lawsynth/docs/guides/information-diffusion.md

2.8 KiB

Information diffusion analysis

lawsynth-information-diffusion is a native Rust application for bounded, reproducible analysis of measured network cascades. It estimates a global independent-cascade transmission probability from observed exposures, evaluates it on a chronological holdout when the data permits, and compares baseline and intervention forecasts under the same calibrated model.

The canonical implementation lives in apps/information-diffusion. Hackathon directories contain submission material only.

Data contract

The application reads six normalized UTF-8 TSV relations. Five are required:

  • nodes: node_id;
  • directed edges: source, target;
  • cascades: cascade_id, canonical UTC started_at, and observation_end_step;
  • activations: cascade_id, node_id, and discrete step; and
  • forecast seeds: node_id.

An optional node_id relation defines blocked nodes for the intervention. Headers and column counts are exact. Identifiers, references, uniqueness, observation windows, timestamps, and configured resource bounds are validated before analysis. Inputs larger than 64 MiB per file are rejected at the CLI boundary.

Analysis and receipts

For each inactive node-step with at least one newly active parent, calibration records the one-time incoming edge exposure count and whether the node activates at the next step. This matches the synchronous independent-cascade forecast, where an edge gets one attempt when its source activates. A bounded maximum-likelihood fit estimates the global edge probability. The report keeps its confidence interval, negative log likelihood, and Brier score separate from the seeded Monte Carlo forecast bands.

Cascades are sorted by start time and identifier. With at least three cascades, the first 80 percent calibrate a holdout model and the remaining cascades measure out-of-sample Brier score and log loss. If either partition has no measurable edge exposures, the backtest is unmeasured with an explicit reason.

The baseline and blocked-node/transmission-multiplier intervention use identical seed and simulation settings. Every report includes schema and model versions, normalized data digest, model inputs, deterministic simulation seed, calibrated metrics, forecasts, limitations, and a SHA-256 receipt digest. Output is synced and installed atomically; replacement requires explicit --overwrite.

See apps/information-diffusion/README.md for exact headers and CLI flags.

Interpretation boundary

An edge means activation timing is compatible with a candidate path. It does not establish causal influence; shared causes and missing nodes can produce the same pattern. Absence is interpreted only inside the declared observation window. Process bands do not include parameter uncertainty, which remains a separate calibration interval.