Source-linked AI summary
Fairness-Aware Explainable Recommendation over Knowledge Graphs
Zuohui Fu, Yikun Xian, Ruoyuan Gao, Jieyu Zhao, Qiaoying Huang, Yingqiang Ge, Shuyuan Xu, Shijie Geng, Chirag Shah, Yongfeng Zhang, Gerard de Melo
TL;DR
The paper addresses unfair recommendation performance and explanation diversity caused by unequal user visibility and interaction data in knowledge-graph explainable recommendation. It applies fairness-constrained heuristic path reranking to existing models, and reports improved recommendation quality alongside reduced disparities across group and individual fairness measures.
Problem
Explainable recommendation can exhibit performance and explanation-diversity disparities because users with fewer interactions have lower visibility and less-captured preferences.
Method
The paper reranks candidate user–item paths using a fairness-constrained 0–1 integer-programming formulation solved with heuristic techniques.
Results
The fairness-aware algorithm improves recommendation and fairness evaluations across group and individual levels, including higher Clothing NDCG and lower group unfairness over PGPR.
Takeaways & Limitations
Fairness-aware path reranking can provide high-quality explainable recommendations while reducing disparities in performance and explanation diversity.
Takeaways & Limitations
Evaluation treats each Amazon dataset as an independent domain, so results are not comparable across domains; baseline results also have slight implementation discrepancies.
Abstract
from arXiv · showhide
There has been growing attention on fairness considerations recently, especially in the context of intelligent decision making systems. Explainable recommendation systems, in particular, may suffer from both explanation bias and performance disparity. In this paper, we analyze different groups of users according to their level of activity, and find that bias exists in recommendation performance between different groups. We show that inactive users may be more susceptible to receiving unsatisfactory recommendations, due to insufficient training data for the inactive users, and that their recommendations may be biased by the training records of more active users, due to the nature of collaborative filtering, which leads to an unfair treatment by the system. We propose a fairness constrained approach via heuristic re-ranking to mitigate this unfairness problem in the context of explainable recommendation over knowledge graphs. We experiment on several real-world datasets with state-of-the-art knowledge graph-based explainable recommendation algorithms. The promising results show that our algorithm is not only able to provide high-quality explainable recommendations, but also reduces the recommendation unfairness in several respects.
1 INTRODUCTION
The paper examines fairness in explainable recommendation, focusing on disparities in recommendation performance and explanation diversity among users with different visibility or activity levels. It proposes fairness-aware path reranking to improve recommendation quality while narrowing these disparities.
- Explainable recommendation adds personalized, intuitive explanations to high-quality recommendations but raises concerns about performance and explanation-diversity fairness.
- Users with fewer historical interactions may have poorly captured preferences, making them vulnerable to weaker recommendations and less diverse explanations.
- The paper treats user visibility to the recommender as an observable fairness-related property linked to platform activity.
- The study investigates how to quantify unfairness, improve recommendation quality and explanation diversity, assess group and individual fairness, and test generalizability.
- Experiments on four Amazon e-commerce datasets characterize data imbalance and its relationship to unfairness in knowledge-graph explainable recommendation.
- The proposed approach formulates fairness-aware path selection as a 0–1 integer program solved with heuristic techniques rather than retraining existing path-finding models.
2 RELATED WORK
Prior fairness research studies bias across decision-making and recommendation settings, while explainable recommendation over knowledge graphs uses relational paths to justify recommendations. This paper focuses specifically on disparities in user visibility to ranking algorithms.
- Fairness in Decision Making: Group fairness seeks equal treatment for protected and advantaged groups, whereas individual fairness seeks similar treatment for individuals with similar attributes.
- Fairness-aware Ranking and Recommendation: Prior work addresses recommendation fairness across consumers and providers, selection, marketing, popularity, and ranking bias.
- Fairness-aware Ranking and Recommendation: Causal-graph methods have been reported to outperform statistical-parity approaches for identifying and mitigating rank discrimination.
- Explainable Recommendation with Knowledge Graphs: Knowledge-graph explainable recommendation uses entities, relationships, and paths to support explanations across latent-factor, deep, social, visual, sequential, and dynamic recommendation settings.
3 PRELIMINARIES
The preliminaries represent recommendation knowledge as a graph of entities and relations, then define relation patterns and user–item paths that connect users to recommended items. These paths provide the explanations associated with recovered recommendations.
- Knowledge Graphs: A knowledge graph consists of triples connecting head and tail entities through relations; users and items form separate entity subsets in recommendation.
- Path Patterns: A relation pattern is a sequence of composed relations, and a path instantiates that pattern through consecutive entities and relations.
- Path Distributions: The user–item path distribution D_u,V assigns each pattern a share based on its occurrence frequency among paths from user u to items.
- Explainable Recommendation: Explainable recommendation over KGs recovers missing user–item facts and associates each recommended item with a user–item path as its explanation.
4 MOTIVATING FAIRNESS CONCERNS
The paper links unfair explainable recommendation to activity-based data imbalance and divergent user–item path distributions. Inactive users, who form the majority, receive lower recommendation scores and less diverse path patterns, while existing methods fit skewed overall distributions.
- Activity imbalance: Activity-based data imbalance leaves inactive users with fewer historical interactions and makes their preferences less visible to explainable recommendation models.The paper frames user activity as an observable visibility property rather than relying on sensitive demographic attributes.
- Activity imbalance: The four Amazon datasets divide users into an active top 5% and an inactive remainder, with inactive users comprising the majority.The analysis covers CDs and Vinyl, Clothing, Cell Phones, and Beauty.
- Path distributions: Inactive users have less diverse user–item path patterns than active users, indicating divergent explanation structures between the groups.Path patterns are treated as relational structures that can provide recommendation explanations and reveal explanation diversity.
- Recommendation disparity: The inactive group obtains far lower recommendation scores despite containing most users, sacrificing performance for the vast majority.The comparison is reported across recommendation algorithms with separate active-user and inactive-user results.
- Path distributions: Existing KG-based methods neglect path-distribution differences and fit the highly skewed overall distribution, contributing to unfair performance and explanation-diversity disparities.The paper identifies divergence between groups’ user–item path distributions as an essential factor in both disparities.
- Quantifying diversity: Simpson’s Index of Diversity is introduced to quantify path-distribution unfairness by accounting for both pattern richness and evenness.In the user analysis, SID measures the probability that two randomly selected paths share a pattern; larger values indicate greater diversity.
5 FAIRNESS OBJECTIVES
The paper formalizes fairness-aware explainable recommendation at group and individual levels, measuring disparities in recommendation performance and explanation diversity. It then treats fairness as constrained path reranking over candidates produced by an existing KG-based model.
- Group fairness: Group fairness compares active and inactive users, defined by historical purchase activity, and seeks parity across their recommendation outcomes.The active and inactive groups are disjoint and represent different levels of user visibility.
- Group fairness: Group recommendation unfairness measures disparity in recommendation quality between user groups, while group explanation diversity unfairness applies SID to their path distributions.Recommendation quality may use metrics such as NDCG@K or F1 score.
- Individual fairness: Individual fairness is adapted as a measurement objective because strict equal treatment is incompatible with addressing bias caused by inherent data imbalance.The paper follows the principle that similar individuals should be treated similarly without requiring absolute parity.
- Individual fairness: Individual recommendation unfairness uses the Gini coefficient to quantify inequality in benefits received across users.The Gini coefficient ranges from 0 for equality to 1 for maximal inequality.
- Problem formulation: The fairness-aware task ranks candidate user–item paths to maximize top-K recommendation quality while satisfying group constraints on recommendation and explanation-diversity unfairness.This avoids retraining existing path-finding models within the fairness-constrained process.
- Problem formulation: The individual fairness-aware task treats each user as a group and aims to minimize unfairness while retaining recommendation quality to the extent possible.The formulation considers both group-level and individual-level fairness.
6 FAIRNESS-AWARE ALGORITHM
The framework re-ranks recommendations under fairness constraints while preserving recommendation relevance and improving explanation-path diversity. It combines path quality, path diversity, and fairness scores, then uses heuristic optimization to select and order top-K items.
- Framework: The framework applies fairness-aware ranking on top of multiple knowledge graph-based explainable recommendation algorithms.It retains original explainable paths while incorporating fairness constraints into recommendation selection and ranking.
- Fairness scores: The algorithm evaluates fairness through path quality and explainable path diversity scores.These scores aggregate personalization and diversity considerations for each user.
- Fairness scores: The path score debiases historical user–item path patterns while retaining path-score weights from the original recommendation process.A regularization factor is used to reduce path-diversity bias between user groups.
- Fairness scores: The diversity score uses retrieved user–item path distributions and Simpson’s Index of Diversity to quantify explanation-path diversity.The method incorporates this diversity measure into fairness-aware recommendation ranking.
- Constrained selection: The recommendation objective selects K items that maximize preference scores while satisfying group fairness constraints.The formulation can also incorporate individual fairness constraints based on Gini coefficient differences.
- Constrained selection: The constrained optimization is a 0-1 integer programming problem solved with fast heuristics, which may reach a local rather than global optimum.After selection, items are ranked by optimized recommendation and fairness scores.
7 EXPERIMENTS
Experiments on four Amazon e-commerce benchmarks evaluate fairness-aware algorithms built on explainable recommendation methods over knowledge graphs. The approach improves recommendation quality while reducing disparities between active and inactive users and across fairness metrics.
- Dataset and Experimental Setup: The evaluation uses four Amazon domains as independent knowledge-graph benchmarks and compares fairness-aware algorithms with existing explainable recommendation approaches.The domains are CDs and Vinyl, Clothing, Cell Phones, and Beauty; results are not comparable across domains.
- Dataset and Experimental Setup: The experiments assess NDCG and F1 recommendation quality alongside group and individual fairness for recommendation performance and explanation diversity.Group Recommendation Unfairness (GRU) measures disparity between inactive and active users, while additional metrics address explainable diversity.
- Main Results: 3.101% NDCG versus 2.856% for vanilla PGPR, while GRU falls from 1.410% to 0.233% on Clothing.The fairness-aware PGPR sacrifices some performance for the most active users while substantially improving performance for inactive users.
- Study of fairness weight: Fairness-aware methods outperform vanilla HeteroEmbed and PGPR across α settings on GEDU and IRU, with unfairness minimized at α = 1.0.An appropriate α can improve IRU while maintaining more reliable GEDU and IRU behavior.
- Study of ranking weight: Increasing β consistently reduces GEDU and IEDU, while the β value minimizing IRU differs between Fair HeteroEmbed and Fair PGPR.The reported optimum is around β = 0.2 for Fair HeteroEmbed and β = 1 for Fair PGPR, partly attributed to PGPR preprocessing and inactive users’ comment-heavy interactions.
- Study of ranking weight: Across changes to α or β, GRU remains better than for the baseline methods.Recommendation quality for inactive users initially improves as β increases and then declines as β approaches 1 when α = 0.75.
- Main Results: The qualitative analysis finds six correct purchase recommendations connected by diverse explainable paths, compared with four items in vanilla PGPR.The fairness-aware method also adds related items through more comprehensive paths.
8 CONCLUSIONS
The paper quantifies fairness in knowledge-graph explainable recommendation at both group and individual levels and proposes a generalized fairness-aware algorithm. Experiments show reduced unfairness with diverse paths while maintaining strong explainable recommendation quality.
- Conclusions: The paper quantifies unfairness at both the individual and group levels for state-of-the-art explainable recommendation algorithms over knowledge graphs.It defines fairness metrics covering path diversity and recommendation-performance disparity.
- Conclusions: The generalized fairness-aware algorithm reduces disparity while maintaining recommendation quality and providing diverse path patterns.The approach is evaluated extensively on several real-world datasets.