Slides for the invited talk "A Topology-aware Analysis of Graph Collaborative Filtering" at the Glasgow Information Retrieval Group (University of Glasgow).
Event link: https://samoa.dcs.gla.ac.uk/events/viewtalk.jsp?id=19320
Paper: https://arxiv.org/pdf/2308.10778.pdf
VIP High Class Call Girls Jamshedpur Anushka 8250192130 Independent Escort Se...
Β
[IRTalks@The University of Glasgow] A Topology-aware Analysis of Graph Collaborative Filtering
1. A Topology-aware
Analysis of Graph
Collaborative Filtering
Daniele Malitesta, PhD
SisInfLab Research Group
Politecnicodi Bari (Bari, Italy)
The Universityof Glasgow (Glasgow, UK)
50 minutes
February 05, 2024 (15.00 UK Time)
IR Talks
3. USEFUL RESOURCES
The content of the following slides is taken from:
β Daniele Malitesta, Claudio Pomo, Vito Walter Anelli, Alberto Carlo
Maria Mancino, Eugenio Di Sciascio, Tommaso Di Noia:
A Topology-aware Analysis of Graph Collaborative Filtering. CoRR
abs/2308.10778 (2023)
β Daniele Malitesta, Claudio Pomo, Tommaso Di Noia: Graph Neural
Networks for Recommendation: Reproducibility, Graph Topology,
and Node Representation. CoRR abs/2310.11270 (2023)
Scan me!
7. A NON-EXHAUSTIVE TIMELINE
Pioneer approaches
proposing GCN-based
aggregation methods
[v an den Berg et al.,
Ying et al.]
Explore inter-dependencies
between nodes and their
neighbours [Wang et al.
(2019a)], use graph attention
networks to recognize
meaningful user-item
interactions at higher-grained
level [Wang et al. (2019b)]
Lighten the graph
convolutional layer [Chen
e t al., He et al.], use graph
attention networks to
recognize meaningful user-
item interactions at higher-
grained level [Wang et al.
(2020b), Tao et al.]
Self-supervised and
contrastive learning
[Wu et al., Yu et al.]
Simplify the message-
passing formulation
[Mao et al., Peng et al.,
Shen et al.]
Explore other latent
spaces [Shen et al., Sun
e t al., Zhang et al.]
Exploit hypergraphs
[We i et al., Xia et al.]
Use graph attention networks
to recognize meaningful user-
item interactions at finer-
grained [Zhang et al.] level
2017-2018
2019
2020
2021-2022
2021-2022
2021-2022
2022
2022
13. Topologically
WHY GRAPH CF WORKS WELL?
β’ The machine learning literature
acknowledges that graph topology
plays a crucial role in GNNs [Wei et
al., Castellana and Errica].
β’ For instance, revisit network
topology to enhance the learning
ability of GCNs [Shi et al.].
β’ Established approaches such as
LightGCN and DGCF adapt the GCN
layer to suit the CF rationale.
β’ Recent graph-based RSs including
UltraGCN and SVD-GCN go beyond
the concept of message-passing.
Algorithmically
14. [Adomavicius andZhang, Deldjoo et al.]
Classical dataset characteristics
such as dataset sparsity may
impact the performance of
recommendation models in
diverse scenarios and tasks.
15. The topological nature of the user-
item data and the new graph CF wave
make it imperative to unveil the
dependencies between topological
properties and recommendation
performance.
16. 3) Build an explanatory model
OUR CONTRIBUTIONS
that studies the impact of classical and
topological dataset characteristics on the
performance of SOTA graph CF.
which are recent and across-the-board in
the literature.
to calculate the linear dependences among
characteristics and recommendation
metrics.
on their statistical significance, with
varying settings of graph samplings.
2) Select 4 graph CF approaches
4) Validate the explanations
1) First analysis in the literature
18. NODE DEGREE EXTENSION (1/3)
DEFINITION
Let ππ’
(π)
and ππ
(π)
be the sets of neighborhood nodes for user π’ and item π at π distance hops. The extended
definition of node degrees for π’ and π is:
ππ’ = ππ’
(1)
, ππ = ππ
(1)
.
The average user and item node degrees on the whole user and item sets (π and πΌ):
ππ =
1
π
Οπ’βπ ππ’
1
, ππΌ =
1
πΌ
ΟπβπΌ ππ
(1)
.
19. NODE DEGREE EXTENSION (2/3)
RECSYS RE-INTERPRETATION
The node degree in the user-item graph stands for the number of items (users) interacted by a user
(item). This is related to the cold-start issue in recommendation, where cold users denote low activity on
the platform, while cold items are niche products.
π’1
π’2
π’3
π’4
π’5
π’6
π’7
π1
π2
π3
π4
π5
π6
π7
π8
ππ’2
= 5 (active user)
ππ’7
= 1 (non-active user)
20. NODE DEGREE EXTENSION (3/3)
RECSYS RE-INTERPRETATION
The node degree in the user-item graph stands for the number of items (users) interacted by a user
(item). This is related to the cold-start issue in recommendation, where cold users denote low activity on
the platform, while cold items are niche products.
π’1
π’2
π’3
π’4
π’5
π’6
π’7
π1
π2
π3
π4
π5
π6
π7
π8
ππ7
= 4 (popular item)
ππ4
= 1 (niche item)
26. DEGREE ASSORTATIVITY (1/3)
DEFINITION
Let π· = {π1,π2, β¦} be the set of unique node degrees in the graph, and let ππβ,ππ
be the fraction of edges
connecting nodes with degrees πβ and ππ. Then, let ππβ
be the probability distribution to choose a node
with degree πβ after having selected a node with the same degree (i.e., the excess degree distribution).
The degree assortativity coefficient [M. E. J. Newman] is:
π =
Οπβππ
πβππ ππβ,ππ
β ππβ
πππ
π π‘ππ
2
,
where π π‘ππ is the standard deviation of the distribution π.
27. DEGREE ASSORTATIVITY (2/3)
RECSYS RE-INTERPRETATION
The degree assortativity calculated user- and item-wise is a proxy to represent the tendency of users
with the same activity level on the platform and items with the same popularity to gather, respectively.
To visualize the degree assortativity, we need the projected user-user and item-item graphs.
π’1
π’2
π’3
π’4
π’5
π’6
π’7
π1
π2
π3
π4
π5
π6
π7
π8
π’1
π’2
π’3
π’4
π’5
π’6
π’7
user-item graph user-user projected graph
28. DEGREE ASSORTATIVITY (3/3)
RECSYS RE-INTERPRETATION
The degree assortativity calculated user- and item-wise is a proxy to represent the tendency of users
with the same activity level on the platform and items with the same popularity to gather, respectively.
To visualize the degree assortativity, we need the projected user-user and item-item graphs.
π’1
π’2
π’3
π’4
π’5
π’6
π’7
π1
π2
π3
π4
π5
π6
π7
π8
π’1
π’2
π’3
π’4
π’5
π’6
π’7
π = β0.191
(high degree dis-assortativity)
29. Density
CLASSICAL DATASET CHARACTERISTICS
Estimates the number of all possible
interactions:
π = ππΌ.
Defines the ratio between the number
of users and items:
π =
π
πΌ
.
Measures the ratio of actual user-item
interactions to all possible interactions:
πΏ =
πΈ
ππΌ
.
Calculates the interactionsβ
concentration for users (items):
π π =
Οπ’=1
πβ1 Οπ£=π’+1
π πππ ππ’βππ£
π Οπ’=1
π ππ’
.
Shape
Gini coefficient
Space size
31. UltraGCN (CIKM 2021)
SELECTED GRAPH CF APPROACHES
Node degree used to normalize the
adjacency matrix in the message-passing.
Node degree used to normalize the
adjacency matrix in the message-
passing.
Node degree used for normalization in
the infinite layer message-passing. The
model also learns from the item-
projected graph.
Node embeddings involve the largest
singular values of the normalized
user-item interaction matrix, whose
maximum value is related to the
maximum node degree of the user-
item graph. The model learns from
user- and item-projected graphs.
DGCF (SIGIR 2020)
SVD-GCN (CIKM 2022)
LightGCN (SIGIR 2020)
32. The selected graph CF approaches
explicitly utilize the node degree in
the representation learning. However,
clustering coefficient and degree
assortativity do not have an evident
representation in the formulations.
33. β’ Which topological aspects graph-based
models can (un)intentionally capture?
β’ Are (topological) dataset characteristics
influencing the recommendation
performance of graph CF models?
35. Our goal is to understand whether there
exist dependencies among (topological)
dataset characteristics and the performance
of graph-based recommendation models.
To this aim, we decide to build and fit an
explanatory framework.
62. GENERATE (LINEAR) EXPLANATIONS
Characteristics
Performance
We aim to fit the following multivariate linear
regression model:
π = π + π0 + πππΏπ.
Predicted
performance
Prediction
error
Global
bias
Weight of
characteristic c
Characteristic c
for all samples
63. GENERATE (LINEAR) EXPLANATIONS
Characteristics
Performance
We aim to fit the following multivariate linear
regression model:
π = π + π0 + πππΏπ.
We use the ordinary least squares (OLS)
optimization model:
π0
β,π½π
β = min
π0,ππ
1
2
π β π0 β π½ππΏπ 2
2
.
67. IMPACT OF CHARACTERISTICS
How good is the
linear regression
at making
predictions?
π 2
is, for many settings, above 95%.
The regression model can provide good explanations.
71. IMPACT OF CHARACTERISTICS
direct
This confirms what observed in the modelsβ formulations.
High usersβ interactions correspond to better performance.
76. In the worst-case scenario, node-
dropout drops many high-degree
nodes; edge-dropout, drops all the
edges connected to several nodes and
thus disconnect them from the graph.
86. WHAT WE HAVE LEARNED
01
LATENT-FACTORS vs.
NEIGHBORHOOD
87. WHAT WE HAVE LEARNED
01 02
LATENT-FACTORS vs.
NEIGHBORHOOD
NODE DEGREE
88. WHAT WE HAVE LEARNED
01 02 03
LATENT-FACTORS vs.
NEIGHBORHOOD
NODE DEGREE CLUSTERING
COEFFICIENT
89. WHAT WE HAVE LEARNED
01
04
02 03
LATENT-FACTORS vs.
NEIGHBORHOOD
NODE DEGREE CLUSTERING
COEFFICIENT
DEGREE ASSORTATIVITY
90. WHAT WE HAVE LEARNED
01
04
02
05
03
LATENT-FACTORS vs.
NEIGHBORHOOD
NODE DEGREE CLUSTERING
COEFFICIENT
DEGREE ASSORTATIVITY NODE- vs. EDGE-
DROPOUT
91. REFERENCES (1/3)
[Malitesta et al.] A Topology-aware Analysis of Graph Collaborative Filtering. CoRR abs/2308.10778
(2023)
[Malitesta et al.] Graph Neural Networks for Recommendation: Reproducibility, Graph Topology, and
Node Representation. CoRR abs/2310.11270 (2023)
[Wang et al. (2020a)] Learning and Reasoning on Graph for Recommendation. WSDM 2020: 890-893
[El-Kishky et al.] Graph-based Representation Learning for Web-scale Recommender Systems. KDD
2022: 4784-4785
[Gao et al.] Graph Neural Networks for Recommender System. WSDM 2022: 1623-1625
[Purificato et al. (2023a)] Tutorial on User Profiling with Graph Neural Networks and Related Beyond-
Accuracy Perspectives. UMAP 2023: 309-312
[Purificato et al. (2023b)] Leveraging Graph Neural Networks for User Profiling: Recent Advances and
Open Challenges. CIKM 2023: 5216-5219
[van den Berg et al.] Graph Convolutional Matrix Completion. CoRR abs/1706.02263 (2017)
[Ying et al.] Graph Convolutional Neural Networks for Web-Scale Recommender Systems. KDD 2018:
974-983
[Wang et al. (2019a)] Neural Graph Collaborative Filtering. SIGIR 2019: 165-174
[Wang et al. (2019b)] KGAT: Knowledge Graph Attention Network for Recommendation. KDD 2019:
950-958
92. REFERENCES (2/3)
[Chen et al.] Revisiting Graph Based Collaborative Filtering: A Linear Residual Graph Convolutional
Network Approach. AAAI 2020: 27-34
[He et al.] LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation
LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. SIGIR 2020:
639-648
[Wang et al. (2020b)] Disentangled Graph Collaborative Filtering. SIGIR 2020: 1001-1010
[Tao et al.] MGAT: Multimodal Graph Attention Network for Recommendation. Inf. Process. Manag. 57(5):
102277 (2020)
[Wu et al.] Self-supervised Graph Learning for Recommendation. SIGIR 2021: 726-735
[Yu et al.] Are Graph Augmentations Necessary?: Simple Graph Contrastive Learning for
Recommendation. SIGIR 2022: 1294-1303
[Mao et al.] UltraGCN: Ultra Simplification of Graph Convolutional Networks for Recommendation.
CIKM 2021: 1253-1262
[Peng et al.] SVD-GCN: A Simplified Graph Convolution Paradigm for Recommendation. CIKM 2022:
1625-1634
[Shen et al.] How Powerful is Graph Convolution for Recommendation? CIKM 2021: 1619-1629
[Sun et al.] HGCF: Hyperbolic Graph Convolution Networks for Collaborative Filtering. WWW 2021:
593-601
93. REFERENCES (3/3)
[Zhang et al.] Geometric Disentangled Collaborative Filtering. SIGIR 2022: 80-90
[Wei et al.] Dynamic Hypergraph Learning for Collaborative Filtering. CIKM 2022: 2108-2117
[Xia et al.] Hypergraph Contrastive Collaborative Filtering. SIGIR 2022: 70-79
[Wei et al.] Designing the Topology of Graph Neural Networks: A Novel Feature Fusion Perspective.
WWW 2022: 1381-1391
[Castellana and Errica] Investigating the Interplay between Features and Structures in Graph Learning.
CoRR abs/2308.09570 (2023)
[Shi et al.] Topology and Content Co-Alignment Graph Convolutional Learning. IEEE Trans. Neural
Networks Learn. Syst. 33(12): 7899-7907 (2022)
[Adomavicius and Zhang] Impactof data characteristics on recommender systems performance. ACM
Trans. Manag. Inf. Syst. 3(1): 3:1-3:17 (2012)
[Deldjoo et al.] How Dataset Characteristics Affect the Robustness of Collaborative Recommendation
Models. SIGIR 2020: 951-960
[Latapy et al.] Basic notions for the analysis of large two-mode networks. Soc. Networks 30(1): 31-48
(2008)
[M. E. J. Newman] Mixing patterns in networks. Phys. Rev. E, 67:026126, Feb 2003. doi: 10.
1103/PhysRevE.67.026126
94. CREDITS: This presentation template was created by Slidesgo, and
includes icons by Flaticon,and infographics & images by Freepik
Q&A Time
Scan me!