AI VK Open-Sources HEGM: A Model That Predicts the Full Distribution of Video Watch Time

If you build recommendations for short videos, you have almost certainly predicted watch time. The classic setup is simple: take user and clip features, train a regression, and get an "expected 12 seconds." The problem is that in reality almost nobody watches exactly 12 seconds. The AI VK team has published a paper that tackles this problem in practice.

Why Average Watch Time Lies

Imagine two clips, each 30 seconds long. For the first, half of the audience swipes past within a second, while the other half watches to the end. For the second, almost everyone watches for around 15 seconds. Their average estimates may turn out identical, yet these are completely different clips with different fates in the feed.

The watch-time distribution for short videos is awkward. First there is a sharp spike of quick skips in the opening seconds, then a long tail, and for some clips there is also a peak at full completion. Regression with squared error pulls the prediction toward the middle, which is exactly where there are few real views. A ranking built on such a number poorly distinguishes "it landed" from "swiped away."

What HEGM Proposes

HEGM stands for Hierarchical Exponential-Gaussian Mixture. From the name and the authors' description, the general idea is clear: the model predicts not a single point but an entire distribution, as a mixture of components of two types. The exponential component suits quick skips well: the probability drops sharply with every second. The Gaussian describes "normal" viewing around some characteristic time. The word "hierarchical" indicates the components are organized in multiple levels, but the details of this structure are not disclosed in the announcement and must be sought in the paper.

The authors also mention problems that arise when training probabilistic models. A familiar example: mixtures of distributions tend to "collapse," where one component takes everything and the others are left dead. How exactly the team solved this will have to be read in the preprint.

What to Do With This in Practice

The first step does not require HEGM. Build a histogram of actual watch time across your clips and see whether it has several humps. If it does, a single point regression is almost certainly hurting you.

Second: from the distribution you can derive not one number but several different signals. These are the probability of a skip in the first seconds, the probability of a full view, and the expected time given that the person did not leave immediately. For a product manager, such features are easier to understand and easier to tie to product goals than an abstract "expected time."

Third: the HEGM code and models are on GitHub, so you can test the hypothesis on your own logs without building from scratch. The research was accepted to the main track of IEEE ICDM 2026, which means it has passed peer review.

Limitations

It is important to understand that the announcement contains no concrete figures. We do not know by what percentage HEGM beat the baselines or what exactly changed in the A/B test on VK traffic. The authors only write that a test was conducted and that there were changes. This is a claim of a strong result, not proof of it, and the numbers should be sought in the paper.

Moreover, the model was tested on short videos, and it is unknown how well the approach transfers to long videos, podcasts, or streaming, where the shape of the distribution is different. A mixture of distributions is also harder to train and debug than ordinary regression, and requires more computation.

The bottom line: the idea of predicting a distribution rather than the mean is entirely reasonable and applicable beyond VK. Whether you need HEGM specifically is something to decide after reading the paper and trial-running the code on your own data.