What makes substitute graphs fundamentally different from complement graphs in recommendation systems?
This explores why a graph of items that *replace* each other (alternatives a shopper picks between) behaves differently from a graph of items that *go together* (things bought as a set) — and what that difference means for how each graph is built and what it does to recommendations.
This explores why a graph of substitutes (items that compete for the same purchase — two phones, two pairs of running shoes) is a fundamentally different object than a graph of complements (items consumed together — a phone and its case). The short version: complements are revealed by what people buy *together*, while substitutes are revealed by what people *consider but don't buy together*. A shopper viewing five laptops and buying one is broadcasting a substitute signal across all five; a shopper buying a laptop and a sleeve is broadcasting a complement signal. The same browsing session feeds both graphs, but the relationships point in opposite directions, and that changes everything downstream.
The deepest consequence is how noisy each signal is. Complement relations sit on direct co-purchase edges — relatively clean, because money changed hands on both items. Substitute relations have no such anchor; they're inferred from softer behavioral overlap (co-viewing, co-considering), which is far noisier. Taobao's Swing algorithm exists precisely because of this: rather than trust any single co-view edge, it looks for *quasi-local bipartite structure* — patterns where multiple independent users repeat the same substitution. Structural patterns resist noise because several noisy edges rarely align by accident, whereas one edge can be a fluke Can graph structure patterns outperform direct edge signals in noisy data?. So substitute graphs almost demand structure-level construction, while complement graphs can lean more on direct edges.
The difference isn't just engineering — it reshapes user behavior and even opinion. Work on opinion convergence finds that frequently-bought-together networks (complements) and co-viewed networks (substitutes) drive ratings in measurably different directions, because each network type pulls in a different audience with different prior expectations. Recommending a complement says "complete your purchase"; recommending a substitute says "reconsider your choice" — and those two framings attract and satisfy different people, converging or diverging their opinions accordingly Do different recommender types shape opinion convergence differently?.
There's a quieter implication worth surfacing: substitute graphs are where *diversity* lives. A complement graph, followed blindly, narrows you toward one coherent bundle. The same intuition shows up in social recommendation, where friends with *different* tastes — not similar ones — generate the most valuable suggestions by pushing you toward choices outside your usual orbit Can friends with different tastes improve recommendations?. Substitution is the structural cousin of that idea: it maps the space of alternatives you *could* have chosen, which is exactly the raw material for exploration and diversity rather than confirmation.
One honest caveat: this corpus doesn't contain a paper that sits down and formally contrasts substitute-graph and complement-graph construction side by side. What it gives you instead is the substitution-specific machinery (Swing), the behavioral consequence (opinion convergence by network type), and the diversity angle (diverse-taste signals) — enough to see that the substitute/complement split is a real architectural fork, not a labeling choice.
Sources 3 notes
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.
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.
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.