TRIX: A More Expressive Model for Zero-shot Domain Transfer in Knowledge Graphs

TL;DR

TRIX addresses zero-shot domain transfer in knowledge graphs by enabling fully inductive learning with expressive triplet embeddings. It extends prior fully inductive models by supporting both entity and relation prediction in inductive settings and introduces a relation graph and iterative updates. Empirically, TRIX surpasses state-of-the-art fully inductive models in zero-shot and in-domain predictions and even large-context LLMs in out-of-domain scenarios, across 57 KGs. Training minimizes binary cross-entropy loss over triplets, as $ \text{Loss} = -\log p(h,r,t) - \sum_{i=1}^{n} \frac{1}{n} \log(1 - p(h_i^\prime, r, t_i^\prime)) $ for entities and $ $ \text{Loss} = -\log p(h,r,t) - \sum_{i=1}^{n} \frac{1}{n} \log(1 - p(h, r_i^\prime, t)) $ for relations, with $n$ negatives per positive. The authors provide code at https://github.com/yuchengz99/TRIX for reproducibility and further exploration.

Abstract

Fully inductive knowledge graph models can be trained on multiple domains and subsequently perform zero-shot knowledge graph completion (KGC) in new unseen domains. This is an important capability towards the goal of having foundation models for knowledge graphs. In this work, we introduce a more expressive and capable fully inductive model, dubbed TRIX, which not only yields strictly more expressive triplet embeddings (head entity, relation, tail entity) compared to state-of-the-art methods, but also introduces a new capability: directly handling both entity and relation prediction tasks in inductive settings. Empirically, we show that TRIX outperforms the state-of-the-art fully inductive models in zero-shot entity and relation predictions in new domains, and outperforms large-context LLMs in out-of-domain predictions. The source code is available at https://github.com/yuchengz99/TRIX.