How YouTube’s Recommendation Engine Evolved: From Graph Walks to Deep Neural Networks

First Stage: User‑Video Graph Walk (2008)

YouTube initially recommended videos similar to those a user had already watched by constructing a User‑Video co‑view graph and applying an Adsorption algorithm that propagates label information across the graph until convergence.

The algorithm iteratively updates each node’s label distribution based on the weighted contributions of its neighbors, resembling PageRank or a Markov‑chain walk.

Second Stage: Video‑Video Graph Walk (2010)

YouTube shifted to recommending videos similar to those a user had consumed, defining similarity through co‑view counts, session co‑occurrence, and ordered co‑views.

Videos watched by a certain number of common users.

Videos frequently watched together within the same session.

Videos co‑watched with order information considered.

Candidate videos are generated by aggregating similar‑video sets, optionally expanding to “similar‑of‑similar” videos, followed by ranking based on video quality metrics and user‑specific signals.

Third Stage: Search‑Based Methods & Collaborative Filtering (2014)

YouTube introduced a topic‑based representation for videos, extracting topic‑weight pairs from descriptions, keywords, search queries, and playlist names.

Similarity between two videos is computed using two formulas: one based on traditional IR weighting (including IDF‑like damping) and another that learns topic weights via a pairwise ranking loss.

Online A/B tests showed improvements in watch time, completion rate, and reduced drop‑off compared with pure collaborative filtering.

Fourth Stage: Deep Neural Networks (2016)

Facing massive scale, freshness, and noisy feedback, YouTube adopted deep learning for both candidate generation and ranking.

Videos and users are embedded into vectors (e.g., via Word2Vec or TensorFlow embeddings). A multi‑layer ReLU network with a Softmax output predicts the probability of a user watching a video in a given context.

The system combines offline candidate generation with online ranking that incorporates user context, video quality signals, and temporal features, achieving a dramatic improvement in recommendation relevance.

Summary

Across four research phases, YouTube’s recommendation pipeline progressed from simple graph‑based label propagation to sophisticated deep‑learning models, each step addressing scalability, freshness, and the need to reduce bias toward popular content.

Video Suggestion and Discovery for YouTube: Taking Random Walks Through the View Graph

The YouTube Video Recommendation System

Up Next: Retrieval Methods for Large‑Scale Related Video Suggestion

Deep Neural Networks for YouTube Recommendations