# Conditional Attention Networks for Distilling Knowledge Graphs in Recommendation

**Authors:** Ke Tu, Peng Cui, Daixin Wang, Zhiqiang Zhang, Jun Zhou, Yuan Qi, Wenwu Zhu  
**arXiv:** [2111.02100](https://arxiv.org/abs/2111.02100)  
**Date:** 2021-11-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2111.02100v1/

## Key points


## Summary

In this paper, the authors propose a model called Knowledge-aware Conditional Attention Networks (KCAN) to incorporate knowledge graphs into recommender systems. They address the issue of generalization and scale in knowledge graphs, where most knowledge relationships are not helpful for target user-item predictions. To overcome this, they propose distilling the knowledge graph to reserve useful information and refine the knowledge to capture users' preferences. The KCAN model uses a knowledge-aware attention propagation method to obtain node representations that capture global semantic similarity on both the user-item network and the knowledge graph. Then, for a given target user-item pair, they automatically distill the knowledge graph into a target-specific subgraph based on knowledge-aware attention. By applying conditional attention aggregation on this subgraph, they refine the knowledge graph to obtain target-specific node representations. The authors highlight two key challenges: knowledge graph distillation and refinement. Knowledge graph distillation involves transforming the full massive knowledge graph into a small concentrated one that accurately captures users' preferences. On the other hand, knowledge graph refinement aims to give different weights to all knowledge relationships instead of just considering their neighbors when making recommendations. They argue that previous embedding-based models fail to solve these challenges effectively. Instead, they propose path-based methods that explore different paths in the knowledge graph to capture more specific and relevant information for recommendation tasks. The rest of the paper is structured as follows: Section 2 provides a brief review of related works focusing specifically on Knowledge Graph Embedding (KGE). Section 3 formally defines the solved problems and introduces details of their proposed KCAN model. In Section 4, experimental results using real-world datasets are reported demonstrating effectiveness of their framework compared with state-of-the art algorithms. Finally in Section 5 conclusions are drawn from findings presented in this paper.