The VK engineering team has moved away from classical gradient boosting in its recommendation systems, switching to the DCDN neural network architecture based on Deep & Cross Network. The fundamental technical shift is that the model independently computes CTR and derived metrics directly from raw event counters. Manual feature engineering—a process that traditionally consumes months of data science resources—is now largely obsolete for this workflow.
Instead of collapsing complex user behavior into a single monolithic score, the new ranker decomposes its predictions. The architecture simultaneously estimates the probabilities of likes and dislikes while modeling watch time distributions, clearly separating instant skips from deep, engaged viewing sessions. These signals are subsequently combined via Learning to Rank to produce the final feed sorting.
Deployment results in VK Clips highlight the clear business payoff of abandoning monolithic scoring: total watch time increased by 5.5%, likes rose by 5%, and reposts surged by 15%. For digital product leaders, this is a clear takeaway: end-to-end probabilistic neural ranking delivers ROI by slashing expensive manual engineering while directly boosting core engagement and retention metrics.