CMKL: Modality-Aware Continual Learning for Evolving Biomedical Knowledge Graphs
Researchers have introduced the Continual Multimodal Knowledge Graph Learner (CMKL), a novel framework designed to address challenges in processing large, dynamic, and multimodal biomedical knowledge graphs. Existing methods often assume static structures or fail to effectively utilize multimodal data under evolving distributions, applying uniform regularization that ignores distinct forgetting dynamics across modalities. CMKL natively encodes structure, text, and molecular data, fusing them via a Mixture-of-Experts router while protecting prior knowledge through Elastic Weight Consolidation and a K-means-diverse replay buffer. Evaluated on a benchmark with 129,000 entities across ten tasks, CMKL achieved an Average Precision of 0.591 in entity classification, representing a 60% improvement over structural baselines with near-zero forgetting. In relationship prediction, it matched top sequential models and outperformed joint training approaches. The study highlights that modality asymmetry exists at the representation level, managed effectively by the model's routing mechanism. This advancement supports more accurate inference of unobserved biomedical relationships, driven by rapid biotechnological progress such as high-throughput sequencing.
Wire timeline
CMKL: Modality-Aware Continual Learning for Evolving Biomedical Knowledge Graphs
Researchers have introduced the Continual Multimodal Knowledge Graph Learner (CMKL), a novel framework designed to address challenges in processing large, dynamic, and multimodal biomedical knowledge graphs. Existing methods often assume static structures or fail to effectively utilize multimodal data under evolving distributions, applying uniform regularization that ignores distinct forgetting dynamics across modalities. CMKL natively encodes structure, text, and molecular data, fusing them via a Mixture-of-Experts router while protecting prior knowledge through Elastic Weight Consolidation and a K-means-diverse replay buffer. Evaluated on a benchmark with 129,000 entities across ten tasks, CMKL achieved an Average Precision of 0.591 in entity classification, representing a 60% improvement over structural baselines with near-zero forgetting. In relationship prediction, it matched top sequential models and outperformed joint training approaches. The study highlights that modality asymmetry exists at the representation level, managed effectively by the model's routing mechanism. This advancement supports more accurate inference of unobserved biomedical relationships, driven by rapid biotechnological progress such as high-throughput sequencing.
cs.AI updates on arXiv.org