A social graph with 28 billion links is a wonderful thing to own and an expensive thing to compute on. Researchers at AI VK Research say they have found a way to make that bill much smaller, and the paper describing it has been accepted to the international CIKM 2026 conference.
The problem with graphs
Graph neural networks work on objects and the connections between them. In a social network, the objects are people and the connections are friendships. The model reads who is linked to whom and uses that to suggest, say, other people you might know.
The catch is cost. Connected data at this scale demands a great deal of computing power, and the memory needed just to hold the graph during training is a large part of that bill. That is the part the VK team went after.
What the team reports
The researchers tested their approach on an anonymized graph of 194 million users and 28 billion connections. Two results stand out:
- Memory: their method of storing the data cut memory consumption by more than 100 times.
- Data preparation: getting the information ready for training became 2.5 times faster.
They also say the method takes into account how social connections change over time. That matters because a friendship graph is not a fixed map. People add and drop contacts constantly, and a model trained on a frozen snapshot is quietly training on the past.
From paper to the friends tab
The team then tried the technology in friend recommendations on VKontakte. The reported results:
| Metric | Change |
|---|---|
| Friend additions coming from recommendations | up 16% |
| Share of users who added at least one person from recommendations | up 11.5% |
The two figures measure different things, so they do not contradict each other. One counts additions overall, the other counts how many users added at least one person. It is plausible, though the report does not say, that some users simply added more people than before.
What we do not know
The report is brief, so a few things are missing. It does not explain how the storage method works. It does not name the baseline the 100x saving is measured against. It does not say how long the live test ran, how many users were in it, or how the comparison group was set up. The 16% and 11.5% figures are therefore the team's own headline numbers, and the full paper is where the details will be.
It is also worth keeping in mind that more friend additions is a proxy for better recommendations, not proof of it. A user who adds someone from a suggestion has at least found the suggestion plausible, which is a respectable signal, if not a final verdict.
Beyond friend suggestions
The researchers say the work could apply to other tasks that involve large volumes of interconnected data. They do not list any, so that remains a claim rather than a demonstration.
Still, the pitch is easy to follow. Graph models are only useful if you can afford to train them on the graph you actually have, rather than a politely shrunken sample of it. Cutting memory by two orders of magnitude is the kind of improvement that turns a research prototype into something a production team can justify running. Whether the full paper backs up the numbers will be clear when it appears at CIKM 2026.