Assessing Unfairness in GNN-Based Recommender Systems: A Focus on
Metrics for Demographic Sub-Groups
Nikzad Chizari
1 a
, Keywan Tajfar
2 b
and Mar
´
ıa N. Moreno-Garc
´
ıa
1 c
1
Department of Computer Science and Automation, University of Salamanca, Plaza de los Ca
´
ıdos sn,
37008 Salamanca, Spain
2
College of Science, School of Mathematics, Statistics, and Computer Science, Department of Statistics,
University of Tehran, Tehran, Iran
Keywords:
Recommender Systems, Bias, Fairness, Graph Neural Networks, Metrics.
Abstract:
Recommender Systems (RS) have become a central tool for providing personalized suggestions, yet the grow-
ing complexity of modern methods, such as Graph Neural Networks (GNNs), has introduced new challenges
related to bias and fairness. While these methods excel at capturing intricate relationships between users and
items, they often amplify biases present in the data, leading to discriminatory outcomes especially against
protected demographic groups like gender and age. This study evaluates and measures fairness in GNN-based
RS by investigating the extent of unfairness towards various groups and su bgroups within these systems. By
employing performance metrics like NDCG, this research highlights disparities in recommendation quality
across different demographic groups, emphasizing the importance of accurate, group-level measurement. This
analysis not only sheds light on how these biases manifest but also lays the groundwork for developing more
equitable recommendation systems that ensure fair treatment across all user groups.
1 INTRODUCTION
Recommender systems (RS) are advanced algorithms
that suggest relevant items to users by analyzing their
preferences and behaviors. By examining user-item
interactions, among other data, these systems de-
liver personalized recommendations, helping users
discover new products and services while mitigating
information overload, thus enhancing user engage-
ment (Chen et al., 2020; Zheng and Wang, 2022).
GNN-based RS use Graph Neural Networks
(GNNs) to improve recommendation quality by cap-
turing complex relationships between users and items
in graph-structured data (Zhou et al., 2020; Zhang
et al., 2021). GNNs aggregate information from
neighboring nodes, learning patterns that lead to
more accurate, personalized recommendations, out-
performing traditional methods (Steck et al., 2021;
Khan et al., 2021; Mu, 2018). However, this approach
can unintentionally amplify biases, as the clustering
of similar sensitive attributes in social graphs may re-
sult in biased representations and recommendations,
a
https://orcid.org/0000-0002-7300-6126
b
https://orcid.org/0000-0001-7624-5328
c
https://orcid.org/0000-0003-2809-3707
exacerbating fairness issues (Dai and Wang, 2021).
Sensitive attributes such as race, gender, religion,
age, and disability are protected by privacy laws to
prevent discrimination, making it essential for RS
to consider these factors to avoid biased recommen-
dations under European or US regulations (Di Noia
et al., 2022; Floridi et al., 2022). Failing to do so
can result in significant economic, legal, ethical, and
security risks for both companies and users (Di Noia
et al., 2022; Fahse et al., 2021; Wang et al., 2023).
In response, international organizations have stressed
the need to understand, measure, and mitigate bias,
particularly in sensitive areas, and have implemented
regulations to address these concerns (Di Noia et al.,
2022). In GNN-based RS, bias measurement focuses
on assessing disparities between user groups based on
protected attributes, with this study concentrating on
group fairness and the evaluation of bias toward spe-
cific protected groups.
1.1 Group Fairness
Group fairness ensures that algorithms do not produce
biased predictions or decisions against individuals in
any specific sensitive group. In this section, we dis-
Chizari, N., Tajfar, K. and Moreno-García, M.
Assessing Unfairness in GNN-Based Recommender Systems: A Focus on Metrics for Demographic Sub-Groups.
DOI: 10.5220/0013069400003825
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Conference on Web Information Systems and Technologies (WEBIST 2024), pages 433-440
ISBN: 978-989-758-718-4; ISSN: 2184-3252
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
433

cuss common fairness concepts under group fairness.
Fairness notations under group fairness are described
in Figure 1.
Figure 1: Group fairness notations.
Several key metrics are commonly used to mea-
sure group fairness in GNN-based RS. Demographic
Parity ensures even distribution of recommendations
across groups, with Generalized Demographic Parity
(GDP) addressing continuous sensitive features (Rah-
man et al., 2019; Spinelli et al., 2021). Equality of
Odds and Equality of Opportunity focus on balancing
True Positive and False Positive Rates across groups
(Hardt et al., 2016; Spinelli et al., 2021). Distribution-
based fairness uses metrics like the Wasserstein dis-
tance to maintain fairness in node embeddings, while
model-based fairness makes predicting sensitive fea-
tures difficult (Du et al., 2020; Navarin et al., 2020).
Other fairness measures include Balance Score for
clustering, Maxmin Fairness for influence maximiza-
tion (Dong et al., 2023; Rahmattalabi et al., 2021),
and Disparate Impact (DI) and Treatment Equality
(TE) for comparing outcomes across groups (Purifi-
cato et al., 2022). Additionally, NDCG (Normalized
Discounted Cumulative Gain) is adapted to assess
ranking fairness across demographic groups (Chizari
et al., 2023b).
Traditional bias evaluation metrics are often un-
suitable for RS due to the unique characteristics and
goals of these systems compared to general machine
learning models (Chen et al., 2020; Chizari et al.,
2023a). While typical bias metrics focus on con-
sistent predictions across demographic groups, RS-
specific metrics emphasize identifying biases in rec-
ommended items, reflecting the system’s focus on
predicting user preferences, which naturally vary
across groups (Wang et al., 2022). RS bias evalua-
tion also requires more nuanced approaches, such as
addressing long-tail effects, popularity bias, and per-
sonalization discrepancies. Research by Ekstrand et
al. (2018), Chizari et al. (2022), Chen et al. (2023),
and Gao et al. (2023) underscores the significance
of fairness in enhancing recommendation quality and
user experience.
2 LITERATURE REVIEW
This section provides a review of the state-of-the-art
literature, focusing on the challenges of bias and un-
fairness in RS and specifically GNN-based RS. The
objective is to thoroughly explore and understand the
measurement techniques employed in prior studies,
while also identifying their strengths and weaknesses.
GNN-based RS can enhance accuracy but may
also exacerbate bias and fairness issues due to their
graph structures and message-passing mechanisms
(Steck et al., 2021; Khan et al., 2021; Mu, 2018;
Chizari et al., 2022; Dai and Wang, 2021). For exam-
ple, homophily in social networks, where nodes with
similar sensitive attributes (e.g., age, gender) are more
likely to connect, can lead to biased representations
and unfair outcomes (Dai and Wang, 2021).
To measure fairness in these systems, various
studies have utilized metrics like statistical parity and
classification performance (e.g., NDCG, MRR, AUC)
to assess the influence of sensitive attributes on rec-
ommendations (Rahman et al., 2019; Wu et al., 2020;
Neophytou et al., 2022; Wu et al., 2021). Some
approaches focus on measuring performance across
multiple groups using variance (Rahman et al., 2019)
or the largest gap between groups (Spinelli et al.,
2021). However, these techniques can lead to infor-
mation loss, as variance is sensitive to outliers and
does not reveal which groups are most disadvantaged.
Additionally, focusing only on the two groups with
the highest and lowest accuracy can overlook signifi-
cant disparities among other groups.
The study by (Boratto et al., 2024) examines the
robustness of recommendations in GNN-based RS
from both user and provider perspectives, using De-
mographic Parity (DP) to evaluate group fairness.
It also defines consumer preference and satisfaction
metrics using NDCG and precision. However, this re-
search is limited by the range of models considered
and primarily addresses fairness against poisoning-
like attacks based on edge-level perturbations.
Several studies have explored fairness in RS us-
ing metrics like NDCG, MRR, AUC, and RBP, but
they face challenges such as lack of focus on multi-
ple demographic groups, information loss, and accu-
racy issues. (G
´
omez et al., 2022) focus on traditional
RS without addressing multiple groups, while (Rah-
man et al., 2019; Neophytou et al., 2022; Wu et al.,
2021) note the absence of sensitive attribute combina-
tions. (Spinelli et al., 2021) highlights fairness gaps
but struggles with accuracy, and (Boratto et al., 2024)
limit their analysis to Demographic Parity across a
small model range. Addressing these issues requires
choosing appropriate fairness metrics for GNN-based
RS based on the task and data characteristics.
WEBIST 2024 - 20th International Conference on Web Information Systems and Technologies
434

3 METHODOLOGY
This study evaluates group and subgroup unfairness
in GNN-based RS, focusing on age and gender. Age
is divided into four groups (less than 20, 20 to 40,
40 to 60, and above 60) to assess accuracy consis-
tency across categories. Subgroups like young men,
young women, old men, and old women are also an-
alyzed to identify biases at the intersection of these
attributes. Various performance metrics are applied
to each group and subgroup to assess recommenda-
tion quality. Three real-world datasets are used for a
comprehensive analysis of potential disparities in ac-
curacy.
3.1 Benchmark Datasets
This study uses three well-known real-world datasets,
including MovieLens (Harper and Konstan, 2019)
100K, LastFM 100K (Celma, 2010), and Book Rec-
ommendation (Mobius, 2020). These datasets have
been the main sources of the majority of research in
this section.
3.2 Recommendation Approaches
This study uses various approaches with different
models including Collaborative Filtering (CF), Matrix
Factorization (MF), and GNN-based models:
Table 1: Model Categories and References.
Category Model References
CF ItemKNN (Item K Near-
est Neighbour)
(Al-Ghamdi et al.,
2021; Airen and
Agrawal, 2022)
NNCF (Neural Network
Collaborative Filtering)
(Sang et al., 2021;
Girsang et al., 2021)
MF DMF (Deep Matrix Fac-
torization)
(Xue et al., 2017; Yi
et al., 2019; Liang
et al., 2022)
NeuMF (Neural Collabo-
rative Filtering)
(Kuang et al., 2021;
Zhang et al., 2016)
GNN-
based
LightGCN (Light Graph
Convolutional Network)
(Broman, 2021;
Ding et al., 2022)
NGCF (Neural Graph
Collaborative Filtering)
(Wang et al., 2021;
Sun et al., 2021)
SGL (Self-Supervised
Graph Learning)
(Yang, 2022; Tang
et al., 2021)
DGCF (Disentangled
Graph Collaborative
Filtering)
(Bourhim et al.,
2022; Sha et al.,
2021)
3.3 Evaluation Metrics
Two types of metrics, evaluation and fairness, are
used to assess model performance. This approach
helps evaluate both model accuracy and its behav-
ior toward protected groups. All metrics are applied
based on the top-K ranked items in the recommenda-
tion list (K represents the list size).
3.3.1 Model Evaluation Metrics
The metrics for evaluating the top K recommendation
lists are defined below using the notation described in
Table 2.
Table 2: Table of notations.
Notation Definition
U Set of users
I Set of items
u User
i Item
R(u) Ground-truth set of items that user u in-
teracted with
ˆ
R(u) Ranked item list produced by a model
K Length of the recommendation list
• Recall@K: Recall is a measure that computes the
fraction of relevant items out of all relevant items
on the top-K list.
Recall@K =
1
|U|
∑
u∈U
|
ˆ
R(u) ∩ R(u)|
|R(u)|
(1)
• Precision@K: Precision or positive predictive
value is a measure that calculates the fraction of
relevant items out of all the recommended items
on the list. The average is calculated for each user
to gather the final result. |
ˆ
R(u)| denotes the item
count of
ˆ
R(u).
Precision@K =
1
|U|
∑
u∈U
|
ˆ
R(u) ∩ R(u)|
|
ˆ
R(u)|
(2)
• Mean Reciprocal Rank (MRR)@K: The MRR
computes the corresponding rank of the first rele-
vant item found in the top-k list. Rank
∗
u
represents
be the position of that item in the list provided by
a given algorithm for the user u.
MRR@K =
1
|U|
∑
u∈U
1
Rank
∗
u
(3)
• Hit Ratio (HR)@K: HR or Hit calculates how
many ‘hits’ are in a top-K recommendation list.
HR requires at least one item that falls in the
ground-truth set. δ(0) is an indicator function.
δ(b) = 1 if b is true; otherwise it would be 0.
/
0
denotes the empty set.
HR@K =
1
|U|
∑
u∈U
δ(
ˆ
R(u) ∩ R(u) ̸=
/
0) (4)
• Normalized Discounted Cumulative Gain
(NDCG)@K: NDCG is a measure of ranking
quality, where positions are discounted loga-
rithmically. It accounts for the position of the
Assessing Unfairness in GNN-Based Recommender Systems: A Focus on Metrics for Demographic Sub-Groups
435

hit by assigning higher scores to hits at the top
ranks. NDCG is the ratio between DCG and
the maximum possible DCG.δ(0) is an indicator
function.
NDCG@K =
1
|U|
∑
u∈U
1
∑
min(|R(u)|,K)
i=1
1
log
2
(i+1)
·
K
∑
i=1
δ(i ∈ R(u))
1
log
2
(i + 1)
(5)
4 RESULTS
This section presents the experimental results from
three real-world datasets, along with a brief EDA to
illustrate user distribution within each group. It first
displays performance results for each group, followed
by a detailed analysis of the subgroups.
4.1 Exploratory Data Analysis (EDA)
This section presents an exploratory data analysis
(EDA) of user distribution (not interaction) across
groups, focusing on user count rather than interac-
tions. In Figure 2 the gender distribution for Movie-
Lens and LastFM datasets shows a higher number of
male users compared to female users, indicating po-
tential bias in the data that could lead to unfairness.
Figure 3 indicates the age distribution of the users
in the groups for the three datasets. Here also the
charts show an unequal distribution across groups,
with the senior user group being in the minority, at
a great distance from the other groups.
4.2 Results of Overall Performance
In this part, the results of all of the used metrics on
all models across the datasets can be seen. All charts
present the performance of the models on the @10
recommendation list. Figure 4 shows the overall re-
sults without considering any subgroups using differ-
ent metrics on all datasets. It can be seen that On the
LastFM and the Book Recommendation datasets, the
overall performance of GNN models is lower.
4.3 Results of Each Group
In the first part of the experiment, the results of the
NDCG metric are calculated for each group for the
three datasets.
Figure 5 shows the NDCG results for gender on
the MovieLens and LastFM datasets. In the Movie-
Lens dataset, SGL has the lowest performance and
the largest difference between groups, indicating it
(a) MovieLens
(b) LastFM
Figure 2: Gender distribution for MovieLens and LastFM
indicating the bias in data toward men.
is prone to discrimination. Other GNN models per-
form better with minor unfairness, though LightGCN
shows slight bias. In the LastFM dataset, GNN mod-
els generally underperform compared to conventional
methods, with ItemKNN also performing poorly. Sig-
nificant group differences are only observed in SGL
and LightGCN.
Figure 6 also represents the results of NDCG
performance for sensitive attribute age including
teenagers (less than 20), young adults (20 to 40),
adults (40 to 60), and seniors (more than 60) on all
datasets.
In the MovieLens dataset, SGL performed poorly
with significant group differences, while NGCF and
DGCF showed strong performance and low discrimi-
nation. Traditional models had higher unfairness. In
the LastFM, ItemKNN and other GNN models un-
derperformed, with NGCF and DGCF showing more
unfairness. Similarly, in the Book Recommendation
dataset, GNN models performed poorly, with NeuMF,
DGCF, DMF, and NNCF showing higher unfairness,
while NGCF had the lowest unfairness.
WEBIST 2024 - 20th International Conference on Web Information Systems and Technologies
436

(a) MovieLens
(b) LastFM
(c) Book Recommendation
Figure 3: Age distribution for the three datasets showing the
bias in data across the groups.
4.4 Bias and Fairness Results for Each
Subgroup
The second stage analyzes NDCG performance
across subgroups, providing insights into model be-
havior within each subgroup and helping to better un-
derstand unfairness toward protected groups.
(a) MovieLens
(b) LastFM
(c) Book Recommendation
Figure 4: Performance results across all datasets based on
various metrics.
In the MovieLens dataset, NGCF, DGCF, and
DMF performed better for a specific subgroup, pos-
sibly due to user distribution, while other models
showed fair results. In the LastFM dataset, the same
subgroup (women seniors) had the poorest results,
highlighting that a lack of users in certain subgroups
can lead to higher unfairness.
Assessing Unfairness in GNN-Based Recommender Systems: A Focus on Metrics for Demographic Sub-Groups
437

(a) MovieLens
(b) LastFM
Figure 5: Results of NDCG performance for sensitive at-
tribute gender on MovieLens and LastFM.
5 CONCLUSIONS AND FUTURE
WORK
This research highlights the challenges of ensuring
fairness in GNN-based recommender systems (RS)
for protected demographic groups, particularly con-
cerning gender and age. The exploratory data analysis
(EDA) reveals uneven user distributions across groups
in the datasets, contributing to biased outcomes and
potential discrimination.
Model results indicate that considering gender and
age groups leads to lower performance of SGL on
the MovieLens, LastFM, and Book Recommendation
datasets, exacerbating unfairness toward unprotected
groups. In the LastFM dataset, GNN models gener-
ally underperformed regarding both gender and age.
Analysis of subgroups reveals that specific subgroups
suffer from unfairness, with NGCF and DGCF show-
ing higher performance in the MovieLens dataset,
while most models exhibit greater unfairness toward
(a) MovieLens
(b) LastFM
(c) Book Recommendation
Figure 6: Results of NDCG performance for sensitive at-
tribute age on MovieLens, LastFM, and Book Recommen-
dation.
certain subgroups in LastFM.
Overall, while GNN models demonstrate strong
performance, user group distribution significantly im-
pacts fairness. This necessitates further investiga-
tion into data distribution and the application of pre-
processing methods to mitigate biases. Future work
WEBIST 2024 - 20th International Conference on Web Information Systems and Technologies
438

(a) MovieLens
(b) LastFM
Figure 7: Results of NDCG performance for subgroups on
Movielens and LastFM.
should focus on selecting models that align with data
behavior and task requirements, incorporating tech-
niques like counterfactual fairness, adversarial train-
ing, and personalized approaches. Additionally, ex-
ploring hybrid solutions that integrate graph struc-
tures with debiasing mechanisms could enhance both
performance and fairness in various recommendation
tasks.
REFERENCES
Airen, S. and Agrawal, J. (2022). Movie recommender
system using k-nearest neighbors variants. National
Academy Science Letters, 45(1):75–82.
Al-Ghamdi, M., Elazhary, H., and Mojahed, A. (2021).
Evaluation of collaborative filtering for recommender
systems. International Journal of Advanced Com-
puter Science and Applications, 12(3).
Boratto, L., Fabbri, F., Fenu, G., Marras, M., and Medda,
G. (2024). Robustness in fairness against edge-level
perturbations in gnn-based recommendation. In Euro-
pean Conference on Information Retrieval, pages 38–
55. Springer.
Bourhim, S., Benhiba, L., and Idrissi, M. J. (2022). A
community-driven deep collaborative approach for
recommender systems. IEEE Access.
Broman, N. (2021). Comparasion of recommender systems
for stock inspiration.
Celma, O. (2010). Music Recommendation and Discovery
in the Long Tail. Springer.
Chen, J., Dong, H., Wang, X., Feng, F., Wang, M., and
He, X. (2020). Bias and debias in recommender sys-
tem: A survey and future directions. arXiv preprint
arXiv:2010.03240.
Chizari, N., Shoeibi, N., and Moreno-Garc
´
ıa, M. N. (2022).
A comparative analysis of bias amplification in graph
neural network approaches for recommender systems.
Electronics, 11(20):3301.
Chizari, N., Tajfar, K., and Moreno-Garc
´
ıa, M. N. (2023a).
Bias assessment approaches for addressing user-
centered fairness in gnn-based recommender systems.
Information, 14(2):131.
Chizari, N., Tajfar, K., Shoeibi, N., and Moreno-Garc
´
ıa,
M. N. (2023b). Quantifying fairness disparities in
graph-based neural network recommender systems for
protected groups. In WEBIST, pages 176–187.
Dai, E. and Wang, S. (2021). Say no to the discrimina-
tion: Learning fair graph neural networks with limited
sensitive attribute information. In Proceedings of the
14th ACM International Conference on Web Search
and Data Mining, pages 680–688.
Di Noia, T., Tintarev, N., Fatourou, P., and Schedl, M.
(2022). Recommender systems under european ai reg-
ulations. Communications of the ACM, 65(4):69–73.
Ding, S., Feng, F., He, X., Liao, Y., Shi, J., and Zhang,
Y. (2022). Causal incremental graph convolution for
recommender system retraining. IEEE Transactions
on Neural Networks and Learning Systems.
Dong, Y., Kose, O. D., Shen, Y., and Li, J. (2023). Fairness
in graph machine learning: Recent advances and fu-
ture prospectives. In Proceedings of the 29th ACM
SIGKDD Conference on Knowledge Discovery and
Data Mining, pages 5794–5795.
Du, M., Yang, F., Zou, N., and Hu, X. (2020). Fairness
in deep learning: A computational perspective. IEEE
Intelligent Systems, 36(4):25–34.
Fahse, T., Huber, V., and Giffen, B. v. (2021). Managing
bias in machine learning projects. In International
Conference on Wirtschaftsinformatik, pages 94–109.
Springer.
Floridi, L., Holweg, M., Taddeo, M., Amaya Silva, J.,
M
¨
okander, J., and Wen, Y. (2022). capai-a procedure
for conducting conformity assessment of ai systems in
line with the eu artificial intelligence act. Available at
SSRN 4064091.
Girsang, A. S., Wibowo, A., et al. (2021). Neural collabora-
tive for music recommendation system. In IOP Con-
ference Series: Materials Science and Engineering,
volume 1071, page 012021. IOP Publishing.
G
´
omez, E., Boratto, L., and Salam
´
o, M. (2022). Provider
fairness across continents in collaborative recom-
mender systems. Information Processing & Manage-
ment, 59(1):102719.
Hardt, M., Price, E., and Srebro, N. (2016). Equality of op-
portunity in supervised learning. Advances in neural
information processing systems, 29.
Harper, F. M. and Konstan, J. A. (2019). Movie-
Assessing Unfairness in GNN-Based Recommender Systems: A Focus on Metrics for Demographic Sub-Groups
439

Lens 25m dataset. https://grouplens.org/datasets/
movielens/25m/.
Khan, Z. Y., Niu, Z., Sandiwarno, S., and Prince, R. (2021).
Deep learning techniques for rating prediction: a sur-
vey of the state-of-the-art. Artificial Intelligence Re-
view, 54(1):95–135.
Kuang, H., Xia, W., Ma, X., and Liu, X. (2021). Deep
matrix factorization for cross-domain recommenda-
tion. In 2021 IEEE 5th Advanced Information Tech-
nology, Electronic and Automation Control Confer-
ence (IAEAC), volume 5, pages 2171–2175. IEEE.
Liang, G., Sun, C., Zhou, J., Luo, F., Wen, J., and Li,
X. (2022). A general matrix factorization framework
for recommender systems in multi-access edge com-
puting network. Mobile Networks and Applications,
pages 1–13.
Mobius, A. (2020). Book recommendation dataset.
Mu, R. (2018). A survey of recommender systems based on
deep learning. Ieee Access, 6:69009–69022.
Navarin, N., Van Tran, D., and Sperduti, A. (2020). Learn-
ing kernel-based embeddings in graph neural net-
works. In ECAI 2020, pages 1387–1394. IOS Press.
Neophytou, N., Mitra, B., and Stinson, C. (2022). Re-
visiting popularity and demographic biases in recom-
mender evaluation and effectiveness. In European
Conference on Information Retrieval, pages 641–654.
Springer.
Purificato, E., Boratto, L., and De Luca, E. W. (2022). Do
graph neural networks build fair user models? assess-
ing disparate impact and mistreatment in behavioural
user profiling. In Proceedings of the 31st ACM In-
ternational Conference on Information & Knowledge
Management, pages 4399–4403.
Rahman, T., Surma, B., Backes, M., and Zhang, Y. (2019).
Fairwalk: Towards fair graph embedding.
Rahmattalabi, A., Jabbari, S., Lakkaraju, H., Vayanos, P.,
Izenberg, M., Brown, R., Rice, E., and Tambe, M.
(2021). Fair influence maximization: A welfare opti-
mization approach. In Proceedings of the AAAI Con-
ference on Artificial Intelligence, volume 35, pages
11630–11638.
Sang, L., Xu, M., Qian, S., and Wu, X. (2021). Knowl-
edge graph enhanced neural collaborative filtering
with residual recurrent network. Neurocomputing,
454:417–429.
Sha, X., Sun, Z., and Zhang, J. (2021). Disentangling
multi-facet social relations for recommendation. IEEE
Transactions on Computational Social Systems.
Spinelli, I., Scardapane, S., Hussain, A., and Uncini, A.
(2021). Fairdrop: Biased edge dropout for enhanc-
ing fairness in graph representation learning. IEEE
Transactions on Artificial Intelligence, 3(3):344–354.
Steck, H., Baltrunas, L., Elahi, E., Liang, D., Raimond,
Y., and Basilico, J. (2021). Deep learning for recom-
mender systems: A netflix case study. AI Magazine,
42(3):7–18.
Sun, W., Chang, K., Zhang, L., and Meng, K. (2021). In-
gcf: An improved recommendation algorithm based
on ngcf. In International Conference on Algorithms
and Architectures for Parallel Processing, pages 116–
129. Springer.
Tang, H., Zhao, G., Wu, Y., and Qian, X. (2021).
Multisample-based contrastive loss for top-k recom-
mendation. IEEE Transactions on Multimedia.
Wang, S., Hu, L., Wang, Y., He, X., Sheng, Q. Z., Orgun,
M. A., Cao, L., Ricci, F., and Yu, P. S. (2021).
Graph learning based recommender systems: A re-
view. arXiv preprint arXiv:2105.06339.
Wang, Y., Ma, W., Zhang*, M., Liu, Y., and Ma, S. (2022).
A survey on the fairness of recommender systems.
ACM Journal of the ACM (JACM).
Wang, Y., Ma, W., Zhang, M., Liu, Y., and Ma, S. (2023). A
survey on the fairness of recommender systems. ACM
Transactions on Information Systems, 41(3):1–43.
Wu, L., Chen, L., Shao, P., Hong, R., Wang, X., and Wang,
M. (2021). Learning fair representations for recom-
mendation: A graph-based perspective. In Proceed-
ings of the Web Conference 2021, pages 2198–2208.
Wu, S., Sun, F., Zhang, W., Xie, X., and Cui, B. (2020).
Graph neural networks in recommender systems: a
survey. ACM Computing Surveys (CSUR).
Xue, H.-J., Dai, X., Zhang, J., Huang, S., and Chen, J.
(2017). Deep matrix factorization models for recom-
mender systems. In IJCAI, volume 17, pages 3203–
3209. Melbourne, Australia.
Yang, C. (2022). Supervised contrastive learning for rec-
ommendation. arXiv preprint arXiv:2201.03144.
Yi, B., Shen, X., Liu, H., Zhang, Z., Zhang, W., Liu, S., and
Xiong, N. (2019). Deep matrix factorization with im-
plicit feedback embedding for recommendation sys-
tem. IEEE Transactions on Industrial Informatics,
15(8):4591–4601.
Zhang, Q., Wipf, D., Gan, Q., and Song, L. (2021). A bi-
ased graph neural network sampler with near-optimal
regret. Advances in Neural Information Processing
Systems, 34:8833–8844.
Zhang, Z., Liu, Y., Xu, G., and Luo, G. (2016). Recommen-
dation using dmf-based fine tuning method. Journal
of Intelligent Information Systems, 47(2):233–246.
Zheng, Y. and Wang, D. X. (2022). A survey of rec-
ommender systems with multi-objective optimization.
Neurocomputing, 474:141–153.
Zhou, J., Cui, G., Hu, S., Zhang, Z., Yang, C., Liu, Z.,
Wang, L., Li, C., and Sun, M. (2020). Graph neu-
ral networks: A review of methods and applications.
AI Open, 1:57–81.
WEBIST 2024 - 20th International Conference on Web Information Systems and Technologies
440