Differential Semantics of Intervention in Bayesian Networks

Biao QinIJCAI 2015

TL;DR

This paper bridges the gap between differentiation (calculus) and intervention (causality) in Bayesian networks. By introducing "Partial Intervention Tables" (PITs), the author provides a way to calculate causal effects without altering the network's structure, proving that an intervention is essentially the derivative of an observation—much like velocity is the derivative of displacement.

Introduction: The Challenge of Intervention

In causal Bayesian networks, an intervention (denoted by Pearl’s $do$-operator) represents an external agent forcing a variable to a specific value. Traditionally, calculating this requires "mutilating" the graph by deleting incoming edges to the intervened node or creating an "augmented" network with extra nodes.

Traditional intervention methods: (a) mutilated graph and (b) augmented network.

While effective, these structural changes are computationally expensive when testing every possible value of a variable (a "full atomic intervention"), as they often require running inference algorithms multiple times. This paper seeks a more unified framework that treats intervention as a property of the network's existing parameters.

Methods: PITs and the FullDo Algorithm

The paper introduces two major innovations to avoid structural modifications:

  1. Partial Intervention Tables (PIT): Instead of deleting edges, the author encodes intervention by replacing a node’s Conditional Probability Table (CPT) with a PIT. In a PIT, the probability of the chosen value is set to 1, and all others to 0, regardless of the parents. This allows the network structure to remain untouched.
  2. The FullDo Algorithm: A jointree-based algorithm that computes the effects of intervening on every possible value of a variable simultaneously. By caching intermediate factors in the jointree, the algorithm avoids redundant calculations.
  3. Differential Semantics: The paper mathematically proves that if a variable $V_i$ can reach the evidence nodes $e$, the causal effect $P(e | do(V_i))$ is equal to the sum of the partial derivatives of the network polynomial with respect to that variable's parameters.
A standard Bayesian network and its associated CPTs, which the PIT method aims to modify without structural changes.

Key Insights and Results

The most profound insight is that causal intervention has "differential semantics." Action (intervention) is to Prediction (observation) what Velocity is to Displacement. Just as velocity describes the rate of change of position, intervention describes the rate of change of probabilistic outcomes.

The proposed FullDo algorithm was evaluated using 16 standard benchmark Bayesian networks, including alarm, diabetes, and pathfinder. It consistently outperformed Pearl’s standard $do$-calculus approach. The speedup was particularly noticeable in networks with high cardinality; for the diabetes network, FullDo was roughly 5x faster than traditional methods.

Furthermore, the paper addresses a flaw in previous state-of-the-art algorithms regarding reachability. If an intervened node cannot "reach" the evidence nodes, standard differentiation fails. The author's framework identifies when to use the derivative and when to default to the prior probability.

Conclusion

This work successfully unifies the "differential approach" used for sensitivity analysis with the "causal approach" used for policy intervention. While currently focused on atomic interventions, the relationship between multiple simultaneous interventions and higher-order derivatives remains a promising avenue for future exploration.

Summary updated

BibTeX
@inproceedings{qin2015differential,
  author = {Biao Qin},
  title = {{Differential Semantics of Intervention in Bayesian Networks}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}

Loading related content...

Figure