Source-linked AI summary
Jointly Learning Explainable Rules for Recommendation with Knowledge Graph
Weizhi Ma, Min Zhang, Yue Cao, Woojeong, Jin, Chenyang Wang, Yiqun Liu, Shaoping Ma, Xiang Ren
TL;DR
Existing recommender systems struggle to provide explanations, while knowledge-graph approaches may require manual rule design. RuleRec jointly induces explainable rules from item-centric knowledge graphs and trains rule-guided recommendation models, achieving better performance than baselines while retaining explanations. Its evaluations also expose sensitivity to noisy, loosely organized knowledge-graph information.
Problem
Neural embedding recommenders are difficult to explain, while meta-path methods require manual pattern design and existing explanation methods suffer from data sparsity.
Method
RuleRec jointly learns item-association rules from knowledge graphs and recommendation models that use the induced rules, sharing rule importance between modules.
Results
The four rule-enhanced BPRMF and NCF variants achieve remarkable results across multiple domains and outperform all baseline models while providing rule-based explanations.
Takeaways & Limitations
Induced rules can support accurate, explainable recommendation for different item-association types and improve recommendation models’ ability to handle cold-start cases.
Takeaways & Limitations
Knowledge-graph recommendation performance can degrade when linking datasets to a large but noisy, less topic-focused graph such as Freebase.
Abstract
from arXiv · showhide
Explainability and effectiveness are two key aspects for building recommender systems. Prior efforts mostly focus on incorporating side information to achieve better recommendation performance. However, these methods have some weaknesses: (1) prediction of neural network-based embedding methods are hard to explain and debug; (2) symbolic, graph-based approaches (e.g., meta path-based models) require manual efforts and domain knowledge to define patterns and rules, and ignore the item association types (e.g. substitutable and complementary). In this paper, we propose a novel joint learning framework to integrate \textit{induction of explainable rules from knowledge graph} with \textit{construction of a rule-guided neural recommendation model}. The framework encourages two modules to complement each other in generating effective and explainable recommendation: 1) inductive rules, mined from item-centric knowledge graphs, summarize common multi-hop relational patterns for inferring different item associations and provide human-readable explanation for model prediction; 2) recommendation module can be augmented by induced rules and thus have better generalization ability dealing with the cold-start issue. Extensive experiments\footnote{Code and data can be found at: \url{https://github.com/THUIR/RuleRec}} show that our proposed method has achieved significant improvements in item recommendation over baselines on real-world datasets. Our model demonstrates robust performance over "noisy" item knowledge graphs, generated by linking item names to related entities.
1 INTRODUCTION
The paper motivates explainable recommendation through item associations and proposes jointly learning rules from knowledge graphs with rule-guided recommendation.
- Neural recommendation methods often provide poor explanations, while user- and review-based explanations suffer from sparse social or review information.
- Knowledge-graph methods add structured information, but meta-path approaches require manually defined patterns and domain knowledge.
- Complementary and substitute item associations can distinguish useful recommendations, such as chargers or cases after a cellphone purchase from another cellphone.
- The proposed framework jointly learns explainable rules from knowledge graphs and recommendations guided by those rules.The rules summarize relational patterns and provide explanations for recommendations.
- The paper evaluates the framework extensively on real-world datasets and reports effective, accurate, and explainable recommendation.
2 PRELIMINARIES
The preliminaries define item recommendation, knowledge-graph rules, and the problem setting, then introduce BPRMF and NCF as base recommendation models.
- Item recommendation identifies suitable items for users from historical interactions such as purchases or ratings.
- A knowledge graph represents facts as entities connected by typed relational edges, while an inductive rule focuses on relation sequences rather than specific entities.
- The formal task jointly learns item-association rules and a recommender that uses those rules with each user’s interaction history.
- BPRMF represents users and items with latent vectors and computes preference scores using their inner product.
- NCF extends matrix factorization with generalized matrix-factorization and nonlinear multilayer-perceptron interactions before producing a final score.
3 THE RULEREC FRAMEWORK
RuleRec jointly learns item-association rules from a heterogeneous knowledge graph and uses them to guide recommendation. Shared rule weights connect rule learning with recommendation, producing features that support both prediction and explanation.
- The framework contains rule-learning and recommendation sub-tasks: deriving rules from item associations and recommending items using users’ histories and the derived rules.
- Multi-task Learning: Multi-task learning shares rule weights between association prediction and recommendation, optimizing recommendation and rule-learning objectives with a trade-off parameter.
- Heterogeneous Graph Construction: Items are linked to entities in a heterogeneous graph, including entity-linked nodes and edges for items that lack exact knowledge-graph matches.
- Rule Learning Module: Random walks estimate rule-following path probabilities between item pairs, forming relation-type rules and rule feature vectors.
- Rule Learning Module: Rule selection uses hard-selection methods based on chi-square scores or learned rule importance, alongside linear-regression and sigmoid objectives.
- Item Recommendation Module: The recommendation module combines conventional user-item scores with rule features, and the highest-contributing rule can explain why a candidate item is recommended.
4 RULE SELECTION DETAILS
Rule selection derives explainable rules from item-linked knowledge graphs, filters large-scale paths with support and length constraints, and evaluates their recommendation coverage. Chi-square selection is effective for retaining a compact rule subset.
- Dataset and Implementation Details: Freebase supplies the large knowledge graph used for rule learning, while item associations are obtained from Amazon user logs.The association types include substitute relations (ALV, BAV) and complementary relations (ALB, BT).
- Dataset and Implementation Details: Items are linked to DBPedia and then Freebase using titles, brands, and descriptions, ignoring links with confidence below 0.6.Entity-type constraints reduce the four-hop random-walk search from over 10 million entities to around 100K per domain.
- Rule Selection: Candidate rules must support at least fraction α of training item pairs and have maximum path length β; experiments set α=0.01 and β=4.The length constraint limits paths to at most four edges between entities.
- Rule Selection: Chi-square selection outperforms linear regression, with 50 selected ALB rules covering 87.0% of user purchase history versus the full-rule upper bound.The selected subset is only 2.2 percentage points below using all rules.
- Rule Selection: Multi-task learning does not require separate rule selection because rule effects are learned through the combined recommendation and rule-selection loss.This contrasts with the two-step framework, which explicitly selects rules before recommendation.
5 RECOMMENDATION EXPERIMENTS
Recommendation experiments compare RuleRec with matrix-factorization, neural, and knowledge-graph baselines on Amazon Cellphone and Electronic datasets. Rule-guided models outperform baselines broadly, with multi-task learning generally stronger than two-step learning and learned rules receiving high expert acceptance.
- Experimental Setup: Experiments evaluate RuleRec against traditional matrix factorization, neural-network, and knowledge-graph recommendation baselines.The recommendation datasets are Amazon Cellphone and Amazon Electronic, and evaluation uses leave-one-out ranking with 99 sampled negatives.
- Enhanced Algorithms: RuleRecmulti with BPRMF improves over BPRMF by 6.5% to 11.0% across metrics, while RuleRecmulti with NCF improves over NCF by 3.0% to 6.4% in Cellphone.RuleRecmulti with NCF is best on Cellphone, whereas RuleRecmulti with BPRMF is best on Electronic.
- Overall Performances: Every enhanced algorithm outperforms all baselines in both domains on each metric, with most improvements statistically significant.The results support the usefulness of knowledge-graph-derived rules for ranked item recommendation.
- Overall Performances: Multi-task RuleRec performs better than two-step RuleRec, indicating that jointly combining recommendation and rule-selection losses improves rule-weight learning.This advantage appears despite learning-based selection performing worse than chi-square selection in the separate rule-selection experiment.
- Knowledge-Graph Baselines: On noisy Freebase-linked datasets, HERec and RippleNet perform worse, which the authors attribute to the graph being less clean than compact, on-topic subgraphs.On MovieLens-1M, RuleRec outperforms HERec but remains below RippleNet; the dataset has only seven relation types and more complete connection coverage.
- Explainability Case Study: Over 94% of 100 learned ALB and BT rules are accepted by users, including 87% accepted by all three experts.The expert labeling evaluates whether selected rules would improve recommendation explainability in real scenarios.
- Component Analysis: Multi-task scoring is best across all tested metrics and associations, outperforming hard filtering, equal weighting, selection weighting, and joint learning.The comparison uses RuleRec(BPRMF) on the Cellphone dataset.
4. Study on single association vs. all associations
The study finds that rules derived from individual item associations improve RuleRecmulti with BPRMF, while combining all association types improves performance further. Increasing the rule count does not uniformly help, particularly for RuleRectwo with BPRMF.
- RuleRecmulti with BPRMF significantly outperforms BPRMF when rules are derived from any one of four item associations.
- Combining all association types outperforms using a single association, indicating that the combination contributes to recommendation performance.
- For RuleRectwo with BPRMF, increasing the rule count decreases recall@5 and MRR@10, while NDCG@10 remains stable and Recall@10 increases.
- Overall performance does not improve with more rules because additional rules may include many bad rules that the two-step model handles poorly.
6 RELATED WORK
Related work uses side information from knowledge graphs through meta-path and embedding-based methods, but these approaches have explainability or manual-design limitations. Rule-learning methods instead derive interpretable patterns, while this paper targets cold items through rule-based feature generation.
- Knowledge-graph recommendation: Meta-path methods use predefined paths and require manual effort and domain knowledge to define patterns for feature extraction.
- Additional experiments: Figures 7 and 8 examine performance under different numbers of rules for RuleRectwo and RuleRecmulti with BPRMF.
- Knowledge-graph recommendation: Embedding-based methods learn user and item representations from knowledge-graph structure but give up the graph’s explainable strength.
- Rule learning: Prior item-association prediction methods suffer from cold items, whereas this work mines meaningful rules that generate feature vectors without user reviews.
- Rule learning: Knowledge-graph reasoning research focuses on link prediction, while this study focuses on inducing rules from ground-truth entity pairs.
7 CONCLUSIONS AND FUTURE WORK
The paper presents a joint optimization framework that induces explainable rules from item knowledge graphs and incorporates them into recommendation models. Experiments report strong performance and explainability, while future work seeks an embedding-based combination that preserves explainability.
- The framework jointly induces rules from a knowledge graph and recommends items based on those induced rules.
- It contains rule learning and recommendation modules, implemented through two-step and jointly learning approaches.
- The four rule-enhanced algorithms outperform all baseline models across multiple domains while providing explanations for recommendations.
- Multi-task combinations and rules derived from different associations contribute to better recommendation results.
- Future work will investigate an embedding-learning combination algorithm that retains explainability with the knowledge graph.