Lightweight Inference-Time Personalization for Frozen Knowledge Graph Embeddings

TL;DR

This paper tackles the mismatch between cohort-level KG embeddings and individual user preferences by introducing GatedBias, an inference-time personalization mechanism that operates on frozen KG embeddings. The core idea is to attach interpretable, structure-gated biases computed from training graph structure and profile features, without backpropagating through the backbone, using only around $\sim 300$ trainable parameters. The final score is $s'(h,r,t)=s_\theta(h,r,t)+b_p(t)$, where $b_p(t)$ decomposes across relation groups via gates derived from the training graph. Evaluation on Amazon-Book and Last-FM shows that GatedBias improves personalization metrics like Alignment@k and Counterfactual Responsiveness while preserving standard link-prediction performance, with stronger effects in domains where semantic attributes clearly align with user preferences. The work demonstrates a practical, parameter-efficient pathway to tailor relational models to individual users and provides a framework for causal validation of personalization signals.

Abstract

Foundation models for knowledge graphs (KGs) achieve strong cohort-level performance in link prediction, yet fail to capture individual user preferences; a key disconnect between general relational reasoning and personalized ranking. We propose GatedBias, a lightweight inference-time personalization framework that adapts frozen KG embeddings to individual user contexts without retraining or compromising global accuracy. Our approach introduces structure-gated adaptation: profile-specific features combine with graph-derived binary gates to produce interpretable, per-entity biases, requiring only ${\sim}300$ trainable parameters. We evaluate GatedBias on two benchmark datasets (Amazon-Book and Last-FM), demonstrating statistically significant improvements in alignment metrics while preserving cohort performance. Counterfactual perturbation experiments validate causal responsiveness; entities benefiting from specific preference signals show 6--30$\times$ greater rank improvements when those signals are boosted. These results show that personalized adaptation of foundation models can be both parameter-efficient and causally verifiable, bridging general knowledge representations with individual user needs.

Figures (1)

• Figure 1: Overview of the GatedBias framework. A frozen knowledge-graph model is refined by lightweight gates that combine structural and profile features to re-rank entities, enabling efficient, interpretable personalization without retraining.