FITTER: Vocabulary-Agnostic Cross-Domain Inference on Temporal Knowledge Graphs
This paper introduces FITTER, the first fully-inductive, vocabulary-agnostic structural model for temporal knowledge graph link prediction that enables cross-domain transfer by representing predicates through relative interaction patterns and temporal encodings, thereby outperforming existing baselines without retraining on unseen entities, relations, and timestamps.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine the internet as a massive, ever-growing library where every fact is a card in a giant deck. Some cards tell us who is friends with whom, while others tell us when those friendships started or ended. In the world of computer science, this is called a "Knowledge Graph." It's like a super-organized map of the world's facts. But the real world changes constantly; people meet, wars start, and treaties are signed every second. To keep up, scientists use "Temporal Knowledge Graphs," which are these maps with a built-in clock, tracking not just what happened, but when.
For a long time, computers trying to learn from these maps had a major limitation: they were like students who only studied for one specific test. If a computer learned the rules of a graph about diplomatic events (like peace talks), it couldn't understand a graph about encyclopedia facts (like who was born when), even if the logic was similar. It was as if the computer had memorized the names of every student in one school but was completely lost when walking into a different school with different names. This made it hard to teach computers to be flexible, especially when they encountered new data they had never seen before.
Enter FITTER, a new approach developed by researchers that acts like a universal translator for these time-traveling maps. Instead of memorizing specific names or dates, FITTER learns the patterns of how things connect and change over time. Think of it like learning the rules of a game rather than memorizing the players' names. If you know how a "negotiation" usually follows a "conflict" in a story, you can understand that pattern whether the characters are named "France" and "Germany" or "Japan" and "China."
The paper introduces FITTER as the first model capable of "fully-inductive" inference. This means it can take what it learned from one set of data—say, daily news reports about international conflicts—and apply that knowledge to a completely different set of data, like a graph of historical encyclopedia entries, without needing to be retrained. It does this by ignoring the specific vocabulary (the names of people, places, and exact dates) and focusing entirely on the structure: how events interact with each other and how they are ordered in time.
The researchers tested FITTER by training it on six different datasets that varied wildly in their "flavor." Some datasets tracked events every 15 minutes, while others only updated once a year. Some covered just one year of history, while others spanned nearly 200 years. Despite these massive differences, FITTER successfully transferred its knowledge from one domain to another. For instance, a model trained on the dense, fast-paced data of diplomatic crises could predict relationships in the slow-moving, long-term data of historical biographies.
The results were impressive. In 15 different cross-domain scenarios, FITTER consistently outperformed previous methods that tried to be "inductive" (flexible). It didn't just guess; it learned that certain patterns, like a specific type of event recurring after a set time, were universal. The paper shows that by using a special mathematical trick called "sinusoidal positional encoding"—which is like using a wave pattern to represent time instead of a clock number—the model could understand that "five days later" means the same thing whether the clock ticks every minute or every year.
However, the authors are careful to note that while FITTER is a giant leap for flexibility, it isn't a magic wand that solves everything. It still struggles a bit with extremely dense graphs where there are too many events happening at once, and it currently focuses on predicting missing people or things in a fact, rather than predicting the time or the relationship itself. But the core finding is clear: by stripping away the specific names and dates and focusing on the underlying rhythm of how the world changes, we can build AI that learns from one world and applies it to another, making our digital knowledge much more adaptable and useful.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.