Influence of Backdoor Paths on Causal Link Prediction

TL;DR

CausalLPBack is proposed, a novel approach to causal link prediction that eliminates backdoor paths and uses knowledge graph link prediction methods that extends the representation of causality in a neuro-symbolic framework, enabling the adoption and use of traditional causal AI concepts and methods.

Abstract

The current method for predicting causal links in knowledge graphs uses weighted causal relations. For a given link between cause-effect entities, the presence of a confounder affects the causal link prediction, which can lead to spurious and inaccurate results. We aim to block these confounders using backdoor path adjustment. Backdoor paths are non-causal association flows that connect the \textit{cause-entity} to the \textit{effect-entity} through other variables. Removing these paths ensures a more accurate prediction of causal links. This paper proposes CausalLPBack, a novel approach to causal link prediction that eliminates backdoor paths and uses knowledge graph link prediction methods. It extends the representation of causality in a neuro-symbolic framework, enabling the adoption and use of traditional causal AI concepts and methods. We demonstrate our approach using a causal reasoning benchmark dataset of simulated videos. The evaluation involves a unique dataset splitting method called the Markov-based split that's relevant for causal link prediction. The evaluation of the proposed approach demonstrates atleast 30\% in MRR and 16\% in Hits@K inflated performance for causal link prediction that is due to the bias introduced by backdoor paths for both baseline and weighted causal relations.

Figures (8)

• Figure 1: A causal network where the nodes represent the event and the edges between the nodes represent the causal relations. To estimate the true causal influence of node A on H (marked in orange) we need to block the influence of the common cause node, G which is affecting both A and H. Hence the edges on the path from A to H going via G is removed from the network
• Figure 2: CausalLPBack has five primary phases: 1) encoding the causal associations from data as a causal network, 2) removing the backdoor paths spanning across the nodes in the training and testing set (3) translating the causal network into a CausalKG, compatible with the causal ontology, (4) learning KGE for the CausalKG-W and CausalKG-Base, and (5) using the KG embeddings for predicting new causal links.
• Figure 3: Reified causal relation, causesType and causedByType. The causesType is a reified relation from a cause-entity instance to the type of an effect-entity. The causedByType is a reified relation from an effect-entity instance to the type of a cause-entity.
• Figure 4: Different CLEVRER-Humans CausalKG structures used for evaluating causal explanation and prediction tasks: (a) subgraph C which consists of links with only causal relations, i.e. causes, causedBy, causesType, and causedByType, (b) subgraph CT with causal relations and information about entity types, i.e. rdf:type, (c) subgraph CTP with causal relations, entity type relations, and information about the objects that participate in the causal events.
• Figure 5: The MRR scores KGE models with backdoor (w/ Backdoor), maximum backdoor elimination (w/o Backdoor (max)), and sufficient backdoor elimination (w/o Backdoor (Suff)) for (a) causal explanation with weights, (b) causal explanation without weights