A Causal Framework for Decomposing Spurious Variations
Authors: Drago Plecko, Elias Bareinboim
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The described approach has several applications, ranging from explainable and fair AI to questions in epidemiology and medicine, and we empirically demonstrate its use. |
| Researcher Affiliation | Academia | Drago Plecko and Elias Bareinboim Department of Computer Science Columbia University dp3144@columbia.edu, eb@cs.columbia.edu |
| Pseudocode | Yes | Algorithm 1 (Abduction, Action and Prediction [9]). Given an SCM F, P(u) , the conditional probability P(YC | E = e) of a counterfactual sentence if it were C then Y ", upon observing the evidence E = e, can be evaluated using the following three steps: (i) Abduction update P(u) by the evidence e to obtain P(u | e), (ii) Action modify F by the action do(C), where C is an antecedent of Y , to obtain FC, (iii) Prediction use the model FC, P(u | e) to compute the probability of YC. ... Algorithm 2 (Partial Abduction and Prediction). |
| Open Source Code | Yes | The source code for reproducing the experiments can be found in our code repository. |
| Open Datasets | Yes | We now apply our framework to a synthetic example (called Synthetic A) with a known ground truth, summarized in Tab. 2 where the SCM M and the causal diagram G are given. The source code for the experiment can be found in our repository. |
| Dataset Splits | No | The paper uses a synthetic example where the ground truth is known and compares estimates to this ground truth. It does not describe explicit training, validation, or test splits for this experimental setup. |
| Hardware Specification | No | The paper does not provide any specific details regarding the hardware (e.g., CPU, GPU, memory) used for running the experiments. |
| Software Dependencies | No | The paper does not specify any software dependencies with version numbers. |
| Experiment Setup | Yes | For this example, we set the parameters λ1 = λ2 = λ3 = 0.2. We then vary each parameter λi [0, 0.2] (while keeping the other two parameters fixed), which changes the value of the effect associated with latent variable Ui. The key task is to compute the ground truth values of P(y | x U[i]) for different values of i. According to Def. 1, we want to obtain the conditional distribution of Y given X = x but subject to not updating U[i] according to the evidence X = x. Based on the true SCM, this can be done efficiently using rejection sampling as follows: (1) Take N samples from the SCM M in Tab. 2, (2) For all samples k {1, . . . , N} with u(k) such that X(u(k)) = x, re-sample the part of the unit u(k) that is not included in U[i] (e.g., if U[i] = {U1, U2}, latent u(k) 1 , u(k) 2 are not re-sampled but u(k) 3 is) and replace u(k) with this new sample, (3) Evaluate the mechanisms F of M for all units u(k), (4) If there exists a sample k with X(u(k)) = x go back to Step (2), (5) Return the mean of the Y variables 1N PN k=1 Y (k). |