Multiplex graphs

are particularly suitable for scenarios in which the same set of nodes exists across different

layers, but these nodes may be interconnected by different types of relationships in each

layer. They have proven to be effective when it comes to capturing different types of

interactions in different areas such as e-commerce, biomedicine and urban transportation

(Bielak and Kajdanowicz, 2024;Zhang et al., 2020;Lee et al., 2014;Zhang et al., 2022).

This is particularly useful when analysing phenomena where the same entities are involved

in multiple types of interactions, and allows for consistent and coherent analysis across

different contexts. In contrast, multi-layer graphs with different types of nodes are better

suited for representing systems in which different entities that cannot be directly compared

with each other or are of different nature interact within the same system. The advantage of

multiplex graphs lies in their ability to simplify the complexity of data analysis by

maintaining consistency of nodes across layers, which is crucial for direct comparison and

aggregation of insights across multiple dimensions of interaction. This multi-dimensional

analysis of same entities is particularly effective in RS, where understanding the interplay

between various factors is crucial for making personalisedand relevant recommendations. It

provides a detailed understanding of the multi-layered interactions between the same set of

entities,thus providing more tailored and contextually relevantinsights for recommendations.

As such, our research builds on the foundational work in graph theory (Cozzo et al.,2018;

Boccaletti etal.,2014;Hashemi and Behrouz, 2024)a nd the operational capabiliti es of graph

databases (Negro, 2021;Fernandes and Bernardino, 2018), which demonstrate the potential

for enhanced pattern recognition and recommendation accuracy through multi-dimensional

data integration.

Our approach proposesto mix both multiplex and graph databases to provide a structured

way to manipulate graphdatabases. Thus, our objectives are summarised as follows:

•formalising the construction of a multiplex graph based on both interaction and

content-based information;

•developing an algebra dedicated to multiplex graphs to ease graph manipulations

and recommendation queries;

•implementing these algebraic operations within the Neo4j [3] graph database

system; and

•analysing the impact of our approach on various data sets against traditional matrix-

based methods.

The structure of the work is as follows: Section 2 providesan overview of existing literature

on RS in general and in specif‌icc ontexts. Section 3 outlines o ur methodology, presenting our

framework and modelling techniques along with supporting theoretical concepts such as

theorems,def‌initions and lemmas. Section 4 describesour implementation in detail,including

International

Journal of Web

Information

Systems

539