How do feature-based approaches compare to aggregation methods for cold-start?

This explores how recommenders handle the cold-start gap, and the corpus frames it as a genuine fork: when you have no interaction history to aggregate, you either fall back on what you *know about* the user or item (features) or you find cleverer ways to keep aggregating. The honest answer the collection suggests is that the best systems refuse to choose — they fuse the two. GHRS is the clearest example: it builds graph features from rating structure *and* side information, then runs them through a deep autoencoder to discover non-linear relationships that a plain hybrid would miss Can autoencoders solve the cold-start problem in recommendations?. The lesson is that 'feature-based' and 'aggregation-based' aren't rivals so much as two inputs to a representation that works precisely because new users and items can be placed by their attributes before any behavior accumulates.

The purest feature-based answer to cold-start is to treat recommendation as a decision under uncertainty rather than a memory lookup. LinUCB does exactly this: it casts news recommendation as a contextual bandit, using article and user *features* to estimate value and explicitly balancing trying uncertain items against exploiting proven ones Can bandit algorithms beat collaborative filtering for news?. This beats collaborative filtering precisely where aggregation is weakest — fast-churning content where every item is effectively cold and there's no time to accumulate the co-occurrence statistics CF depends on. So the trade is legible: aggregation is powerful when history is dense and slow-moving; feature-driven exploration wins when history is thin or expires before it's useful.

There's a subtler question underneath, which is *how rich your features need to be*. TransRec argues that no single identifier facet is enough — pure IDs give you distinctiveness but no meaning, pure text gives you semantics but poor grounding — so it combines numeric IDs, titles, and attributes into one structured identifier Can item identifiers balance uniqueness and semantic meaning?. That's directly relevant to cold-start: a cold item has no behavioral ID signal worth aggregating, so the title and attribute facets are what carry it until interactions arrive. The feature side, in other words, is what buys you a graceful degradation path rather than a hard wall.

The aggregation family, meanwhile, has been quietly reinventing itself at a higher level — aggregating *models* instead of *ratings*. Avengers-Pro routes each query to a specialized model by semantic cluster and beats a single frontier model, suggesting that selection can be a stronger lever than building one bigger thing Can routing beat building one better model?. The interesting transfer to cold-start is conceptual: routing-by-cluster is itself a feature-based gate over an ensemble, which is to say the two paradigms collapse into each other once you look closely — you use features to *decide* which aggregation to trust.

So the takeaway you might not have gone looking for: the framing of 'features vs. aggregation' dissolves under the corpus's best work. Features are how you bootstrap and route; aggregation is how you exploit once signal exists; and the systems that win on cold-start — GHRS fusing both, LinUCB using features to manage exploration, TransRec hedging across identifier types — are the ones that treat the boundary as a dial rather than a wall.