Skip to content
Needs Verification

Explanatory Breadth (Consilience & Evidence Coverage)

Definition & Conceptual Goal

The Explanatory Breadth (Consilience & Evidence Coverage) metric favors hypotheses and theories that explain a broader spectrum and larger volume of distinct empirical evidence (Thagard, 1989, pp. 437, 442; Schurz, 2024, sec. 5.3).

Originally emphasized by William Whewell and formalized computationally in Thagard's Theory of Explanatory Coherence (TEC), consilience operates on the principle that a hypothesis gains epistemic credibility in proportion to the diversity and quantity of independent empirical facts it successfully accounts for.

Conversely, this metric directly incorporates the Unexplained Evidence Ratio (UER): penalizing theoretical networks that achieve artificial internal harmony merely by cherry-picking a tiny subset of data while leaving the majority of established empirical observations unexplained.

Theoretical Grounding & Model Formulation

Let \((G = (V_H \cup V_E, R_{\text{sup}} \cup \mathcal{R}_{\text{att}}))\) be a directed knowledge/coherence graph, where:

  • \((V_H)\) is the set of hypothesis nodes representing theoretical claims.
  • \((V_E)\) is the set of established evidence nodes within the empirical domain scope.
  • \((R_{\text{sup}} \subseteq V_E \times V_H)\) is the set of directed support relations \(E \to H)\) indicating evidence \((E)\) supports or is explained by \((H\).
  • \((\mathcal{R}_{\text{att}} \subseteq V_E \times V_H)\) is the set of directed attack/refutation relations \(E \to H)\) indicating evidence \((E)\) directly contradicts or refutes \((H\).

While Thagard's ECHO connectionist relaxation model required symmetric edge weights \(w_{ij} = w_{ji}\) for Lyapunov energy convergence, structural graph metrics operate on directed edges to accurately model explanatory asymmetry without numeric oscillation.

Mathematical Specification

Consilience Out-Set \(Cons\)

For an individual hypothesis \((H \in V_H)\), consilience is defined set-theoretically over incoming support edges:

\[Cons (H) = |\{ E \in V_E \mid (E, H) \in R_{\text{sup}} \}|\]

Relative Node Coverage \(REB_{\text{NODE}}\)

For a theory \((T)\) composed of hypothesis nodes \((V_H (T) \subseteq V_H)\), relative node coverage is computed over the union of explained evidence to prevent double-counting shared facts:

\[REB_{\text{NODE}} (T) = \frac{\left| \bigcup_{H \in V_H (T)} \{ E \in V_E \mid (E, H) \in R_{\text{sup}} \} \right|}{|V_E|} = \frac{|V_E^{\text{exp}} (T)|}{|V_E|}\]

Domain-Clustered Explanatory Breadth \(REB_{\text{CLUSTER}}\)

To capture Whewellian qualitative consilience across distinct empirical fields:

  1. Let \((\mathcal{C} = \{C_1, C_2, \dots, C_k\})\) be the partition of evidence nodes \((V_E)\) into topological/topical communities identified by a clustering algorithm (e.g., Leiden modularity optimization).
  2. Domain coverage ratio \((REB_{\text{CLUSTER}} (T))\) evaluates the proportion of empirical clusters containing at least one evidence node explained by theory \((T)\):
\[REB_{\text{CLUSTER}} (T) = \frac{|\{ C_i \in \mathcal{C} \mid C_i \cap V_E^{\text{exp}} (T) \neq \emptyset \}|}{|\mathcal{C}|}\]

Note: In evaluation pipelines, \((REB_{\text{NODE}} (T))\) and \((REB_{\text{CLUSTER}} (T))\) can be combined via geometric mean or presented as complementary metrics.

Unexplained Evidence Ratio \(UER\) & Empirical Conflict Ratio \(ECR\)

We separate passive evidence omission from active empirical falsification:

  1. Unexplained Evidence Ratio (Passive Gap): $$ UER (T) = \frac{|V_E \setminus V_E^{\text{exp}} (T)|}{|V_E|} = 1 - REB_{\text{NODE}} (T) $$

  2. Empirical Conflict Ratio (Active Refutation): $$ ECR (T) = \frac{\left| \bigcup_{H \in V_H (T)} { E \in V_E \mid (E, H) \in \mathcal{R}_{\text{att}} } \right|}{|V_E|} $$

  3. Net Explanatory Coverage \(REB_{\text{net}}\): $$ REB_{\text{net}} (T) = REB_{\text{NODE}} (T) - \alpha \cdot ECR (T) $$ where \((\alpha \ge 1.0)\) is the falsification severity multiplier.

Dynamic Skepticism Penalty

To penalize theories that cherry-pick data or ignore active refutations, the systemic decay/skepticism rate \((\theta_ {\text{effective}})\) is scaled by both omission and conflict ratios:

\[\theta_{\text{effective}} = \theta_{\text{base}} \cdot \left (1 + \gamma_{\text{unexp}} \cdot UER (T) + \gamma_{\text{att}} \cdot ECR (T) \right)\]

where:

  • \((\theta_{\text{base}})\) is the baseline network decay rate.
  • \((\gamma_{\text{unexp}} \ge 1.0)\) is the evidence omission penalty weight (default = \((1.0\).
  • \((\gamma_{\text{att}} \gg \gamma_{\text{unexp}})\) is the active refutation penalty weight (default = \((3.0\).

Measurement & Graph Implementation

Diagnostic & Metascientific Value

\(REB(T)\) Score \(UER(T)\) Score Epistemic Status Metascientific Interpretation
\(REB > 0.6\) \(UER < 0.4\) Broadly Consilient Strong explanatory coverage; accounts for diverse phenomena across scope.
\(0.2 \le REB \le 0.6\) \(0.4 \le UER \le 0.8\) Partial Scope Moderate coverage with notable unaddressed observational pockets.
\(REB < 0.1\) \(UER > 0.9\) Isolated / Cherry-Picking Explains only isolated observations; high decay penalty applied due to omitted data.

Grounding References

  • [Balzer et al., 1987] Balzer, W., Moulines, C. U., & Sneed, J. D. An Architectonic for Science. Reidel, Ch. VIII (Directed intertheoretical links and theory-holons).
  • [Newman, 2018] Newman, M. Networks. Oxford University Press, Ch. 14 (Community structure and modularity algorithms).
  • [Nováček et al., 2015] Nováček, V. et al. Formalising Hypothesis Virtues for Knowledge Graphs. arXiv: 1503.09137v2 (Metric constellations and modesty/simplicity measures).
  • [Schurz, 2024] Schurz, G. Philosophy of Science: A Unified Approach. Routledge, sec. 3.12, 5.3 & 6.4 (Content elements, tacking paradoxes, and unification).
  • [Thagard, 1989] Thagard, P. Explanatory Coherence. Behavioral and Brain Sciences, 12 (3), pp. 435–467.