Jooho-Lee
School of Computer Science and
Information Engineering, The Catholic
University of Korea
E-mail: jooho414@gmail.com
2023-03-16

1
 Introduction
• Problem Statement
• Contribution
 Method
 Experiment
 Conclusion

2
Introduction
Problem Statement
• User Intents
• Relational Paths
 because of GNN aggregator (they just do summation or something)

3
Introduction
Contribution
• Revealing user intents behind the interactions in KG-based recommendation, for better model capacity and
interpretability
• Proposing a new model, KGIN, which consider user-item relationships at the finer granularity of intents and
long-range semantics of relational paths under the GNN paradigm

4
Method
KGIN : Model Architecture

5
Method
User Intent Modeling
• Representation Learning of Intents
• Independence Modeling of Intents

6
Method
Relational Path-aware Aggregation
• Aggregation Layer over Intent Graph

7
Method
Relational Path-aware Aggregation
• Aggregation Layer over Knowledge Graph

8
Method
Capturing Relational Paths

9
Method
Model Prediction & Optimization

10
Experiment
Experimental Setting
• Amazon-Book
• Last-FM
• Alibaba-iFashion

11
Experiment
Overall performance comparison

12
Experiment
Impact of presence of user intents and KG relations

13
Experiment
Impact of the number of layers

14
Experiment
Explanation of Intent

15
Conclusion
• uncovering user-item relationships at the granularity of intents, which are coupled with KG relations to exhibit
the explainable semantics
• relational path-aware aggregation, which integrates relational information from multi-hop paths to refine the
representations