Skip to content
TheChatArchiveThe Chat Archive
o3-2025-04-16· OpenAIHugging Face LMArena Human Preference 140K

When multiplying an adjacency matrix of a bipartite network by its transpose in order to…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Coherecommand-a-03-2025’s answer.
  1. User

    When multiplying an adjacency matrix of a bipartite network by its transpose in order to obtain the projection network, is there any standard way to normalize?

  2. Assistant

    Short answer
    There is no single “standard” normalisation that everybody agrees on, but there are a handful of very common ones. Which one you pick should be driven by what you want the projected edges to mean ( “number of shared neighbours”, “similarity”, “statistically significant overlap”, …).

    Below is a quick → full overview.

    ─────────────────────────────────

    1. The raw projection
      ─────────────────────────────────
      Let

    • U = {u₁ … u_m}, V = {v₁ … v_n}
    • B (m × n)  = bipartite adjacency (B_{ik}=1 if u_i is linked to v_k)

    Projection on U (rows)
    P = B Bᵀ    P_{ij} = Σ_k B_{ik} B_{jk} = “# common V-neighbours”

    Good if you only care about counts, but it is heavily biased by the degrees of u_i and u_j as well as by the degrees of the v_k that they meet at.

    ─────────────────────────────────
    2. Degree-based normalisations
    ─────────────────────────────────
    Below k_i = row degree of u_i (k_i = Σ_k B_{ik}), ℓ_k = column degree of v_k.

    A) Cosine (a row/column ½-normalisation)
    B̃ = D_U^{-½} B D_V^{-½}
    P_cos = B̃ B̃ᵀ  P_{ij} = (# common)/√(k_i k_j)

    • Range: [0,1] if B is binary.
    • Intuition: standard “cosine similarity” between the two row vectors.

    B) Jaccard
    P_jac_{ij} = (# common) / (k_i + k_j − # common)

    • Penalises high-degree nodes even more.
    • Good when you need a similarity that is ≤1 and treats full co-incidence as 1.

    C) Overlap coefficient (= containment)
    (# common)/min(k_i, k_j)

    • Useful if you want to know whether the smaller of the two sets is “included” in the other.

    D) Inverse-degree weight of the middle node (Newman 2001)
    P_invdeg_{ij} = Σ_k 1/(ℓ_k − 1)

    • Each shared v_k contributes less if it is a hub.
    • Very popular for collaboration networks (co-authorship, actor–movie, etc.).

    E) Leicht–Holme–Newman / configuration-model normalisation
    Expected overlap under a random‐bipartite null model:
    E[A_{ik}A_{jk}] = k_i k_j ℓ_k / (2m)²
    Then use a z-score or a ratio
    z_{ij} = (P_{ij} − E)/√Var

    • Gives you statistical significance rather than similarity per se.

    ─────────────────────────────────
    3. A general template (covers many of the above)
    ─────────────────────────────────
    Choose exponents α, β ∈ [0,1] and compute

    B* = D_U^{-α} B D_V^{-β}  
    P* = B* B*ᵀ
    

    α = β = ½ → cosine
    α = 1, β = 0 → divide by row degree only, etc.

    ─────────────────────────────────
    4. When do you need to normalise?
    ─────────────────────────────────
    • If you care about pure co-occurrence counts (e.g. “two drugs share 15 targets”), keep P = B Bᵀ.
    • If you want a similarity measure, use cosine, Jaccard, overlap, etc.
    • If hub nodes (either side) create spurious connections, weight by 1/ℓ_k or use a significance filter.
    • If you need a statistically significant backbone, do a null-model test or disparity filter after the projection.

    ─────────────────────────────────
    5. References for deeper dives
    ─────────────────────────────────
    • Newman, “Scientific collaboration networks: II. Shortest paths, weighted networks and centrality.” Phys. Rev. E (2001).
    • Guillaume & Latapy, “Bipartite structure of all complex networks.” Inf. Proc. Lett. (2004).
    • Latapy, Magnien & Del Vecchio, “Basic notions for the analysis of large two-mode networks.” Social Networks (2008).
    • Neal, “Efficient simulation of bipartite configuration models.” Phys. Rev. E (2020) – for null-model expectations.

    ─────────────────────────────────
    Take-home message
    There is no single canonical normalisation; instead pick (or design) the one whose semantics match your research question—cosine for similarity, inverse-degree weighting for hub correction, Jaccard for overlap proportion, or a null-model based z-score for statistical significance.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.