Recommendation systems have traditionally been explained through individual signals: a user clicks, likes, watches or comments, and the platform updates its estimate of what that person wants. VK’s new “neural profile” approach points toward a more informative representation — one built from sequences rather than isolated events.
The distinction is important in machine learning. A single observation is often ambiguous. A video view can be accidental. A like can be ironic. A comment can express disagreement rather than preference. Temporal sequences add context by showing whether behaviour repeats, changes direction or clusters around related topics.
Materials describing the new VK model say it considers activity across VKontakte, VK Video and VK Clips and tries to separate persistent interests from temporary ones. That is consistent with VK’s broader Discovery architecture. The company’s shared Discovery platform combines recommendation and search technologies, while Discovery AI, announced in July 2026, is designed to use interests from multiple services.
VK has also reported that its multimodal systems can predict user reactions and form a “viewer portrait” from interactions with content. Together, these components indicate a shift from simple event logging toward higher-level behavioural representation.
The rollout materials cite a 15% increase in content sharing, a 10% increase in full-screen opens and a 5% increase in likes. These figures are attributed rollout claims. Without a published experimental design alongside them — sample size, control groups, exposure windows and statistical uncertainty — they should not be interpreted as a peer-reviewed causal estimate.
The underlying scientific idea, however, is well supported: digital behaviour can contain signals about stable human characteristics. A 2024 study in Scientific Reports analysed data from 1,358 Russian-speaking VK users who had provided informed consent. Researchers used social-network information to model personality traits and cognitive abilities and then applied explainable-AI methods to examine what contributed to the predictions.
That study is not evidence about VK’s commercial recommendation system, and its ethical design is materially different from large-scale platform profiling because participants consented to the research. It is nevertheless relevant because it demonstrates the inference capacity hidden inside ordinary social-media traces.
The risk of inference grows when multiple kinds of behaviour are linked over time. Views can reflect attention. Comments can reflect engagement or conflict. Shares can reflect identity signaling. Repeated patterns across different products can provide a stronger estimate of preference than any one action.
In Russia, the significance of those models is amplified by the surrounding surveillance environment. Article 10.1 of the country’s information law requires certain internet-service operators to retain categories of user and communications data. From January 2026, specified metadata and user-action information covered by the rules must be stored for three years, while message content may be retained for up to six months under separate requirements.
Russia also operates SORM, a technical interception system. A 2026 EU regulation describes SORM as monitoring internet and social-network activity and providing the FSB with access to copied traffic; it also references location and behavioural patterns.
There is no public evidence in the cited material that VK exports its finished neural profile directly into SORM or to the FSB. That should not be inferred from the existence of data-retention laws alone. A proprietary machine-learning representation and the underlying event data are different technical objects.
But from a scientific and policy perspective, the boundary matters less than it might appear. If enough underlying events are retained or accessible, many higher-level inferences can be reconstructed without possessing the exact commercial model. The power lies in the density, duration and linkability of the data.
VK’s new system is therefore a useful case study in modern behavioural computing. Sequence models can make recommendations better because they reconstruct context. The same capability can make digital histories more revealing because context is precisely what turns scattered clicks into a model of a person.
