PaperlayerView exact revision on arXiv

Source-linked AI summary

Contextualized Graph Attention Network for Recommendation with Item Knowledge Graph

Susen Yang, Yong Liu, Yonghui Xu, Chunyan Miao, Min Wu, Juyong Zhang

arXiv:2004.11529v1cs.IR

TL;DR

GNN-based recommendation may not capture non-local knowledge-graph context explicitly. CGAT combines user-specific attention for local context, random walks and an RNN for non-local context, and item-specific attention to historical items. It is reported to usually achieve the best performance across datasets and metrics, with superiority over state-of-the-art baselines validated on three datasets.

  • Problem

    GNN-based recommendation may not explicitly capture non-local KG context, the most related high-order neighbors of an entity.

    Source: §1 · p. 2

  • Method

    CGAT uses user-specific graph attention for local context, biased random walks and an RNN for non-local context, and item-specific attention to historical items.

    Source: §1 · p. 2

  • Results

    CGAT usually achieves the best performance across datasets and metrics, and its superiority over state-of-the-art baselines is reported on three datasets.

    Source: §4.2 · p. 5, §5 · p. 7

  • Takeaways & Limitations

    The reported results support CGAT’s effectiveness against state-of-the-art KG-based recommendation methods on three datasets.

    Source: §1 · p. 2, §5 · p. 7

    The framework assumes an item knowledge graph is available.

    Source: §3 · p. 2

Abstract

from arXiv · show

Graph neural networks (GNN) have recently been applied to exploit knowledge graph (KG) for recommendation. Existing GNN-based methods explicitly model the dependency between an entity and its local graph context in KG (i.e., the set of its first-order neighbors), but may not be effective in capturing its non-local graph context (i.e., the set of most related high-order neighbors). In this paper, we propose a novel recommendation framework, named Contextualized Graph Attention Network (CGAT), which can explicitly exploit both local and non-local graph context information of an entity in KG. Specifically, CGAT captures the local context information by a user-specific graph attention mechanism, considering a user's personalized preferences on entities. Moreover, CGAT employs a biased random walk sampling process to extract the non-local context of an entity, and utilizes a Recurrent Neural Network (RNN) to model the dependency between the entity and its non-local contextual entities. To capture the user's personalized preferences on items, an item-specific attention mechanism is also developed to model the dependency between a target item and the contextual items extracted from the user's historical behaviors. Experimental results on real datasets demonstrate the effectiveness of CGAT, compared with state-of-the-art KG-based recommendation methods.

1 Introduction

The introduction identifies two gaps in GNN-based recommendation with knowledge graphs: insufficient personalization of local context and weak explicit modeling of non-local context. CGAT addresses these gaps with user-specific local attention and explicit non-local context modeling.

Motivation

  • KGs offer semantic relations between items, but recommendation methods still face the challenge of exploiting those relations and graph structure effectively.The passage describes a KG as a heterogeneous network of entities and relations.Source: §1 · p. 1

Research gaps

  • Many GNN-based methods do not model users’ different preferences for the same entity when aggregating its first-order neighbors.The example contrasts users who prefer the same item for different KG attributes.Source: §1 · p. 1
  • Existing GNN methods may miss important high-order neighbors, while layer-by-layer propagation can weaken distant entities’ effects or introduce noise.The example describes an item with one directly linked entity but a relevant entity connected through multiple paths.Source: §1 · p. 2

Proposed approach

  • CGAT combines user-specific attention for local KG context, biased random walks and an RNN for non-local context, and item-specific attention to users’ historical items.The framework explicitly exploits local and non-local entity context as well as context from users’ historical data.Source: §1 · p. 2

2 Related Work

The related work groups KG-based recommendation into regularization-, path-, and GNN-based methods. It describes differences in how these approaches use KG structure and context.

KG-based recommendation methods

  • Regularization-based methods incorporate KG structure into the embedding-loss objective, but do not explicitly model KG semantic relations [Cao et al., 2019].Examples include CKE, KTUP, and cross&compress units [Wang et al., 2019b].Source: §2 · p. 2
  • Path-based methods use entity connection patterns, including meta-path similarities and paths selected by rules or item associations [Yu et al., 2013; Shi et al., 2015; Zhao et al., 2017; Wang et al., 2018; Sun et al., 2018; Ma et al., 2019].Some approaches use path-length conditions or item associations to select multi-hop relational paths.Source: §2 · p. 2
  • GNN-based methods support end-to-end KG recommendation, but aggregate high-order neighborhoods implicitly through layer-by-layer propagation [Wang et al., 2019a; Wang et al., 2019c].KGNN-LS learns user-specific relation weights, while another method applies graph attention to local neighborhoods.Source: §2 · p. 2

3 Contextualized Graph Attention Network

CGAT combines local and non-local knowledge-graph context with context from a user’s interaction history to predict preference scores for candidate items.

Local Graph Context

  • User-specific attention aggregates one-hop KG neighbors, incorporating relation embeddings to represent how each neighbor matters to a target user.The resulting local neighborhood embedding is combined with the target entity’s embedding.Source: Local Graph Context · pp. 2–3

Non-Local Graph Context

  • Biased random walks extract non-local context by generating fixed-length paths and ranking encountered entities by their frequency across walks.The method uses a GRU to model dependencies between the target entity and this context, processing the frequency-ranked entities in reverse order.Source: Non-Local Graph Context · pp. 3–4
  • A learned gate combines local and non-local KG context, producing an item context embedding that is concatenated with the item embedding.The gate learns weights for integrating the two contextual embeddings.Source: Non-Local Graph Context · p. 4

3.2 Exploiting Interaction Graph Context

  • Item-specific attention weights a user’s historical items according to their relevance to the target item, forming interaction context for that prediction.The method assumes a historical item can have different importance for different candidate items.Source: §3.2 · p. 4

3.3 Learning Algorithm

  • BPR ranks interacted items above unobserved items, while KG regularization contrasts observed triples with sampled non-neighbors during joint parameter optimization.The model uses gradient descent to solve the combined objective.Source: §3.3 · p. 4
  • Each mini-batch iteration has complexity O(BNSd^2) for KG context and O(BNd^2) for interaction context; random-walk sampling is performed before training.The sampling procedure has complexity O(|I|SML).Source: §3.3 · p. 5

4 Experiments

Experiments compare CGAT with KG-based baselines on three datasets and examine the contributions of its context components and parameter choices.

4.1 Experimental Settings

  • Experiments use Last-FM, Movielens-1M, and Book-Crossing, with 60% of interactions for training, 20% for tuning, and 20% for testing.Ratings are treated as implicit feedback on FM and BC; ML uses ratings above 4.Source: §4.1 · p. 5

4.2 Performance Comparison

  • CGAT usually achieves the best performance across datasets and metrics, significantly outperforming baselines in 23 of 27 comparisons (p < 0.05).Averaged over datasets, CGAT exceeds CFKG, RippleNet, MKR, KGNN-LS, and KGAT by 26.07%, 21.32%, 22.29%, 21.92%, and 9.56%, respectively, in HR@20.Source: §4.2 · p. 5, Table 2 · p. 6

4.3 Ablation Study

  • CGAT outperforms variants without local context, non-local context, or user-specific attention; non-local-only is slightly better than local-only on ML and BC.The ablation results are reported only for HR@20.Source: §4.3 · pp. 5–6, Table 3 · p. 6

4.4 Parameter Sensitivity Study

  • On FM, the best settings include S=4, N=16, λ1=5 × 10−5 or 10−4, and M=15; L between 4 and 12 performs better.Larger values can increase training time and sometimes reduce recommendation performance; increasing N beyond 16 also reduces performance.Source: §4.4 · p. 6

5 Conclusion and Future Work

CGAT combines local and non-local knowledge-graph context with users’ historical interaction context. The paper reports superiority over state-of-the-art baselines on three datasets and identifies context integration as future work.

5 Conclusion and Future Work

  • CGAT uses user-specific graph attention for local KG context, biased random walks and a GRU for non-local entities, and item-specific attention for item influences.The model also incorporates interaction context from users’ historical behaviors.Source: §5 · p. 6
  • CGAT’s superiority is validated against state-of-the-art baselines on three datasets.Source: §5 · p. 7
  • Future work will develop aggregation strategies to integrate KG and interaction-graph context to improve recommendation accuracy.Source: §5 · p. 7

Related papers

Loading 2004.11529v1…