How do different feed-weighting schemes construct distinct network topologies at population scale?
This explores how the scoring rules a recommendation feed uses to rank content (its 'weights') don't just change what individuals see — they reshape the connective structure of who-is-exposed-to-what across an entire population.
This explores how the scoring rules a recommendation feed uses to rank content don't just change individual experiences but actively wire up the population into different network shapes. The clearest statement of this in the corpus is the argument that recommendation feeds are persuasion infrastructure, not neutral pipes: feed weights influence what producers make, the resulting network topology pushes opinions to converge, and these effects compound at population scale through rating contamination and selection biases How do recommendation feeds shape what people see and believe?. The key move is recognizing that a weighting scheme is a topology-construction policy — change what the feed rewards and you change the graph of exposure.
The mechanism by which weights harden into topology is feedback. A ranker trained on its own past outputs converges on degenerate equilibria that amplify earlier decisions unless selection bias is explicitly modeled out — YouTube's system needs both a multi-gate objective mixer and a separate position tower precisely to break this self-reinforcing loop Why do ranking systems need to model selection bias explicitly?. Without that correction, the weighting scheme doesn't describe the network; it sculpts it over time, each round of exposure narrowing the next.
A subtler lever is the representational machinery underneath the weights. When user and item embeddings are too low-dimensional, the system overfits toward popular items to maximize ranking quality, and niche content gets starved of exposure — a structural skew that compounds over time and can't be patched after the fact Does embedding dimensionality secretly drive popularity bias in recommenders?. So 'feed-weighting' includes not just the visible objective weights but the dimensionality of the latent space, which silently decides whether the population graph collapses toward a popular core or stays diverse.
The corpus also shows that the *structure* of the signal you weight changes how stable the resulting topology is. Taobao's Swing algorithm builds product graphs from quasi-local bipartite patterns rather than single edges, and those structural signals resist noise because multiple independent noisy edges rarely align by chance Can graph structure patterns outperform direct edge signals in noisy data?. Weighting on structure versus weighting on raw edges yields measurably different — and differently robust — networks. There's even a hint that embedding geometry itself imposes a coarse-to-fine organization, where leading eigenvectors split broad categories before fine ones Do embedding eigenvectors organize taxonomy from coarse to fine?, suggesting the topology a feed produces is partly inherited from how its representation space is shaped before any explicit weight is tuned.
What you might not expect: the corpus frames the strongest topology lever as *selection* rather than *scale*. Routing queries to specialized models per semantic cluster outperforms one larger model Can routing beat building one better model? — and the same logic applies to feeds. How you partition and route a population matters more than how big the model is, which means feed-weighting schemes are best understood as routing topologies, where the choice of clustering quietly determines which communities ever see each other.
Sources 6 notes
Research shows recommendation systems operate as political actors: feed weights influence producer behavior, network topology drives opinion convergence, and automation enables targeted persuasion at population scale. These effects compound through rating contamination and selection biases.
YouTube's multi-objective ranker uses MMoE for conflicting objectives and a shallow position tower to remove selection bias from training data. Without both mechanisms, models converge on degenerate equilibria that amplify their own past decisions.
Research shows that when user/item embedding dimensions are too small, recommender systems overfit toward popular items to maximize ranking quality. This compounds over time as niche items receive insufficient exposure, and cannot be fixed post-hoc without treating dimensionality as a fairness hyperparameter.
Taobao's Swing algorithm constructs more robust product substitute graphs by exploiting quasi-local bipartite patterns rather than single edges. Structural signals are inherently noise-resistant because they require multiple independent noisy edges to coincidentally align, which rarely happens by chance.
Leading eigenvectors of embedding Gram matrices separate broad taxonomic branches first, then progressively finer sub-branches—a coarse-to-fine spectral order that tracks the WordNet hypernym tree level by level, confirming predictions from co-occurrence statistics.
Avengers-Pro achieves 7% higher accuracy than GPT-5-medium by routing queries to optimal models per semantic cluster, or matches its performance at 27% lower cost. Ten 7B models with routing previously surpassed GPT-4.1 and 4.5, suggesting selection is a stronger lever than scaling.