Can social graph structure and behavioral co-occurrence both improve recommendation accuracy?
This explores whether two different signals for recommendations — who you're connected to (social graph) and what items tend to be acted on together (behavioral co-occurrence) — each genuinely improve accuracy, and what makes them work.
This explores whether two different signals for recommendations — who you're connected to (social graph) and what items tend to be acted on together (behavioral co-occurrence) — each genuinely improve accuracy, and what makes them work. The corpus says yes to both, but it complicates the obvious intuition about *why* in interesting ways.
On the social side, the surprising finding is that friends help you least when they're like you. Social Poisson Factorization shows that the value of a social network comes not from taste similarity but from influence — friends with *different* preferences pull you toward items outside your usual zone, and that outperforms methods that assume your friends should look like you Can friends with different tastes improve recommendations?. So the social graph improves accuracy, but through diversity and influence on anomalous choices, not the homophily you'd expect.
On the behavioral co-occurrence side, the strongest lesson is that *structure* beats raw edges. Taobao's Swing algorithm builds substitute-product relations from quasi-local patterns in the bipartite user-item graph rather than trusting any single co-purchase signal — and it's more robust precisely because a structural pattern requires multiple independent noisy edges to line up by chance, which rarely happens Can graph structure patterns outperform direct edge signals in noisy data?. Co-occurrence works, but you get accuracy from the graph's shape, not from counting how often two things appear together.
The most interesting answer to your question is that you don't have to choose. Knowledge graph attention networks fold user-item interactions and item-side relationships into a single Collaborative Knowledge Graph, then use attention-based propagation to pull both user-similarity and attribute-similarity signals — including high-order connections that ripple several hops out — from one unified structure Can graphs unify collaborative filtering and side information?. This is the corpus's clearest statement that social and behavioral signals are the same kind of object (a graph) and can be combined rather than bolted together.
Two cautions worth knowing before you trust either signal. Co-occurrence networks aren't neutral mirrors of preference — "frequently bought together" and "co-viewed" networks actually shape how people rate connected products, because each recommender type draws a different audience with different expectations Do different recommender types shape opinion convergence differently?. And the broader recommender literature warns that accuracy gains come more from problem-specific design — constraints, the right likelihood, inductive bias — than from throwing graph capacity at the problem What architectural choices actually improve recommender system performance?. The graph helps; how you constrain it helps more.
Sources 5 notes
Social Poisson Factorization uses friends' diverse tastes to recommend items outside users' usual preferences, outperforming methods that pull friends' representations together. Networks add value through influence on anomalous choices, not taste similarity.
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.
KGAT merges user-item interaction graphs with item knowledge graphs into a Collaborative Knowledge Graph, using attention-based propagation to capture both user-similarity and attribute-similarity signals simultaneously—including high-order connections that standard supervised learning methods miss.
Research shows that frequently-bought-together and co-viewed recommendation networks produce different opinion convergence patterns. The mechanism: each recommender type attracts different audience segments with different prior expectations, shaping both who sees products together and how they rate them.
Research shows that architectural choices like removing hidden layers, enforcing constraints on self-similarity, and using appropriate likelihood functions deliver better results than deeper or more complex models. This suggests that problem-specific design decisions matter more than raw representational capacity.