It is 2003, and social networks have become data: millions of nodes, timestamped edges, and a question with both scientific and commercial teeth — given the network today, which pairs not yet connected will be connected a year from now? Common neighbors, path structure, and community membership all whisper predictions; make them speak precisely. Define the link-prediction problem, compare principled scores against the right null — because on a sparse graph, predicting "no edge" everywhere is deceptively accurate — and say what the achievable accuracy reveals about how much of social life is structurally determined. Get it wrong and recommender engines wire the social graph by folklore; get it right and you must face the second question — what it means that the future of acquaintance is predictable at all.
Chose The causal diagram and do-calculus — right call.
Pearl's Bayesian networks and structural causal models show deep fluency in representing relationships between entities as a formal graph with quantifiable dependencies, conceptually adjacent to this problem's demand to represent social relationships as a graph and reason about which absent edges are likely to form. His causal framework also offers a genuinely useful caution this problem's second question implicitly raises, whether predicted future connections are structurally determined or merely correlated with current network position, a causal-versus-correlational distinction his do-calculus was built to sharpen. But his graphical models represent probabilistic and causal dependencies rather than social-network edges specifically, and he did not work on link prediction as an applied problem. His score reflects a real structural and causal-reasoning kinship, short of direct technical authorship.
Blei's topic models and his broader work making variational inference practical for large-scale latent-structure discovery are technically close kin to this problem's link-prediction task: both infer hidden structure, latent topics or latent community affinities, from sparse, noisy observed data and use that structure to predict what is not directly observed. Latent-factor approaches to link prediction, inferring node embeddings whose inner products predict edge probability, are a direct descendant of the probabilistic latent-variable machinery his research popularized. He did not work on social-network link prediction as a named applied problem himself. His score reflects strong transferable latent-variable-modeling technique directly applicable to this problem's mathematical structure, short of domain-specific authorship of the link-prediction literature itself.
Battle #168 · 8/10/2026, 11:41:32 AM · this result is deterministic: the same two personas on this problem always resolve the same way.