Graph-Based Personalized Recommendation in Intelligent Educational

Platforms: A Case Study in Engineering Education

Soﬁa Merino Costa

, Rui Pinto

2 a

and Gil Gonc¸alves

2 b

Departamento Engenharia Inform

atica, Faculdade de Engenharia, Universidade do Porto, Porto, Portugal

SYSTEC-ARISE, Faculdade de Engenharia, Universidade do Porto, Porto, Portugal

Keywords:

Graph-Based Recommendation, Knowledge Management System, Education 5.0, Engineering Education.

Abstract:

The fragmentation of digital learning materials in engineering education makes it difﬁcult for students to ﬁnd

relevant content. This paper presents a graph-based recommender system integrated into an intelligent Knowl-

edge Management System (KMS) to support personalized learning. Using Neo4j, the system models users,

learning objects, and semantic relationships to generate contextualized recommendations across dashboard,

module, and Learning Path (LP) views. Its scoring mechanism combines semantic similarity, interaction his-

tory, and graph proximity to provide adaptive, explainable suggestions. A mixed-methods evaluation with

engineering students showed high alignment with user interests and positive perceptions of transparency and

personalization. The system effectively transitioned from fallback to tailored recommendations as user inter-

actions increased. Results highlight the potential of graph-based approaches to improve content relevance,

discovery, and learner engagement in web-based educational platforms, in line with Education 5.0 principles.

1 INTRODUCTION

Engineering students increasingly face the challenge

of navigating vast and complex digital learning

ecosystems. While abundant resources, such as lec-

ture notes, academic papers, and interactive tools, of-

fer rich learning opportunities, they also contribute

to information overload, making it difﬁcult for learn-

ers to identify relevant, trustworthy, and pedagogi-

cally aligned content (Dalkir, 2017). Traditional plat-

forms often provide limited support for this discov-

ery process, relying on static listings, keyword-based

search, and non-personalized navigation (Kopeyev

et al., 2020).

To address this, there is growing interest in in-

tegrating intelligent features into web-based educa-

tional environments. Among these, recommender

systems offer promise by matching content to learn-

ers based on usage patterns, metadata, and learning

objectives (G. M. et al., 2024). However, their ap-

plication in education remains limited, particularly in

terms of explainability, semantic richness, and adapt-

ability to evolving learner needs (Sahu et al., 2024;

Liu et al., 2024a).

https://orcid.org/0000-0002-0345-1208

https://orcid.org/0000-0001-7757-7308

The main contributions of this paper are as fol-

lows:

• A graph-based recommendation approach for

an intelligent Knowledge Management System

(KMS), tailored for educational resource discov-

ery using Neo4j (Webber and Robinson, 2018) to

model learners, content, and pedagogical relation-

ships.

• Multi-context recommendation logic tailored to

user states (e.g., new, returning, active) and inter-

face views (dashboard, module, learning path).

• A mixed-methods evaluation combining precision

metrics and user feedback to assess recommenda-

tion quality and perceived usefulness.

• A discussion on how explainability and contextual

relevance foster learner trust and engagement in

alignment with Education 5.0.

The remainder of this paper is structured as fol-

lows: Section 2 reviews related work on educational

recommender systems and graph-based approaches.

Section 3 presents the system architecture and rec-

ommendation logic. Section 4 details the evaluation

methodology and results. Section 5 discusses the

ﬁndings, limitations, and implications. Finally, Sec-

tion 6 concludes the paper and outlines directions for

future work.

Costa, S. M., Pinto, R. and Gonçalves, G.

Graph-Based Personalized Recommendation in Intelligent Educational Platforms: A Case Study in Engineering Education.

DOI: 10.5220/0013830200003985

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 21st International Conference on Web Information Systems and Technologies (WEBIST 2025), pages 439-446

ISBN: 978-989-758-772-6; ISSN: 2184-3252

439

2 RELATED WORK

Recommender systems are central to intelligent edu-

cational platforms, offering personalized content de-

livery and helping students navigate large volumes

of heterogeneous learning materials. In engineer-

ing education, this need is particularly acute, as stu-

dents must identify relevant resources across com-

plex, interdisciplinary domains, often without struc-

tured guidance (Urdaneta-Ponte et al., 2021).

Traditional recommendation techniques fall into

two main categories: collaborative ﬁltering, which

identiﬁes patterns based on user-item interactions,

and content-based ﬁltering, which relies on seman-

tic features and metadata (Lops et al., 2011; Ricci

et al., 2011). While effective in domains such as

e-commerce, these approaches face limitations in

education due to sparse behavioral data, misalign-

ment with pedagogical goals, and the cold-start prob-

lem (Burke, 2002).

To address these challenges, hybrid systems have

emerged, combining collaborative and content-based

methods. A signiﬁcant advancement is the use of

graph-based representations, where users, resources,

and educational concepts are modeled as intercon-

nected nodes. This structure enables semantic en-

richment, multi-hop reasoning, and explainable rec-

ommendations—essential features for transparency

and trust in learning environments (Markchom et al.,

2023).

Recent research reﬂects this shift toward graph-

enhanced educational recommenders. (Lu and Feng,

2024) combined graph convolutional networks with

user behavior modeling for dynamic adaptation. (Liu

et al., 2024b) used Neo4j to model user-resource rela-

tionships, integrating NLP to enhance semantic query

interpretation(Webber and Robinson, 2018). (Hu

et al., 2021) addressed information overload by ﬁl-

tering resources based on rule-based matching and

user proﬁles. (Imamah et al., 2024) applied Ant

Colony Optimization to personalize Learning Paths

(LPs) based on difﬁculty and learner preferences.

Meanwhile, Skillify (Dhairya et al., 2024) integrated

generative AI features to recommend content based

on learner skill gaps.

These contributions highlight the pedagogical

value of recommender systems, particularly when en-

hanced by graph structures, semantic reasoning, and

adaptive logic. However, key limitations remain:

• Many systems struggle with cold-start scenarios

and sparse data.

• Few support multi-context recommendations

aligned with user state and platform usage.

• Explainability, though feasible via graph-based

logic, is underutilized.

• Alignment with Education 5.0 values—human-

centricity, personalization, and intelligent sup-

port—remains limited.

This work addresses these gaps by introducing a

modular graph-based recommender embedded in an

educational KMS for engineering students. The sys-

tem provides context-aware recommendations across

dashboard, module, and LP views, leveraging seman-

tic links for pedagogical relevance. Its graph-based

foundation also enables future enhancements in ex-

plainability, learner modeling, and adaptive reason-

ing, supporting Education 5.0 principles (Pinto et al.,

2023; Pinto et al., 2024).

3 SYSTEM DESIGN

The platform developed in this work serves as a pro-

totype for an intelligent KMS designed to enhance en-

gineering education. The system is composed of four

main components: I) a frontend interface for user in-

teraction; II) a backend server that handles applica-

tion logic, data management, and API communica-

tion; III) a data layer including a relational database

and a graph-based recommendation system; and IV)

external services such as a semantic search engine and

cloud storage. This modular architecture, as repre-

sented in Figure 1, ensures a seamless user experience

and provides scalability for future growth. Although

the system is not fully adaptive or self-learning, it in-

tegrates foundational intelligent features such as se-

mantic information retrieval and graph-based person-

alization, which represent an important step toward

creating more advanced, intelligent systems.

The recommendation system is a core compo-

nent of the broader intelligent knowledge manage-

ment platform tailored for engineering students. This

platform supports modular content organization, se-

mantic search, LPs, and personalized recommenda-

tions to guide students through relevant learning ma-

terials efﬁciently. To meet these objectives, the archi-

tecture adopts a hybrid multi-database backend, com-

bining PostgreSQL (Obe and Hsu, 2017) for relational

content and user management, Elasticsearch (Konda,

2024) for semantic retrieval, and Neo4j (Webber and

Robinson, 2018) (Community Edition) as a graph-

based recommendation engine. Unlike triple-store

Resource Description Frameworks (RDF), this ap-

proach was designed as a database model (rather than

a data exchange format), and easily handles multiple

relationships of the same type between the same two

nodes.

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

440

Figure 1: Component and Data Flow Diagram.

3.1 Data Model & Synchronization

The platform’s ability to support intelligent fea-

tures such as personalized recommendations, seman-

tic search, and adaptive LPs relies heavily on a ro-

bust and well-structured data model. The choice

to combine relational, search, and graph databases

stems from their complementary strengths. First,

PostgreSQL ensures reliable storage and transactional

consistency of structured data such as user proﬁles,

content metadata, and LP structures. Second, Elas-

ticsearch enables fast and ﬂexible semantic search

across textual resource descriptions, supporting rich

query capabilities beyond keyword matching. Finally,

Neo4j, a property graph database, powers the rec-

ommendation engine by efﬁciently modeling com-

plex, evolving relationships between users, learning

resources, modules, and LPs, enabling context-aware

traversal queries that adapt to learner behaviors.

To integrate data from the relational backend into

the graph, a custom synchronization pipeline is im-

plemented via a Node.js script (Mardan et al., 2018).

This script extracts relevant entities and relationships

from PostgreSQL and incrementally populates the

Neo4j graph. It clears existing data and recreates

nodes for users, resources, modules, LPs, and clas-

siﬁcation nodes such as categories and tags, estab-

lishing their interconnections accordingly. User in-

teractions, such as resource views, module starts,

bookmarks, and completions, are stored relationally

and then represented as explicit Interaction nodes

in the graph, linked to the corresponding User and

Resource, Module, or LearningPath targets. This

design enables rich traversal patterns and weighted

personalization.

Graph-Based Personalized Recommendation in Intelligent Educational Platforms: A Case Study in Engineering Education

441

3.1.1 Neo4j Graph Schema

To support personalized recommendations, the plat-

form incorporates a graph-based model implemented

with Neo4j. This structure enables efﬁcient traversal

of interconnected entities, including users, resources,

modules, and LPs, capturing complex relationships

not easily expressed in relational databases (Figure 2).

Key educational elements are modeled as nodes, such

as:

• User: Learners with proﬁle attributes (e.g., ed-

ucation level, ﬁeld of study, topic interests, pre-

ferred content types, language preferences).

• Resource: Learning materials enriched with

metadata such as categories, tags, type, title, and

description.

• Module: Groupings of related resources repre-

senting standalone instructional units.

• Learning Path: Ordered sequences of modules

guiding structured progression.

• Category, Tag, ResourceType: Classiﬁcation

nodes supporting semantic connections.

• Interaction: User actions (e.g., view, start,

bookmark, complete) linked to targets, with

timestamps and weights.

Figure 2: Graph-based recommendation schema.

Relationships encode these connections, such as

(:Module)-[:HAS RESOURCE]->(:Resource) and

(:LearningPath)-[:HAS MODULE]->(:Module)

for content hierarchy, or

(:User)-[:PERFORMED]->(:Interaction)

and (:Interaction)-[:TARGET]->(:Resource)

for tracking user behavior. Each Interaction

node contains properties such as interaction type,

timestamp, and a computed weight.

The graph structure allows the system to per-

form personalized content recommendations based

on traversals of relationships such as shared tags,

co-interacted resources, or similar LPs. Cypher

queries (He et al., 2022) are used to retrieve relevant

content for each user based on both direct and indi-

rect connections, enabling a more contextualized and

intelligent experience.

3.2 Recommendation and Scoring Logic

The recommendation engine integrates three core

dimensions: Content Similarity (CS), Interaction

Weighting (IW), and Graph Proximity (GP). When a

learner interacts with a resource, module, or LP, the

system traverses the graph to identify candidate con-

tent and computes a ﬁnal recommendation score using

a weighted sum, as represented in Equation 1:

Score = w

× CS +w

× IW +w

× GP (1)

The weights w

, w

, and w

are empirically tuned

to balance semantic, behavioral, and structural sig-

nals.

Content Similarity (CS) captures thematic close-

ness by comparing tags, categories, and meta-

data. The system employs the Sørensen–Dice coef-

ﬁcient (Gragera and Suppakitpaisarn, 2016), imple-

mented via Neo4j’s Awesome Procedures on Cypher

(APOC) plugin (Shatnawi and Saquer, 2024), to com-

pute semantic similarity between resources.

Interaction Weighting (IW) reﬂects the pedagog-

ical signiﬁcance and recency of user actions. Each

interaction type is assigned a base weight (w

base

), ad-

justed by a linear decay function that prioritizes recent

activity, as represented in Equation 2:

= w

base

× max(1, 10− DaysSinceInteraction) (2)

The interaction types and corresponding w

base

val-

ues are: 1 - Default (other interactions); 2 - Re-

source/module/LP views; 3 - Module/LP started; 4 -

Content bookmarked; 5 - Module completed; 6 - LP

completed.

This approach ensures that recent, pedagogically

meaningful actions have greater inﬂuence on recom-

mendations, while older actions still contribute with

diminished weight.

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

442

Graph Proximity (GP) encodes the structural re-

lationship between users and content. Shorter path

lengths in the graph indicate stronger contextual rel-

evance, enabling the system to blend content-based

and collaborative ﬁltering perspectives.

Together, these three dimensions enable the sys-

tem to deliver adaptive, context-aware, and explain-

able recommendations aligned with each learner’s

evolving proﬁle and engagement history.

3.3 Recommendation Query

Formulation

The recommendation engine generates context-

speciﬁc suggestions using tailored Cypher queries ex-

ecuted on the underlying graph. These queries are

designed to balance personalization, semantic rele-

vance, and robustness to sparse data. Core elements

of the query formulation include:

• Interaction-Based Personalization: When

available, the engine prioritizes user interaction

history, excluding previously accessed content

and surfacing items aligned with past engagement

patterns and preferences.

• Proﬁle-Driven Filtering: In the absence of in-

teraction data (e.g., cold-start), recommendations

are derived from explicit proﬁle attributes such as

topic interests and preferred content formats.

• Semantic Matching: Queries incorporate tag

similarity, category alignment, and metadata com-

parisons to surface semantically related content

within the current context (e.g., a viewed module

or resource).

• Popularity Signals: Aggregate usage data across

the platform is used to highlight frequently ac-

cessed or highly rated content, improving diver-

sity and baseline relevance.

• Graph-Based Contextualization: The engine

exploits both direct and indirect relationships

within the graph—such as module-resource links

or shared attributes across LPs—to infer pedagog-

ical and thematic connections.

• Multi-Tier Fallbacks: A hierarchical fallback

strategy ensures continuity in recommendations,

progressing from personalized to proﬁle-based

and ﬁnally to globally popular content as needed.

These dynamic query patterns leverage Neo4j’s

aggregation capabilities to compute ranked recom-

mendations across key platform views, including the

dashboard, resource detail pages, and LP modules.

3.4 Recommendation Contexts and

Pedagogical Impact

The recommendation engine operates across multiple

contexts within the platform, each aligned with a spe-

ciﬁc stage of the learner’s journey:

• Dashboard: Upon login, students receive person-

alized recommendations for resources, modules,

and LPs based on their proﬁle, declared interests,

and recent interactions. This reduces the cogni-

tive load of content discovery and promotes re-

engagement.

• Resource View: When accessing a speciﬁc re-

source, the system suggests semantically related

materials, encouraging exploratory learning and

thematic deepening.

• Module View: Within a module, recommenda-

tions highlight complementary modules or LPs

that support conceptual progression and curricu-

lum continuity.

• Learning Path View: For learners navigating a

structured LP, the system offers reinforcing re-

sources and related modules to strengthen under-

standing and align with learning objectives.

Each context employs a tailored version of the

scoring logic, ensuring that recommendations are

both contextually relevant and pedagogically aligned.

4 EVALUATION

This section presents an evaluation of the intelligent

KMS developed in this work, with a focus on assess-

ing the performance of its personalized recommenda-

tions component. The evaluation was designed to re-

ﬂect realistic usage within the context of engineering

education, by testing the system across all major con-

texts where recommendations appear within the plat-

form, as represented in Table 1.

Table 1: Recommendation contexts across the platform.

Context Description

Dashboard Personalized suggestions based

on proﬁle and interaction data.

Resource Page Related resources and modules

extending the current topic.

Module Page Additional modules and LPs

supporting progression.

LP Page Resources that deepen knowl-

edge in active LPs.

Each context was tested under multiple user states,

namely: I) cold start, i.e., new user with no data

Graph-Based Personalized Recommendation in Intelligent Educational Platforms: A Case Study in Engineering Education

443

regarding preferences (content topics and type); II)

proﬁle-based (static preferences selected by a new

or returning user); and III) personalized (interaction-

aware, after multiple interactions of a returning user),

to evaluate system behavior in both ideal and con-

strained conditions. These controlled conditions al-

lowed for consistent comparison of fallback and per-

sonalized recommendation strategies.

Ten participants, students from computer science

programs, completed a structured sequence of 3 tasks

designed to simulate the system’s intended use. Also,

the system wasn’t used in real courses. The tasks

are: I) Participants registered into the system, selected

their preferences, and conﬁgured their proﬁles. This

task was used to test the cold-start logic of the rec-

ommendation engine and capture proﬁle-based per-

sonalization; II) The new users were asked to interact

with their personalized dashboard. They were encour-

aged to bookmark useful resources, interact with rec-

ommendations, and rate their relevance based on the

submitted preferences at sign-up. This task was used

to test the performance of the Neo4j-based graph rec-

ommendation system and assess the perceived qual-

ity of personalization; III) Participants selected one

of two LPs and completed its modules. Each module

included resources and an assessment to verify knowl-

edge gained.

The system was evaluated using the standard

quantitative metric Precision@k (P@k), which mea-

sures the proportion of relevant items among the top k

recommendations, calculated using Equation 3. This

metric was selected due to its widespread use and suit-

ability for top-k recommendation tasks. High preci-

sion indicates strong immediate relevance.

P@k =

# relevant items in top k

(3)

4.1 Results Summary

The following presents the system’s performance

across the different recommendation scenarios identi-

ﬁed in Table 1. Metrics are reported for each context

to reﬂect how well the system adapts to different user

states and content conﬁgurations.

4.1.1 Dashboard

Table 2 shows the results for the dashboard context,

where user-level personalization is expected to be

most impactful. The system showed strong perfor-

mance in proﬁle-based and personalized scenarios,

i.e., achieving a perfect P@k of 1.00. As expected, no

metrics are reported for cold start users, where fall-

back to popular items was applied instead of person-

alized ranking. These results demonstrate the value

of user modeling: when preferences or interaction

data are available, highly relevant recommendations

are consistently retrieved.

Table 2: Dashboard recommendation relevance by user

state.

State Fallback P@k

Cold Start Popular –

Proﬁle-Based Preferences 1.00

Personalized N/A 1.00

4.1.2 Resource Page

Table 3 presents the results for module recommenda-

tions on resource pages.

Table 3: Resource page module recommendation results.

Resource Type P@k Fallback

With Module 1.00 No

Standalone 1.00 Yes

Even when fallback logic was needed (for stan-

dalone resources), the system maintained high rel-

evance, achieving perfect scores in both contexts.

These results suggest that the semantic similarity

logic and graph-based expansion strategies used for

recommending adjacent modules are effective even in

the absence of direct contextual anchors.

4.1.3 Module Page

Table 4 shows how well the system suggested LPs

when users viewed modules.

Table 4: Module page LP recommendation relevance.

Module Type P@k Fallback

With Path 1.00 No

Without Path 0.67 Yes

P@k dropped to 0.67 for standalone modules due

to fallback recommendations, highlighting some gaps

in topic alignment. These ﬁndings suggest a need for

improved fallback mechanisms in cases where mod-

ules are not yet connected to curated LPs.

4.1.4 Learning Path Page

Table 5 summarizes the relevance of resources rec-

ommended on LP pages. The three evaluated LPs

were: Introduction to Knowledge Management Sys-

tems (LP1), Getting Started with Arduino (LP2), and

Cyber-Physical Systems and Internet of Things (LP3).

While the system returned relevant items for all

LPs, performance varied considerably. Notably, LP1

and LP2 exhibited no unique recommendations, with

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

444

Table 5: LP page resource recommendation results.

LP Items Unique Relevant P@k

LP1 6 0 3 0.50

LP2 6 0 2 0.33

LP3 6 2 4 0.67

multiple duplicate items across paths, resulting in

lower P@k scores of 0.50 and 0.33 respectively. In

contrast, LP3, which contained more distinct and

well-tagged content, achieved better diversity with

two unique recommendations and a higher P@k of

0.67. This variation indicates that while the system

can surface relevant content, there is room to improve

its ability to balance relevance and novelty, particu-

larly when multiple paths are thematically similar.

5 DISCUSSION

The evaluation highlighted both the strengths and

limitations of the recommendation system, showing

how performance is shaped by user context, interac-

tion history, and the structure of the underlying con-

tent graph. In personalized scenarios, particularly

the dashboard view, recommendations consistently

achieved perfect P@k scores when user proﬁle data

and interaction history were available. This validates

the scoring strategy, which combines semantic sim-

ilarity, weighted user interactions, and graph prox-

imity to produce contextually relevant suggestions.

The graph-based design, together with scoring logic,

allowed dynamic adaptation to evolving user states,

transitioning smoothly from fallback suggestions to

personalized results. Users positively noted this pro-

gression, perceiving a clear improvement in recom-

mendation quality over the course of their session.

Despite these strengths, several limitations were

observed. In cold-start scenarios, fallback recommen-

dations relied on content popularity. While this en-

sured a baseline level of relevance, it often failed to

align with users’ speciﬁc interests. Additionally, in

the LP detail view, content overlap was observed be-

tween LPs that addressed different topics. This issue,

particularly between LP1 and LP2, with P@6 scores

of 0.50 and 0.33, reﬂected insufﬁcient differentiation

in resource tagging and semantic descriptors. LP3,

featuring more distinctive and well-annotated content,

performed better (P@6 = 0.67), reinforcing the im-

portance of metadata richness.

These ﬁndings suggest that recommendation qual-

ity is not solely dependent on algorithmic logic, but

also on the breadth, granularity, and semantic qual-

ity of the content graph. Sparse metadata and limited

content variety constrain the system’s ability to gener-

ate diverse or specialized suggestions. Enhancing se-

mantic modeling—via knowledge graph embeddings

or NLP-based descriptors—could improve thematic

sensitivity and recommendation diversity. Moreover,

although the system’s graph structure inherently sup-

ports explainability, the lack of a user-facing explana-

tion layer limits its impact on trust and transparency,

which are critical in educational settings.

In summary, the system’s strong performance

in personalized contexts supports the validity of its

graph-based architecture and scoring logic. However,

the modest results under cold-start and semantically

sparse conditions point to areas for future work, in-

cluding improved metadata enrichment, integration of

explainable interfaces, and scalable strategies for ex-

panding and maintaining the content graph. These en-

hancements are key to deploying intelligent, learner-

centered recommendation systems in real-world edu-

cational environments.

6 CONCLUSION & FUTURE

WORK

This paper presented a graph-based recommender

system integrated into an intelligent KMS for en-

gineering education. Aligned with Education 5.0

principles, the system delivers personalized, context-

aware recommendations across multiple learning

views—including the dashboard, module pages, and

learning paths. Leveraging Neo4j to model users,

content, and semantic relationships, the system dy-

namically adapts to learner proﬁles and activity states,

using a hybrid scoring strategy to provide relevant

suggestions—even during early-stage interactions.

Evaluation results conﬁrmed the system’s ef-

fectiveness in personalized contexts, validating the

multi-factor scoring logic that combines semantic

similarity, interaction history, and graph proximity.

Users experienced a clear progression from fallback

to personalized recommendations during sessions.

However, reduced performance in cold-start and se-

mantically sparse scenarios underscored the need for

richer metadata, broader content coverage, and en-

hanced similarity modeling.

This work contributes a practical, explainable,

and modular recommendation framework for educa-

tional platforms, balancing adaptability with integra-

tion ﬂexibility. Future development will focus on ex-

panding the semantic layer, incorporating real-time

behavioral signals, and implementing visual explana-

tion features to enhance transparency, learner trust,

and long-term engagement in self-directed learning.

Graph-Based Personalized Recommendation in Intelligent Educational Platforms: A Case Study in Engineering Education

445

ACKNOWLEDGEMENTS

The authors gratefully acknowledge the support pro-

vided by the Foundation for Science and Technol-

ogy (FCT/MCTES) within the scope of the Associ-

ated Laboratory ARISE (LA/P/0112/2020), the R&D

Unit SYSTEC through Base (UIDB/00147/2020) and

Programmatic (UIDP/00147/2020) funds

REFERENCES

Burke, R. (2002). Hybrid recommender systems: Survey

and experiments. User modeling and user-adapted in-

teraction, 12(4):331–370.

Dalkir, K. (2017). Knowledge Management in Theory and

Practice. MIT Press, 3rd edition.

Dhairya, Hrishikesh, J., Sonu, G. P., Vaishnavi, D., Sana-

pala, S., and Pallavi, L. (2024). Skillify: En-

hanced Learning Management System Using Genera-

tive AI. In 2024 7th International Conference on Cir-

cuit Power and Computing Technologies (ICCPCT),

volume 1, pages 1527–1532.

G. M., D., Goudar, R. H., Kulkarni, A. A., Rathod, V. N.,

and Hukkeri, G. S. (2024). A digital recommendation

system for personalized learning to enhance online ed-

ucation: A review. IEEE Access, 12:34019–34041.

Gragera, A. and Suppakitpaisarn, V. (2016). Semimetric

properties of sørensen-dice and tversky indexes. In

Kaykobad, M. and Petreschi, R., editors, WALCOM:

Algorithms and Computation, pages 339–350, Cham.

Springer International Publishing.

He, Z., Yu, J., and Guo, B. (2022). Execution time predic-

tion for cypher queries in the neo4j database using a

learning approach. Symmetry, 14(1).

Hu, H., Ma, X., Pan, H., Zhang, H., and Gu, F. (2021). Re-

search on Intelligent Recommendation System Based

on Knowledge Management. In 2021 International

Conference on Intelligent Computing, Automation and

Applications (ICAA), pages 195–198.

Imamah, Laili Yuhana, U., Djunaidy, A., and Purnomo,

M. H. (2024). Development of Dynamic Personalized

Learning Paths Based on Knowledge Preferences and

the Ant Colony Algorithm. IEEE Access, 12:144193–

144207. Conference Name: IEEE Access.

Konda, M. (2024). Elasticsearch in action. Simon and

Schuster.

Kopeyev, Z., Mubarakov, A., Kultan, J., Aimicheva, G., and

Tuyakov, Y. (2020). Using a personalized learning

style and google classroom technology to bridge the

knowledge gap on computer science. International

Journal of Emerging Technologies in Learning (IJET),

15(2):218–229.

Liu, J., Luan, M., and Jiang, N. (2024a). Personalized aca-

demic communication recommendation system based

on knowledge graph. In 2024 5th International Con-

ference on Electronic Communication and Artiﬁcial

Intelligence (ICECAI), pages 610–614.

Liu, J., Luan, M., and Jiang, N. (2024b). Personalized

Academic Communication Recommendation System

Based on Knowledge Graph. In 2024 5th Interna-

tional Conference on Electronic Communication and

Artiﬁcial Intelligence (ICECAI), pages 610–614.

Lops, P., Gemmis, M. d., and Semeraro, G. (2011). Content-

based recommender systems: State of the art and

trends. In Recommender Systems Handbook, pages

73–105. Springer.

Lu, P. and Feng, X. (2024). Personalized Recommenda-

tion Algorithm in AI-Assisted Learning System. In

2024 6th International Conference on Communica-

tions, Information System and Computer Engineering

(CISCE), pages 1289–1294. ISSN: 2833-2423.

Mardan, A., Mardan, and Corrigan (2018). Practical Node.

js. Springer.

Markchom, T., Liang, H., and Ferryman, J. (2023). Re-

view of explainable graph-based recommender sys-

tems. ACM Computing Surveys.

Obe, R. O. and Hsu, L. S. (2017). PostgreSQL: up and run-

ning: a practical guide to the advanced open source

database. ” O’Reilly Media, Inc.”.

Pinto, R., Pinheiro, J., Gonc¸alves, G., and Ribeiro, A.

(2023). Towards industry 5.0: A capacitation ap-

proach for upskilling and technology transfer. In

Mehmood, R., Alves, V., Prac¸a, I., Wikarek, J., Parra-

Dom

ınguez, J., Loukanova, R., de Miguel, I., Pinto,

T., Nunes, R., and Ricca, M., editors, Distributed

Computing and Artiﬁcial Intelligence, Special Ses-

sions I, 20th International Conference, pages 342–

351, Cham. Springer Nature Switzerland.

Pinto, R.,

Zilka, M., Zanoli, T., Kolesnikov, M. V., and

Gonc¸alves, G. (2024). Enabling professionals for in-

dustry 5.0: The self-made programme. In 5th Interna-

tional Conference on Industry 4.0 and Smart Manu-

facturing (ISM 2023), volume 232, pages 2911–2920.

Ricci, F., Rokach, L., and Shapira, B. (2011). Introduc-

tion to recommender systems handbook, pages 1–35.

Springer.

Sahu, S. S., Kumar, R., Sahoo, S., Kumar, B., and Mo-

hanta, P. (2024). Machine Learning-Enabled Inte-

grated Information Platform for Educational Univer-

sities, chapter 2, pages 29–47. John Wiley & Sons,

Ltd.

Shatnawi, H. and Saquer, J. (2024). Encoding feature mod-

els in neo4j graph database. In Proceedings of the

2024 ACM Southeast Conference, ACMSE ’24, page

157–166, New York, NY, USA. Association for Com-

puting Machinery.

Urdaneta-Ponte, M. C., Mendez-Zorrilla, A., and

Oleagordia-Ruiz, I. (2021). Recommendation

systems for education: Systematic review. Electron-

ics, 10(14).

Webber, J. and Robinson, I. (2018). A Programmatic Intro-

duction to Neo4j. Addison-Wesley Professional, 1st

edition.

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

446