Source-linked AI summary
RippleNet: Propagating User Preferences on the Knowledge Graph for Recommender Systems
Hongwei Wang, Fuzheng Zhang, Jialin Wang, Miao Zhao, Wenjie Li, Xing Xie, Minyi Guo
TL;DR
RippleNet addresses knowledge-graph-aware recommendation by propagating users’ preferences from historical interests through knowledge-graph links to predict click probabilities. Across movie, book, and news recommendation, it outperforms state-of-the-art baselines and offers a knowledge-graph-based perspective on explainability.
Problem
Knowledge-graph-aware recommendation seeks to predict whether a user will click an unseen item from interaction data and a knowledge graph, amid limitations in existing embedding-based and path-based methods.
Method
RippleNet uses historical interests as seeds, iteratively propagates preferences through knowledge-graph links into ripple sets, and combines their responses with item embeddings for click-through-rate prediction.
Results
RippleNet achieves AUC gains of 2.0% to 40.6% in movie, 2.5% to 17.4% in book, and 2.6% to 22.4% in news recommendation over state-of-the-art baselines.
Takeaways & Limitations
RippleNet provides a knowledge-graph-based explainability perspective by tracking paths from users’ histories to highly relevant items.
Takeaways & Limitations
The framework’s future work includes characterizing entity-relation interactions and designing non-uniform samplers to better explore potential interests.
Abstract
from arXiv · showhide
To address the sparsity and cold start problem of collaborative filtering, researchers usually make use of side information, such as social networks or item attributes, to improve recommendation performance. This paper considers the knowledge graph as the source of side information. To address the limitations of existing embedding-based and path-based methods for knowledge-graph-aware recommendation, we propose Ripple Network, an end-to-end framework that naturally incorporates the knowledge graph into recommender systems. Similar to actual ripples propagating on the surface of water, Ripple Network stimulates the propagation of user preferences over the set of knowledge entities by automatically and iteratively extending a user's potential interests along links in the knowledge graph. The multiple "ripples" activated by a user's historically clicked items are thus superposed to form the preference distribution of the user with respect to a candidate item, which could be used for predicting the final clicking probability. Through extensive experiments on real-world datasets, we demonstrate that Ripple Network achieves substantial gains in a variety of scenarios, including movie, book and news recommendation, over several state-of-the-art baselines.
1 INTRODUCTION
The paper motivates knowledge-graph-aware recommendation as a response to collaborative filtering’s sparsity and cold-start limitations, then introduces RippleNet to propagate user preferences through the knowledge graph. Experiments across movie, book, and news recommendation report gains over state-of-the-art baselines.
- Collaborative filtering suffers from sparse user-item interactions and the cold-start problem, motivating the use of side information.
- Knowledge graphs provide item facts, semantic connections, diverse relation types, and links between history and recommendations that support precision, diversity, and explainability.
- Existing KG-aware recommenders use either embedding-based methods or path-based methods, with path-based approaches relying heavily on manually designed meta-paths or meta-graphs.
- RippleNet is an end-to-end framework that combines embedding-based and path-based advantages through iterative preference propagation in the knowledge graph.
- 2.0% to 40.6%, 2.5% to 17.4%, and 2.6% to 22.4% AUC gains are reported for movie, book, and news recommendation, respectively, versus state-of-the-art baselines.
2 PROBLEM FORMULATION
The problem formulation combines implicit user-item interactions with a knowledge graph of entity-relation-entity triples. The goal is to predict a user’s potential interest in a previously unseen item.
- The user-item interaction matrix records implicit feedback, assigning 1 to observed interactions and 0 otherwise.
- The knowledge graph consists of triples (h,r,t), where entities are connected by typed relations and items may link to one or more graph entities.
- Given interaction matrix Y and knowledge graph G, the model predicts interest in an item with which the user has not interacted before.
- The learned function outputs ŷuv, the probability that user u will click item v, using model parameters Θ.
3 RIPPLENET
This section discusses RippleNet, including the proposed model and related work.
- The section discusses the proposed RippleNet model.
- The section includes discussion of the model.
- The section introduces related work.
3.1 Framework
RippleNet takes a user-item pair, propagates the user’s historical interests through multiple knowledge-graph ripple sets, and predicts the user’s click probability.
- RippleNet takes user u and item v as input and outputs the predicted probability that user u will click item v.
- The user’s historical interests serve as seeds that are extended along knowledge-graph links to form ripple sets at successive hops.
- Ripple sets interact iteratively with the item embedding to obtain user responses that are combined into a final user embedding.
- The user and item embeddings are combined to compute the predicted click probability.
3.2 Ripple Set
Ripple sets extend a user’s clicked items through knowledge-graph links, organizing potential interests by hop distance. Their size and relevance require practical controls as propagation moves farther from the history.
- The user’s past clicked items form the seed set for recursively identifying k-hop relevant entities in the knowledge graph.
- A k-hop ripple set contains knowledge triples beginning from the user’s (k−1)-hop relevant entities.
- Ripple propagation activates potential interests layer by layer from historical preferences toward more distant entities, with weaker preference strength at larger hops.
- Ripple sets can grow too large with hop number, so scenario-specific relations, limited maximum hops, and fixed-size neighbor sampling reduce relevance or computation costs.
3.3 Preference Propagation
RippleNet propagates a user’s preference through relation-aware knowledge-graph neighborhoods and combines multi-hop responses with the candidate item to predict clicking probability.
- RippleNet assigns each one-hop triple a relevance probability by comparing the candidate item with its head entity in the relation’s embedding space.The relation embedding matters because item–entity similarity can differ across relations.
- It computes a one-hop user response by taking the relevance-weighted sum of tail-entity embeddings in the ripple set.
- The framework takes a user–item pair as input and outputs a predicted click probability using a sigmoid over the user and item embeddings.
- Replacing the candidate item embedding with the previous response repeats preference propagation across ripple sets up to H hops.
- RippleNet combines responses from different propagation orders because shorter-hop information may be diluted in the final-hop response.
3.4 Learning Algorithm
RippleNet learns its parameters by jointly modeling implicit user feedback and knowledge-graph structure with a Bayesian objective optimized by stochastic gradient descent.
- The model maximizes the posterior probability of its parameters after observing the knowledge graph and implicit-feedback matrix.
- A Gaussian prior with zero mean and diagonal covariance is assigned to the model parameters.
- RippleNet uses three-way tensor factorization to define the knowledge-graph likelihood.
- The learning algorithm initializes parameters, samples positive and negative interactions plus true and false graph triples, and computes minibatch gradients by back-propagation.
- The loss combines interaction cross-entropy, knowledge-graph reconstruction error, and parameter regularization.
3.5 Discussion
RippleNet connects knowledge-graph preference propagation with explanation paths and models overlapping paths as superposed interest signals. These mechanisms provide interpretive structure beyond prediction alone.
- Explainability: RippleNet can explain recommendations by tracking highly relevant paths from a user’s history to a candidate item in the knowledge graph.
- Explainability: Unlike path-based methods with manually designed path patterns, RippleNet automatically discovers explanation paths using relevance probabilities.
- Ripple superposition: Overlapping paths from multiple clicked items increase interest in entities reached through those parallel connections, a phenomenon RippleNet calls ripple superposition.
3.6 Links to Existing Work
RippleNet connects to attention, memory-network, and knowledge-graph-embedding research while combining embedding-based and path-based ideas for recommendation.
- Attention mechanisms: RippleNet uses a multi-level attention module based on knowledge triples for preference propagation.
- Memory networks: Unlike typical memory-network applications using entry-level or sentence-level memories, RippleNet propagates preferences over knowledge-graph entities.
- Knowledge graph embedding: RippleNet can be viewed as a recommendation-oriented knowledge-graph-embedding method rather than a method aimed primarily at link prediction or triple classification.
4 EXPERIMENTS
The experiments evaluate RippleNet on movie, book, and news recommendation using knowledge-graph-enhanced datasets, controlled comparisons, and CTR and top-K metrics. Results favor RippleNet across datasets, while sensitivity studies identify moderate ripple-set sizes, two or three hops, and tuned embedding regularization as important settings.
- Datasets: RippleNet is evaluated on MovieLens-1M, Book-Crossing, and Bing-News, combining explicit ratings with implicit news feedback.The datasets contain approximately 1 million, 1,149,780, and 1,025,192 interactions, respectively.
- Experiment setup: The evaluation covers CTR prediction and top-K recommendation using Accuracy, AUC, Precision@K, Recall@K, and F1@K.Experiments use a 6:2:2 train/evaluation/test split, repeat each experiment five times, and report averages.
- Baseline comparisons: DKN performs best for news but worst for movies and books, while PER performs unsatisfactorily on movie and book recommendation.The reported explanations involve ambiguous short movie and book names and manually defined meta-paths that are difficult to optimize.
- Results: RippleNet outperforms baselines by 2.0% to 40.6%, 2.5% to 17.4%, and 2.6% to 22.4% on AUC for movie, book, and news recommendation, respectively.It also achieves outstanding top-K recommendation performance, although Bing-News top-K performance is lower because its item set is much larger.
- Sensitivity analysis: Ripple-set performance initially improves as the set grows but drops when it becomes too large; sizes of 16 or 32 suffice for most datasets.Larger sets encode more knowledge, but excessive size reduces performance.
5 CONCLUSION AND FUTURE WORK
The paper concludes that RippleNet integrates knowledge graphs into recommendation through preference propagation and achieves superiority over strong baselines. Future work targets richer entity-relation modeling and improved sampling during propagation.
- Conclusion: RippleNet integrates preference propagation with knowledge-graph-embedding regularization in a Bayesian framework for click-through-rate prediction.
- Conclusion: Experiments across three recommendation scenarios demonstrate RippleNet’s significant superiority over strong baselines.
- Future work: Future work will investigate entity-relation interactions and non-uniform samplers for better exploration of users’ potential interests.