Beyond modality importance: Reliability-guided tree memory for inductive multimodal knowledge graph completion

Jingchao Wang

doi:10.1016/j.knosys.2026.116953

Beyond modality importance: Reliability-guided tree memory for inductive multimodal knowledge graph completion

2026-09-01

0

PRE

AI

DOI：10.1016/j.knosys.2026.116953

Original

Original request for help

Save

Abstract

En 中文

Inductive multimodal knowledge graph completion (MKGC) aims to predict missing facts involving unseen entities by jointly utilizing structural, textual, and visual information. In recent years, retrieval enhancement methods have improved reasoning capabilities for unseen entities through semantic neighborhood retrieval, but still face two key challenges. First, these methods typically rely on attention or gating mechanisms to measure the importance of different modalities, but modality weights cannot determine whether a modality can reliably distinguish the correct entity from similar candidates under the current relation, potentially leading to high-weight but low-discriminative modality cues misleading the final prediction. Second, existing methods often retrieve semantic neighborhood entities from a flat candidate space memory, which easily introduces semantically similar but irrelevant interfering entities, further weakening the model’s ability to distinguish similar candidate entities. To address these challenges, we propose ReTree, a reliability-guided tree memory framework for inductive MKGC. ReTree first learns unified multimodal representations of target triples and evaluates the reliability of textual, visual, and fused representations from relation-level calibration and local candidate separation. It then builds an adaptive prototype-based tree memory that organizes training samples into relation roots, prototype nodes, and entity nodes. Reliability scores further guide prototype refinement and entity assignment, producing more discriminative candidate organization within each relation. During inference, ReTree combines hierarchical retrieval over the tree memory with query-local insertion for unseen entities, enabling more effective identification of similar candidates. Experiments on multiple inductive MKGC benchmarks show that ReTree outperforms existing methods in both prediction performance and generalization ability.

Keywords:

Knowledge graph

Multimodal knowledge graph

Knowledge graph completion

Multimodal fusion

Retrieval-based reasoning

Beyond modality importance: Reliability-guided tree memory for inductive multimodal knowledge graph completion

Abstract

Journal

Organization

Cited Papers