Abstract:Lifting uses a representative of indistinguishable individuals to exploit symmetries in probabilistic relational models, denoted as parametric factor graphs, to speed up inference while maintaining exact answers. In this paper, we show how lifting can be applied to causal inference in partially directed graphs, i.e., graphs that contain both directed and undirected edges to represent causal relationships between random variables. We present partially directed parametric causal factor graphs (PPCFGs) as a generalisation of previously introduced parametric causal factor graphs, which require a fully directed graph. We further show how causal inference can be performed on a lifted level in PPCFGs, thereby extending the applicability of lifted causal inference to a broader range of models requiring less prior knowledge about causal relationships.
What problem does this paper attempt to address?
The core problem that this paper attempts to solve is: how to efficiently estimate causal effects in partially directed graphs (i.e., graphs containing both directed and undirected edges). Specifically, the authors introduced Partially Directed Parametric Causal Factor Graphs (PPCFGs) to extend the previously proposed Parametric Causal Factor Graphs (PCFGs). PCFGs require a fully directed graph to represent causal relationships, while PPCFGs allow the co - existence of directed and undirected edges, thereby reducing the need for prior knowledge of causal relationships and improving the applicability of the model.
### Main problems
1. **Reducing the need for prior knowledge**: In practical applications, we usually cannot know all causal relationships exactly, so a method is needed to perform causal inference when only some causal relationships are known.
2. **Improving inference efficiency**: By taking advantage of the symmetry between objects (i.e., indistinguishable objects), the inference process can be accelerated while enhancing the expressive power.
3. **Handling intervention queries**: How to accurately calculate the effects of intervention operations in partially directed graphs, especially in the presence of undirected edges.
### Solutions
To address the above challenges, the authors proposed the following solutions:
- **Introducing PPCFGs**: PPCFGs, as a new graph model, can represent both directed and undirected edges, thus representing causal relationships more flexibly.
- **Defining d - separation**: Define the d - separation criterion in PPCFGs for judging conditional independence, which is the basis of causal inference.
- **Developing an efficient causal inference algorithm**: Proposed an algorithm based on lifted inference, which can efficiently estimate causal effects without fully expanding the graph. This algorithm processes intervention queries by enumerating all possible sets of parent nodes to ensure the accuracy of the results.
### Formula representation
Some of the key formulas involved in the paper are as follows:
1. **Full joint distribution**:
\[
P_M=\frac{1}{Z}\prod_{g\in G}\prod_{\phi_k\in gr(g)}\phi_k(A_k)
\]
where \(Z\) is the normalization constant, and \(A_k\) are the parameterized random variables (PRVs) that appear in the parameter list of the factor function \(\phi_k\).
2. **Probability distribution after intervention**:
\[
P(R_1 = r_1,\ldots,R_n = r_n|do(R'_1 = r'_1,\ldots,R'_k = r'_k))=
\begin{cases}
\prod_{R_i\in\{R_1,\ldots,R_n\}\setminus\{R'_1,\ldots,R'_k\}}P(r_i|pa(R_i,gr(M))) & \text{if }\forall j\in\{1,\ldots,k\}:r_j = r'_j\\
0 & \text{otherwise}
\end{cases}
\]
### Summary
The main contribution of this paper is to provide a new graph model (PPCFGs) and its corresponding causal inference algorithm, enabling efficient and accurate causal effect estimation even when only some causal relationships are known. This is of great significance for many practical application scenarios, especially when causal relationships are not completely clear.