Source-linked AI summary
Knowledge Graph Convolutional Networks for Recommender Systems
Hongwei Wang, Miao Zhao, Xing Xie, Wenjie Li, Minyi Guo
TL;DR
Collaborative filtering is limited by sparse interactions and cold-start cases, motivating methods that exploit connected user and item attributes. KGCN aggregates sampled KG neighborhoods with personalized, relation-dependent weighting and extends this process across hops. Across movie, book, and music recommendation, KGCN consistently outperforms state-of-the-art baselines, while the paper identifies uniform neighbor sampling and item-end-only KGs as directions for future work.
Problem
Collaborative filtering faces sparse user-item interactions and cold-start problems, motivating the use of user and item attributes.
Method
KGCN samples fixed-size KG neighborhoods and aggregates them with relation- and user-dependent bias across multiple hops to learn structural, semantic, and personalized representations.
Results
KGCN consistently outperforms state-of-the-art baselines in movie, book, and music recommendation, with average AUC gains of 4.4%, 8.1%, and 6.2%, respectively.
Takeaways & Limitations
KGCN provides an end-to-end, minibatch recommendation framework that operates on large datasets and knowledge graphs while modeling users’ personalized and potential interests.
Takeaways & Limitations
The paper uniformly samples entity neighbors, focuses on item-end KGs, and leaves non-uniform sampling and user-end KG integration for future work.
Abstract
from arXiv · showhide
To alleviate sparsity and cold start problem of collaborative filtering based recommender systems, researchers and engineers usually collect attributes of users and items, and design delicate algorithms to exploit these additional information. In general, the attributes are not isolated but connected with each other, which forms a knowledge graph (KG). In this paper, we propose Knowledge Graph Convolutional Networks (KGCN), an end-to-end framework that captures inter-item relatedness effectively by mining their associated attributes on the KG. To automatically discover both high-order structure information and semantic information of the KG, we sample from the neighbors for each entity in the KG as their receptive field, then combine neighborhood information with bias when calculating the representation of a given entity. The receptive field can be extended to multiple hops away to model high-order proximity information and capture users' potential long-distance interests. Moreover, we implement the proposed KGCN in a minibatch fashion, which enables our model to operate on large datasets and KGs. We apply the proposed model to three datasets about movie, book, and music recommendation, and experiment results demonstrate that our approach outperforms strong recommender baselines.
1 INTRODUCTION
The paper frames KG-aware recommendation as a response to collaborative filtering’s sparsity and cold-start limitations, proposing KGCN to learn personalized, high-order information from knowledge graphs. Experiments across movie, book, and music datasets show gains over strong recommendation baselines.
- Collaborative filtering suffers from sparse user-item interactions and the cold-start problem, motivating the use of user and item attributes.
- Knowledge graphs connect item attributes through typed relations, providing semantic relatedness, interest expansion, and explainability for recommendation.
- KGCN aggregates neighborhood information with a user- and relation-dependent bias to capture KG structure, semantics, and personalized relational interests.
- Fixed-size neighborhood sampling makes KGCN’s computational cost predictable despite variable or very large entity neighborhoods.
- KGCN extends receptive fields across multiple hops to model high-order entity dependencies and users’ potential long-distance interests.
- 4.4%, 8.1%, and 6.2% average AUC gains were achieved on movie, book, and music recommendation, respectively, versus state-of-the-art baselines.
2 RELATED WORK
Related work includes spectral and non-spectral graph convolutional methods, as well as graph-based recommendation models. The paper distinguishes KGCN by targeting heterogeneous knowledge graphs, whereas PinSage and GAT are designed for homogeneous graphs.
- Spectral GCN methods perform graph convolution in spectral space, using Fourier-domain definitions, graph-Laplacian approximations, or eigendecomposition-based constructions.
- PinSage and GAT are related graph-based models, but both are designed for homogeneous graphs rather than heterogeneous knowledge graphs.
3 KNOWLEDGE GRAPH CONVOLUTIONAL NETWORKS
KGCN predicts user interest by aggregating user-biased neighborhood information over sampled, multi-hop receptive fields in a knowledge graph. Its layered algorithm combines entity and neighborhood representations, then uses the resulting item representation for prediction.
- 3.1 Problem Formulation: KGCN formulates recommendation from user-item interactions and a knowledge graph of entity-relation-entity triples.The model predicts whether a user will engage with an unseen item.
- 3.2 KGCN Layer: A user-relation score estimates how important each relation is to that user.These scores personalize neighborhood aggregation, such as distinguishing interest in movie stars from genres.
- 3.2 KGCN Layer: The model combines each entity representation with its user-biased neighborhood representation through sum, concatenation, or neighbor aggregation.The three aggregator choices are evaluated experimentally.
- 3.2 KGCN Layer: KGCN samples a fixed-size neighborhood for each entity, defining a receptive field that can extend across multiple hops.Sampling keeps batch computation fixed and efficient, while deeper layers incorporate entities farther from the target.
- 3.3 Learning Algorithm: After H aggregation iterations, an entity representation mixes initial representations from the entity and neighbors up to H hops away.The final representation is combined with the user representation to predict engagement probability.
13 Function Get-Receptive-Field (v)
The training procedure avoids traversing every possible user-item pair by using negative sampling and optimizing a regularized cross-entropy objective. Negative samples are drawn uniformly, with the number per user matched to that user’s positive interactions.
- 13 Function Get-Receptive-Field (v): KGCN uses negative sampling during training to make computation more efficient.The procedure otherwise traverses all possible user-item pairs.
- 13 Function Get-Receptive-Field (v): The loss combines cross-entropy over sampled examples with an L2 regularizer.J denotes cross-entropy loss and the final term is the L2 regularizer.
- 13 Function Get-Receptive-Field (v): For each user, the number of negative samples equals the number of items with positive feedback, and negatives follow a uniform distribution.This setting is specified through T_u and P in the training objective.
4 EXPERIMENTS
KGCN is evaluated on movie, book, and music recommendation using multiple baselines, datasets, and sensitivity analyses. It generally performs best, with stronger gains in sparse book and music settings and sensitivity to neighborhood size, receptive-field depth, and embedding dimension.
- 4.1 Datasets: KGCN is evaluated on MovieLens-20M, Book-Crossing, and Last.FM for movie, book, and music recommendation.The experiments transform explicit feedback into implicit feedback and construct knowledge graphs from Microsoft Satori.
- 4.2 Baselines: The comparison includes KG-free SVD and LibFM alongside KG-aware LibFM + TransE, PER, CKE, and RippleNet.SVD models user-item interactions with an inner product, while PER uses manually designed meta-path features.
- 4.4 Results: KGCN improves more on book and music than movie, while KG-free SVD and LibFM outperform PER and CKE among the compared baselines.LibFM + TransE is better than LibFM in most cases, and RippleNet also performs strongly while using multi-hop neighborhood structure.
- 4.4 Results: KGCN-sum performs best in general, whereas KGCN-neighbor and KGCN-avg perform worse, especially on sparse Book-Crossing and Last.FM.KGCN-avg directly averages neighborhood representations without user-relation scores, providing an examination of the attention mechanism.
- 4.4 Results: KGCN performs best with K = 4 or 8 and is more sensitive to receptive-field depth, with serious model collapse at H = 3 or 4.The experiments attribute excessive K and H to noise, while H = 1 or 2 is sufficient in the reported real cases; increasing embedding dimension eventually causes overfitting.
5 CONCLUSIONS AND FUTURE WORK
KGCN extends non-spectral GCNs to knowledge graphs, selectively aggregating neighborhood information to learn structural, semantic, and personalized interests. Experiments show consistent improvements over state-of-the-art baselines, while future work targets broader sampling and user-end knowledge graphs.
- KGCN selectively and biasedly aggregates neighborhood information to learn KG structure, semantics, and users’ personalized and potential interests.
- KGCN’s minibatch implementation enables operation on large datasets and knowledge graphs.
- KGCN consistently outperforms state-of-the-art baselines across movie, book, and music recommendation.
- Future work includes non-uniform neighbor sampling, user-end knowledge graphs, and combining knowledge graphs at both ends.