TY - JOUR
T1 - Multimodal knowledge graph completion method for addressing imbalance of modal information
AU - Tian, Ronghua
AU - Yu, Hong
AU - Geng, Yaogang
AU - Wang, Xiaoling
N1 - Publisher Copyright:
© The Author(s), under exclusive licence to Springer Science+Business Media LLC, part of Springer Nature 2026.
PY - 2026/12
Y1 - 2026/12
N2 - Multimodal Knowledge Graph Completion (MMKGC) focuses on predicting missing triples in multimodal knowledge graphs by effectively leveraging both relational triples and multimodal information. However, existing MMKGC methods often overlook the imbalance of modal information among entities, resulting in an overreliance on the dominant modalities and degrading the performance of completion. To address the problem caused by the modality imbalance, a novel Multimodal Knowledge Graph Completion method (shorted by MI-MKGC) is proposed in this paper. It addresses the problem from two aspects: (1) the relation knowledge-guided dynamic modal fusion module dynamically adjusts the weights of each modality by using the relational types between entities as prior knowledge, and (2) the joint modality adversarial training module generates negative samples to enhance weak modality information. The extensive experimental results on three public datasets with 16 state-of-the-art baselines have verified the advancement of the proposed method. Specially, the metric Hit@1 of MI-MKGC increases by up to 5.2% than that of the best exiting compared method. The implementation code is available at https://github.com/yuhongcqupt/MI-MKGC.
AB - Multimodal Knowledge Graph Completion (MMKGC) focuses on predicting missing triples in multimodal knowledge graphs by effectively leveraging both relational triples and multimodal information. However, existing MMKGC methods often overlook the imbalance of modal information among entities, resulting in an overreliance on the dominant modalities and degrading the performance of completion. To address the problem caused by the modality imbalance, a novel Multimodal Knowledge Graph Completion method (shorted by MI-MKGC) is proposed in this paper. It addresses the problem from two aspects: (1) the relation knowledge-guided dynamic modal fusion module dynamically adjusts the weights of each modality by using the relational types between entities as prior knowledge, and (2) the joint modality adversarial training module generates negative samples to enhance weak modality information. The extensive experimental results on three public datasets with 16 state-of-the-art baselines have verified the advancement of the proposed method. Specially, the metric Hit@1 of MI-MKGC increases by up to 5.2% than that of the best exiting compared method. The implementation code is available at https://github.com/yuhongcqupt/MI-MKGC.
KW - Adversarial training
KW - Imbalanced modality information
KW - Knowledge graph completion
KW - Multi-modal knowledge graph
UR - https://www.scopus.com/pages/publications/105051858860
U2 - 10.1007/s10618-026-01272-1
DO - 10.1007/s10618-026-01272-1
M3 - 文章
AN - SCOPUS:105051858860
SN - 1384-5810
VL - 40
JO - Data Mining and Knowledge Discovery
JF - Data Mining and Knowledge Discovery
IS - 6
M1 - 106
ER -