Source-linked AI summary
Learning over Knowledge-Base Embeddings for Recommendation
Yongfeng Zhang, Qingyao Ai, Xu Chen, Pengfei Wang
TL;DR
The paper addresses the difficulty of applying collaborative filtering to large-scale structured user behavior and item-property data while preserving heterogeneous relationships. It constructs a user-item knowledge graph, learns joint entity and relation embeddings, and uses them for recommendation. The resulting CFKG method consistently outperforms the best baseline across four datasets and all evaluation measures, although the improvement may reflect both additional information sources and the structured graph representation.
Problem
Existing collaborative-filtering systems mainly use unstructured data, while integrating heterogeneous structured behaviors and item properties with preserved relationships remains challenging.
Method
The approach constructs a heterogeneous user-item knowledge graph and extends collaborative filtering to learn over joint knowledge-base embeddings for personalized recommendation.
Results
CFKG consistently outperforms the best baseline, JRL, across four datasets and all evaluation measures.
Takeaways & Limitations
The approach supports personalized recommendation using multiple relation types while preserving structured knowledge about users and items.
Takeaways & Limitations
The reported improvement may benefit from both using more information sources and using a better structure to model heterogeneous information.
Abstract
from arXiv · showhide
State-of-the-art recommendation algorithms -- especially the collaborative filtering (CF) based approaches with shallow or deep models -- usually work with various unstructured information sources for recommendation, such as textual reviews, visual images, and various implicit or explicit feedbacks. Though structured knowledge bases were considered in content-based approaches, they have been largely neglected recently due to the availability of vast amount of data, and the learning power of many complex models. However, structured knowledge bases exhibit unique advantages in personalized recommendation systems. When the explicit knowledge about users and items is considered for recommendation, the system could provide highly customized recommendations based on users' historical behaviors. A great challenge for using knowledge bases for recommendation is how to integrated large-scale structured and unstructured data, while taking advantage of collaborative filtering for highly accurate performance. Recent achievements on knowledge base embedding sheds light on this problem, which makes it possible to learn user and item representations while preserving the structure of their relationship with external knowledge. In this work, we propose to reason over knowledge base embeddings for personalized recommendation. Specifically, we propose a knowledge base representation learning approach to embed heterogeneous entities for recommendation. Experimental results on real-world dataset verified the superior performance of our approach compared with state-of-the-art baselines.
1 INTRODUCTION
The paper extends collaborative filtering to structured user-item knowledge graphs, addressing the challenge of integrating heterogeneous behaviors and item properties while preserving their relationships. It proposes reasoning over knowledge-base embeddings and reports consistent gains over state-of-the-art baselines.
- Existing collaborative-filtering recommenders mainly use unstructured ratings, reviews, or images, making explicit relationships among user and item information difficult to model.
- The central challenge is integrating diverse user behaviors and item properties while preserving their internal relationships for personalized recommendation.
- The proposed framework combines traditional collaborative filtering with knowledge-base embeddings over a relational user-item knowledge graph.
- The approach directly reasons over structured knowledge-base embeddings and extends collaborative filtering to heterogeneous entities and relations, capturing user preferences more comprehensively.
- The model consistently outperforms many state-of-the-art baselines on real-world e-commerce datasets.
2 COLLABORATIVE FILTERING WITH KNOWLEDGE-GRAPH
The method represents heterogeneous recommender-system information as a user-item knowledge graph and learns entity and relation embeddings with collaborative filtering. Recommendations are generated by ranking items according to their learned relation-based distance to a user's embedding.
- Knowledge-Graph Construction: Each graph triplet consists of a head entity, tail entity, and directed relation representing a user behavior or item property.For example, buy connects a user to an item, while belong to connects an item to a brand or category.
- Knowledge-Graph Construction: The user-item knowledge graph unifies heterogeneous entities and relations, including products, brands, categories, words, and multiple behavioral or review relations.The system defines five entity types and six relation types, including buy, belong to category, belong to brand, mention word, also bought, and also viewed.
- Embedding Learning: The model projects entities and relations into a shared low-dimensional space and learns embeddings by making translated head embeddings close to their observed tail embeddings.The framework applies collaborative filtering over the graph for user profiling and personalized recommendation.
- Embedding Learning: A margin-based loss distinguishes observed triplets from corrupted triplets formed by replacing heads or tails with random entities.The distance uses the ℓ2-norm, and the translation function is implemented as transe(ei) = er + ei, following TransE.
- Recommendation: The loss is optimized with stochastic gradient descent to produce embeddings for all entities and relations.The learned embeddings are then used for recommendation through the buy relation.
- Recommendation: Personalized recommendations rank candidate items by ascending distance from the translated user embedding under the buy relation.
3 EXPERIMENTS
Experiments compare CFKG with recommendation baselines across Amazon datasets and evaluate how heterogeneous relations affect performance. CFKG consistently outperforms the best baseline, while adding relations improves performance and using all relations performs best.
- Experimental Setup: Experiments use four Amazon sub-datasets—CD, Clothing, Cell Phone, and Beauty—and evaluate top-N recommendation with Precision, Recall, Hit-Ratio, and NDCG.
- Experimental Setup: The baselines cover rating-based, review-based, image-based, and heterogeneous-information recommendation methods, including BPR, HFT, DeepCoNN, VBPR, CKE, and JRL.
- Performance Comparison: CFKG outperforms the best baseline, JRL, consistently across all four datasets and evaluation measures.
- Performance Comparison: Using only the buy relation, CFKG significantly outperforms BPR on all measures and datasets, with NDCG improvements of 90% for CDs, 70% for Clothing, 69% for Phone, and 33% for Beauty.
- Performance Comparison: CFKG buy+mention significantly outperforms BPR HFT and outperforms DeepCoNN except for recall on the CD dataset.
- Further Analysis on Different Relations: Adding any one extra relation improves the basic buy model, while modeling all heterogeneous relations outperforms partial-relation variants and all baselines.
4 RELATED WORK
Knowledge bases have long been viewed as useful for recommendation, but heterogeneous entities and relations make path reasoning difficult. Recent embedding techniques offer a way to preserve knowledge structure while supporting recommendation-related applications.
- Knowledge bases have attracted research attention as a way to enhance recommender systems since the field’s early stages.
- Reasoning over paths in heterogeneous knowledge graphs makes it difficult to apply collaborative filtering across different entities and relation types.
- Heterogeneous knowledge-base embeddings learn representations for different entities and support applications including question answering and relation extraction.
- Preserving knowledge structure during embedding is presented as vital for knowledge-enhanced AI in recommendation systems.
5 CONCLUSIONS AND FUTURE WORK
The paper learns heterogeneous knowledge-base embeddings for personalized recommendation by combining user behaviors and item knowledge in a user-item knowledge graph. Experiments on real-world datasets report superior performance and flexibility across multiple relation types.
- The approach constructs a user-item knowledge graph incorporating user behaviors and knowledge about items.
- It jointly learns knowledge-base embeddings across heterogeneous relations and uses user and item embeddings to generate personalized recommendations.
- Experimental results on real-world datasets verified superior performance and flexibility in incorporating multiple relation types.