Explainability

Every intermediate quantity in a BRB inference has semantic meaning: which rules fired, how strongly, and how their belief distributions were combined. desdeo-brb exposes these with first-class helpers.

For an end-to-end walkthrough with plots, see notebooks/04_explainability.ipynb in the repository.

Describe a single rule

print(model.rule_base.describe_rule(3, attribute_names=["x"], consequent_name="f(x)"))
# Rule 3: IF x is 1.5 THEN f(x) = {1: 0.833, 2: 0.167} [w=0.143]

Options:

• attribute_names — list of human-readable names for the attributes. Defaults to x1, x2, ....
• consequent_name — name for the consequent variable. Defaults to no name.
• show_zero_beliefs — if False (default), consequent values with belief degree < 0.001 are hidden for readability.

Describe all rules

print(model.rule_base.describe_all_rules(
    attribute_names=["FlowDiff", "PressureDiff"],
    consequent_name="LeakSize",
))

Explain a single prediction

model.explain(X) returns a structured explanation:

print(model.explain(
    np.array([[1.5]]),
    top_k=3,
    attribute_names=["x"],
    consequent_name="f(x)",
))

Output:

Prediction: 1.167

Top activated rules:
  Rule 3 (w=1.0000, x=1.5): {1: 0.833, 2: 0.167}

Combined belief distribution:
  {1: 0.833, 2: 0.167}

For multiple samples, use sample_idx to select which one to explain:

X = np.array([[1.5], [2.5]])
print(model.explain(X, sample_idx=1))

Accessing raw inference data

predict() returns an InferenceResult with the full trace as NumPy arrays:

result = model.predict(X)

result.output                    # (n_samples,) scalar predictions
result.activation_weights        # (n_samples, n_rules)
result.combined_belief_degrees   # (n_samples, n_consequents)
result.input_belief_distributions  # list of (n_samples, n_rv_i) arrays

Use result.dominant_rules(top_k=3) for the indices of the most-activated rules per sample, and result.explain(sample_idx, rule_base=model.rule_base) to format the explanation for a given sample.

Why it matters

In decision-support systems, explainability is not a luxury. The INFRINGER method ¹ uses BRB systems to learn a decision maker's preferences during interactive multi-objective optimisation; each preference update must be explainable to the decision maker for them to trust the system.

---

1. Giovanni Misitano. Interactively learning the preferences of a decision maker in multi-objective optimization utilizing belief-rules. In 2020 IEEE Symposium Series on Computational Intelligence (SSCI), 133–140. 2020. ↩