Tubes and bubbles topological confinement of YouTube recommendations
Fig 3
Illustration of the recursive crawl focused on a given seed video.
Recommendations are crawled for the seed until a plateau may be estimated, which defines the direct neighbors of the seed and a set of nodes at depth 1. This process is repeated for all nodes at depth 1 in parallel, thus defining depth 2, and again with nodes at depth 2. In the end, the recommendation graph induced by the seed contains nodes at depths 1 and 2 and potentially includes links towards already explored nodes, i.e. at depth 0 (seed), 1 (seed’s direct neighbors) and 2 (seed’s indirect neighbors). There are on average 23.6 nodes at depth 1 (σ = 5.15), 325 nodes at depth 2 (σ = 98.23) and 2830 nodes at depth 3 (σ = 1160). Some elements are shaded simply to indicate that we do not represent all nodes and links on this figure for the sake of clarity.