PaperlayerView exact revision on arXiv

Source-linked AI summary

Do All Nodes Benefit Equally from Knowledge Graphs? Adaptive Node-Aware KG Fusion for Recommendation

Jaehyun Park, Minseo Jeon, Daewon Gwak, Sunuk Kim, Hanvit Lee, Jinhong Jung

arXiv:2609.05909v1cs.IR

TL;DR

Existing KG-aware recommenders may inject item knowledge without accounting for whether each node’s collaborative signals are reliable. AdaKG estimates node-wise CF instability, adapts KG fusion, and achieves the best average rank across four datasets.

  • Problem

    Existing methods inject KG information without accounting for each node’s need, potentially distorting embeddings when CF signals are already reliable.

    Source: §1 · p. 2

  • Method

    AdaKG separately encodes IG and KG, estimates node-wise CF instability under small adversarial perturbations, and aligns and fuses their embeddings according to node-wise reliance.

    Source: §1 · p. 2, §4.4 · p. 6, §4.4.1 · p. 6

  • Results

    AdaKG achieves the best average rank across four datasets, ranking best on Amazon-Book, MovieLens-1M, and Last-FM, and best on MRR@10 and second-best on Recall@10 for Book-Crossing.

    Source: §5.2 · p. 8

  • Takeaways & Limitations

    AdaKG’s node-wise KG contributions preserve reliable collaborative signals as the primary basis while giving less-stable nodes more KG-enhanced information.

    Source: §6 · p. 10

Abstract

from arXiv · show

KG-aware recommendation has been widely studied to alleviate data sparsity by using knowledge graphs (KGs), which represent items, entities, and their relations as graphs and provide item-side knowledge. However, existing methods incorporate item knowledge without considering how much each user or item node should rely on it. As a result, they apply KG signals indiscriminately across nodes, even to nodes whose collaborative filtering (CF) signals from the interaction graph (IG) are already reliable. In this paper, we propose AdaKG (Adaptive Node-Aware KG Fusion), a novel KG-aware recommendation method that adaptively adjusts the contribution of auxiliary knowledge for each node. Since user-item interactions and item knowledge provide different types of signals, directly mixing them can distort the CF signals. To avoid this, AdaKG separately encodes the IG and KG with view-specific encoders, allowing each view to capture its own information. It then estimates how strongly each node should rely on item knowledge by measuring the stability of its CF signals under small adversarial perturbations, assigning a larger KG contribution to less stable nodes. Finally, AdaKG adaptively aligns the IG and KG embeddings in a shared space and fuses them according to the estimated node-wise reliance. Through experiments, we show that AdaKG achieves strong performance compared with its baselines and the effectiveness of our adaptive fusion strategy.

1 Introduction

Knowledge graphs can provide item-side information to mitigate sparse-interaction and cold-start challenges, but indiscriminate KG fusion may distort reliable collaborative signals. AdaKG estimates node-wise CF instability and adapts KG contributions accordingly.

1 Introduction

  • Knowledge graphs connect sparsely interacted or new items to existing items through item-side knowledge, addressing a limitation of interaction-based collaborative filtering.CF relies on user-item interactions and can produce inaccurate preferences when users or items have few connections.Source: §1 · p. 1
  • Prior KG-aware methods also select potentially useful entities and relations, but AdaKG controls how much KG information each node receives.Existing techniques include attention-based message passing, graph augmentation, and graph reconstruction.Source: §1 · pp. 1–2
  • Existing methods may mix interaction-graph and KG signals despite their different roles, while interaction degree alone cannot identify every node with unstable CF signals.Even nodes with many interactions may have unstable signals when their interactions are not coherent; indiscriminate KG injection may distort reliable CF embeddings.Source: §1 · p. 2
  • AdaKG separately encodes the interaction graph and KG, estimates CF instability under small adversarial perturbations, then aligns and fuses embeddings using node-wise reliance.Greater instability indicates greater reliance on KG information.Source: §1 · p. 2
  • Experiments on four real-world datasets show strong performance against state-of-the-art KG-aware methods, and further analyses validate the adaptive fusion strategy.The reported contributions emphasize controlling KG information according to each node’s CF instability.Source: §1 · p. 2

2 Related Work

Prior KG-aware recommendation work spans translation-based, path-based, GNN-based, and self-supervised methods. The review distinguishes these approaches partly by how they incorporate KG information and whether they control its node-wise contribution.

2 Related Work

  • Translation-based CKE and CFKG [53] use translation objectives, while path-based MCRec and RippleNet [41] exploit multi-hop KG structure without assessing each node’s need.The review notes that translation-based KG signals may interfere with CF signals.Source: §2 · pp. 2–3
  • GNN-based methods such as KGCN [44], KGNN-LS [43], KGAT [45], and KGIN [46] aggregate graph information, but may mix IG and KG signals during propagation.Their neighbor or relation weights do not determine whether a node needs KG information.Source: §2 · p. 3
  • Self-supervised methods use reconstruction, contrastive learning, KG-connection scoring, or diffusion, but mainly refine KG signals or align views rather than control node-wise KG injection.Examples include KGCL [50], MCCLK [55], KGRec [49], DiffKG [13], and LightKG [24].Source: §2 · p. 3

3 Preliminaries

The preliminaries establish the graph notation and recommendation task. They define the interaction and knowledge graphs, a unified collaborative knowledge graph, and the top-K recommendation objective.

3 Preliminaries

  • The preliminaries introduce notation and formulate the KG-aware recommendation problem addressed in the paper.Frequently used symbols are summarized in Table 1.Source: §3 · p. 3, Table 1 · p. 3

3.1 Concepts and Notations

  • The interaction graph consists of users, items, and observed user-item interactions; each user’s neighbors are denoted by N_u.The graph is defined as G_IG = (U, I, E_IG).Source: §3.1 · p. 3
  • The KG represents item-side relationships, while the CKG unifies it with interactions by converting each interaction into a relation-labeled triple.The CKG is a relation-labeled multipartite graph over users, items, and entities.Source: §3.1 · p. 3
  • Given an interaction graph, an item-side KG, and a query user, the task is to learn model parameters for recommending a top-K ranked item list.Source: §3.1 · p. 3

4 Proposed Method

AdaKG separately encodes interaction-graph and knowledge-graph signals, estimates node-wise CF instability, and uses it to adapt KG fusion and align the two views. Its time complexity scales linearly with edges and nodes, and its parameter-space complexity scales linearly with nodes and relations.

4.2 Decoupled Dual-View Graph Encoding.

  • Separate IG and CKG encoders preserve view-specific signals: the IG captures collaborative filtering, while the CKG incorporates item-side knowledge into item and user embeddings.The CKG aggregates relation-aware messages from connected entities and items, then aggregates KG-aware item embeddings for users without propagating user embeddings back to KG nodes; layer-wise L2 normalization constrains CKG embedding magnitudes.Source: §4.2 · p. 4, §4.2.1 · p. 4, §4.2.2 · pp. 4–5

4.3 Estimating the Instability of CF Signals.

  • AdaKG estimates CF instability by perturbing initial embeddings in the direction that weakens pairwise ranking, then measuring how much each node’s final IG embedding changes.A larger mean-centered instability score indicates greater sensitivity to perturbation and a less stable CF signal; the perturbation uses recommendation-aware gradients.Source: §4.3 · p. 5, §4.3.1 · p. 5, §4.3.2 · p. 5

4.4.1 Adaptive fusion gate.

  • AdaKG uses each node’s CF-instability score to control its KG fusion weight, assigning more weight to KG embeddings for more unstable nodes.The learned weight increases monotonically with instability; the IG embedding remains the primary recommendation signal, while the KG embedding supplies auxiliary information.Source: §4.4 · p. 6, §4.4.1 · p. 6

4.4.2 Adaptive contrastive learning.

  • Adaptive contrastive learning aligns each node’s IG representation with its KG representation, using the learned fusion ratio to control alignment strength.Other nodes’ KG representations in the batch are negatives, and a stop-gradient operator prevents the contrastive loss from directly updating the fusion gate.Source: §4.4.2 · p. 6

4.5 Optimization for AdaKG

  • AdaKG optimizes recommendation with Bayesian Personalized Ranking, combining its loss with adaptive contrastive loss and L2 regularization.The coefficients λ_ACL and λ_reg control the contrastive-loss and regularization strengths, respectively.Source: §4.5 · p. 6

4.6 Complexity Analysis

  • AdaKG’s forward pass has time complexity O(Ld(m+n)), and its learnable parameters have space complexity O(d(n+r)).Here, L is the number of propagation layers, d the embedding dimension, m the total number of IG and KG edges, n the number of users, items, and entities, and r the number of relations.Source: §4.6 · p. 7
  • AdaKG’s time complexity is linear in the number of edges and nodes, while parameter-space complexity is linear in the number of nodes and relations.The paper says it compares training efficiency with state-of-the-art methods in Appendix C but provides no comparison results here.Source: §4.6 · p. 7

5 Experiments

The experiments evaluate AdaKG’s recommendation performance, component contributions, hyperparameter sensitivity, CF-stability patterns, and graph encoders.

5.1 Experimental Settings

  • For each user, observed interactions were randomly split into training, validation, and test sets at 8:1:1; evaluation used Recall@10 and MRR@10, and reported test performance averaged five runs with different random seeds.Source: §5.1 · p. 7

5.2 Recommendation Performance (Q1)

  • AdaKG is best on Amazon-Book, MovieLens-1M, and Last-FM; on Book-Crossing it is best in MRR@10 and second in Recall@10, with the best average rank across four datasets.The comparison uses Recall@10 and MRR@10 across four real-world datasets.Source: Table 3 · p. 8, §5.2 · p. 8

5.3 Ablation Study (Q2)

  • Ablations support AF, perturbation-based IE, and ACL: fixed fusion is comparable on MovieLens-1M but worse on sparser datasets, degree-based IE underperforms, and removing ACL degrades results.The ablation variants use a fixed fusion ratio, a node-degree heuristic, or no adaptive contrastive loss.Source: Figure 3 · p. 8, §5.3 · pp. 8–9

5.4 Hyperparameter Sensitivity (Q3)

  • Within the evaluated range, increasing 𝜖 improves performance on Amazon-Book and Book-Crossing; optimal 𝜆ACL varies by dataset, while 𝜆ACL = 0.5 degrades performance.The analysis links excessive contrastive alignment to reduced view-specific information and weaker adaptive fusion.Source: Figure 4 · p. 9, §5.4 · p. 9

5.5 Qualitative Analysis of CF-Stable Users (Q4)

  • Across all datasets, CF-stable users have smaller mean embedding distances; on Last-FM, unstable users appear more dispersed even at comparable interaction degrees.The analysis identifies stable and unstable groups using the top and bottom 10% of users ranked by instability score.Source: Figure 5 · p. 9, §5.5 · p. 9

5.6 Analysis of Graph Encoders (Q5)

  • AdaKG consistently outperforms either encoder alone; CKG helps on Book-Crossing but underperforms IG on MovieLens-1M, supporting complementary views and node-wise fusion.The reported performance gain indicates that adaptive fusion balances IG and CKG contributions for each node.Source: §5.6 · pp. 9–10, Figure 6 · p. 10

6 Conclusion

AdaKG separately encodes interaction and knowledge graphs, estimates each node’s CF reliability through perturbation-based instability, and adjusts KG contributions accordingly. Experiments on four real-world datasets show strong performance against state-of-the-art KG-aware methods and support the adaptive fusion strategy.

6 Conclusion

  • AdaKG assigns larger KG contributions to nodes with less stable CF signals while preserving reliable collaborative signals as the primary basis.This node-wise adjustment avoids indiscriminate use of KG information.Source: §6 · p. 10
  • Across four real-world datasets, AdaKG achieves strong performance compared with state-of-the-art KG-aware methods.The conclusion reports strong comparative performance without specifying metric values.Source: §6 · p. 10
  • The paper concludes that KG contribution strength should vary with the reliability of each node’s collaborative signals.Further analyses validate the effectiveness of AdaKG’s adaptive fusion strategy.Source: §6 · p. 10

A Analysis of Designs of the CKG encoder

The CKG encoder design analysis finds that including initial embeddings and normalizing output embeddings improve performance, while normalizing initial embeddings degrades it.

A Analysis of Designs of the CKG encoder

B Effect of Adversarial Perturbation

The perturbation analysis reports that adversarial perturbations outperform random noise-based perturbations on both datasets and are more effective for identifying CF-unstable nodes.

B Effect of Adversarial Perturbation

C Training Efficiency Analysis

AdaKG is substantially faster than KGRec across datasets, while slightly slower than LightKG; the paper reports that it still maintains practical training efficiency while providing better recommendation.

C Training Efficiency Analysis

  • AdaKG is substantially faster than KGRec across all datasets.The paper reports that adaptive fusion does not introduce heavy computational overhead.Source: C Training Efficiency Analysis · p. 10
  • AdaKG is slightly slower than the highly lightweight LightKG.The comparison concerns wall-clock training efficiency.Source: C Training Efficiency Analysis · p. 10
  • Despite being slightly slower than LightKG, AdaKG maintains practical training efficiency while providing better recommendation.The recommendation comparison is referenced to Table 3.Source: C Training Efficiency Analysis · p. 10

D Details of Hyperparameter Settings

AdaKG’s hyperparameters were tuned over specified ranges, with temperature fixed at 0.2; validated dataset configurations are available in the code repository.

D Details of Hyperparameter Settings

  • The learning rate was searched from 10−2 to 10−4, layer count from 1 to 4, and weight decay from 10−3 to 10−5.The learning-rate candidates included 5×10−3; layer counts were {1, 2, 3, 4}, and weight-decay values were {10−3, 10−4, 10−5}.Source: D Details of Hyperparameter Settings · p. 10
  • The adaptive contrastive learning ratio ranged from 10−2 to 5 × 10−1, while perturbation rate candidates ranged from 0.1 to 1.0.The instability distance function was selected from cosine and normalized Euclidean.Source: D Details of Hyperparameter Settings · p. 10
  • Temperature was fixed at 0.2, and validated configurations for each dataset are available in the code repository.The instability distance function was searched over cosine and normalized Euclidean.Source: D Details of Hyperparameter Settings · p. 10

GenAI Usage Disclosure

The authors report limited generative AI use and state that methodological decisions and experimental-result analysis and verification were their own.

GenAI Usage Disclosure

Loading 2609.05909v1…