Network Topology as an Empirical Precursor to Recursive Survival Weighting

Reading Position

Read this after the probability note and the exposure-gate demonstration. The probability note defines the survival-representation map; the beam-mask note shows a concrete support-gate protocol; this paper asks how empirical network topology might help choose variables before a future survival-loss model is fitted and tested.

Its best use is methodological. It gives RSG a conservative route into real survival-analysis datasets without pretending that a graph centrality score is already a recursive exposure law.

Claim Status

The paper is explicit that network survival analysis does not prove recursive survival theory. It claims only that graph topology can be a useful empirical preprocessing layer for selecting or weighting candidate exposure variables.

network topology:
observed variable structure and centrality rankings

recursive survival weighting:
generated histories, exposed loss, survival weights, represented fractions

The protective distinction is between topology and dynamics. A central node in a graph is not automatically causal, and a minimum spanning tree is not a generated recursive history. The bridge needs an independently declared hazard or exposed-loss law.

Network Survival Method

The source method, from Yusoff, Muhammad, and Liang, begins with a multivariate survival dataset. Variables are converted into a correlation matrix, then into a distance matrix, and then into a minimum spanning tree using Kruskal's algorithm.

clinical variables
  -> correlation matrix
  -> distance matrix
  -> minimum spanning tree
  -> degree and betweenness centrality

In the cervical-cancer case study, survival status is read against variables such as stage at diagnosis, histologic type, distant metastasis, age, lymph-node involvement, primary treatment, and ethnicity. The network analysis identifies stage at diagnosis and histologic type as especially important by centrality, with stage at diagnosis prominent as a bottleneck variable.

Translation Into RSG Terms

RSG separates generated histories from survival selection. A finite history family receives accumulated exposed loss, survival weight, and then a represented fraction:

A_i(t) = integral_0^t Gamma_i(tau) W_i(tau) d tau

S_i(t) = exp(-A_i(t))

p_i(t) = q_i S_i(t) / sum_j q_j S_j(t)

The network paper does not supply Gamma_i W_i. It supplies candidate variables that may later parameterise loss or exposure. The cautious vocabulary is:

network node      -> survival-relevant state feature
network edge      -> empirical coupling between features
centrality score  -> candidate leverage over survival outcome

The word "candidate" matters. A central variable could be a useful covariate, a proxy, a coding artefact, a causal influence, or some mixture. Recursive survival weighting must not treat centrality as exposure without validation.

Centrality-Weighted Hazard Bridge

The most defensible bridge runs through ordinary hazard modelling. Let C(v) be a predeclared centrality score for variable v, normalised before validation, and let x_i,v(t) be the variable value for patient or history i.

lambda_i(t) =
  lambda_0(t) * exp(sum_v beta_v * C_tilde(v) * x_i,v(t))

Only after this statistical model has been declared may the recursive-survival bridge be introduced:

Gamma_i(t) W_i(t) == lambda_i(t)      bridge postulate

That identification is not a theorem. It says that, for this clinical analogue, the hazard rate is being interpreted as an exposed survival-loss rate. The statistical model remains responsible for prediction.

Represented Fractions

Once the hazard bridge is declared, the represented fraction over a finite validation cohort can be written:

H_i(t) = integral_0^t lambda_i(tau) d tau

p_i(t) =
  q_i exp(-H_i(t)) / sum_j q_j exp(-H_j(t))

This represented fraction is not the same as the clinical survival probability of a population group. It is the normalised representation of patient-level or history-level survival weights inside the declared model.

p_i(t) / p_j(t) =
  (q_i / q_j) * exp(-H_i(t) + H_j(t))

Validation Discipline

The bridge becomes meaningful only when tested against ordinary survival models. A valid protocol must fix the variable coding, association estimator, distance rule, MST algorithm, centrality definition, centrality normalisation, hazard specification, train/validation split, and evaluation metrics before looking at the validation result.

baseline:
lambda_i^(0)(t) = lambda_0(t) * exp(sum_v beta_v x_i,v(t))

topology-weighted:
lambda_i(t) = lambda_0(t) * exp(sum_v beta_v C_tilde(v) x_i,v(t))

The bridge earns support only if the topology-weighted model improves held-out prediction under predeclared criteria such as concordance, Brier score, calibration, likelihood comparison, or another appropriate survival metric.

Limits

The source case study is static, correlation-based, and built from observed clinical variables. It depends on coding choices for qualitative categories, and its graph is a reduced empirical skeleton rather than a dynamical history.

• Topology is not dynamics. It says which variables are structurally related in the dataset.
• Hazard modelling is not causality by itself. Causal interpretation requires study design and validation.
• Centrality is not exposure by default. It can suggest candidate exposure variables only after a bridge model is declared.
• Recursive interpretation is conditional. The identification Gamma_i W_i == lambda_i is a modelling convention.

Fit With The Site

This paper gives the PDF library a practical empirical bridge. It sits beside the RSG probability and analogue-test notes because it asks how real-world survival datasets might provide predeclared candidate exposure variables before recursive survival weighting is applied.

In one sentence: network topology supplies candidate structure; recursive survival weighting supplies represented survival fractions only after the loss law has been fixed.