Knowledge Graph Convolutional Networks for Recommender Systems

1. Ho-Beom Kim
Network Science Lab
Dept. of Mathematics
The Catholic University of Korea
E-mail: hobeom2001@catholic.ac.kr
2023 / 12 / 18
WANG, Hongwei, et al.
The world wide web conference. 2019.

2. 2
Introduction
Problem Statements
• The traditional method often used for recommendation systems is collaborative filtering (CF).
• CF is implemented using an adjacency matrix that represents interactions between users and items (or
entities) in matrix form.
• Although this method is quite old, it's still frequently used because of its effectiveness.
• However, it has several problems, as described below:
1. An adjacency matrix A^u×e that represents the interrelations between user u and item e as either 0 or 1 is
required. Therefore, it consumes a significant amount of memory.
2. The created matrix A is a sparse matrix, and the cold start problem occurs when an arbitrary user has
minimal interactions with items, making predictions challenging.

3. 3
Introduction
Contribution
1. The authors propose a recommendation system that uses a Knowledge Graph (KG) to learn the
interactions between users and items, enabling the recommendation of new items. Their approach
not only considers the relationships between users and items but also explores the entities that the items
are part of, inferring meanings and utilizing them for learning.
2. It employs a graph convolutional network (GCN) utilizing a receptive field for end-to-end learning to
explore the KG and identify high-order relations of interest to users
3. It demonstrates superior performance when applied to real datasets

4. 4
Methodology
A two-layer receptive field (green entities) of the blue entity in a KG. / The framework of KGCN.
USER
Entities
The number of layer
When h=1, the recommendation system is made based solely on the meanings of the item and user.
Therefore, in the case of ℎ>1, it explores the Knowledge Graph.
If K≠1, there is a need for a method to converge opinions from each entity and convey them to the
higher-level entity or user.

5. 5
Methodology
A two-layer receptive field (green entities) of the blue entity in a KG. / The framework of KGCN.
 Do binary classification between User u, item v
 Calculate the score between user u and relation r
 Calculate the score between user u and item v
 Normalize pi to make the summation of score between user u and relation
r to 1
 Using the normalized score, we can calculate the score between u, v

6. 6
Methodology
A two-layer receptive field (green entities) of the blue entity in a KG. / The framework of KGCN.
From N(v), a new set S(v) containing k elements is sampled.
S(v) acts like a receptive field, performing convolutional operations.

7. 7
Methodology
KGCN algorithm
Using the get-receptive field, generate sets of neighboring
entities according to depth h, and then calculate scores using
each entity and relation.

8. 8
Q&A
Q&A

9. 9
Experiments
Datasets
• MovieLens-204M
• Book-Crossing
• Last.FM
Baselines
• SVD
• LibFM
• LibFM + TransE
• PER
• CKE
• RippleNet

10. 10
Experiments
The results of AUC and F1 in CTR prediction.
KGCN outperforms all baselines by a significant margin, while their performances are slightly distinct
KGCN-avg performs worse than KGCN-sum, especially in Book-Crossing and Last.FM where interactions are sp

11. 11
Experiments
The results of Recall@K in top-K recommendation

12. 12
Experiments
AUC result of KGCN with different neighbor sampling size K.
KGCN achieves the best performance when K = 4 or 8.
This is because a too small K does not have enough capacity to incorporate neighborhood information,
while a too large K is prone to be misled by noises.

13. 13
Experiments
AUC result of KGCN with different depth of receptive field H.
 The results are shown in this table, which demonstrate that KGCN is more sensitive to H compared to K

14. 14
Experiments
AUC result of KGCN with different dimension of embedding.
Increasing d initially can boost the performance since a larger d can encode more information of users and entities,
while a too large d adversely suffers from overfitting.

15. 15
Conclusion
Conclusion
• This paper proposes knowledge graph convolutional networks for recommender
systems.
• In this work They uniformly sample from the neighbors of an entity to construct its
receptive field. Exploring a non-uniform sampler
• This paper (and all literature) focuses on modeling item-end KGs.
• Designing an algorithm to well combine KGs at the two ends is also a promising
direction.

16. 16
Q&A
Q&A