Time Imprint: new framework treats time as modality, boosting KG accuracy by 58%

Multi-modal knowledge graphs (MMKGs) struggle to distinguish entities with highly similar text and image features. Researchers from the University of Amsterdam propose Time Imprint, a framework that explicitly treats time as a separate, entity-level modality. The key innovation is a three-view contrastive objective that jointly aligns temporal, textual, and visual representations, overcoming the challenge of sparse temporal semantics. To handle entities with multiple timestamps, Time Imprint employs a compact timestamp subset selection mechanism followed by attention pooling, balancing specificity and robustness.

On three standard MMKG benchmarks, Time Imprint achieves state-of-the-art link prediction performance — improving Hits@1 by up to 6.07% overall and yielding up to 58% gains on the top-1% hardest ambiguity samples. The framework adds only modest training overhead, making it practical for production-scale graphs. The team also provides analysis on fusion strategies and timestamp quality sensitivity, clarifying when time-as-modality is most beneficial. Code and models are publicly released.