Explaining Classifiers with Causal Concept Effect (CaCE)

TL;DR

CaCE defines the causal effect of a human-interpretable concept on a classifier's output, addressing confounding that plagues correlation-based explanations. It introduces Ground Truth CaCE and two VaE-based estimators (Dec-CaCE and EncDec-CaCE) using conditional VAEs to approximate do-operations in image generation. Across synthetic and real datasets, VaE-CaCE closely matches GT-CaCE and outperforms non-causal baselines like ConExp and TCAV, with diagnostic tests validating the approach. The work provides a practical, causal framework for interpreting global classifier decisions in high-dimensional settings, paving the way for more trustworthy explanations.

Abstract

How can we understand classification decisions made by deep neural networks? Many existing explainability methods rely solely on correlations and fail to account for confounding, which may result in potentially misleading explanations. To overcome this problem, we define the Causal Concept Effect (CaCE) as the causal effect of (the presence or absence of) a human-interpretable concept on a deep neural net's predictions. We show that the CaCE measure can avoid errors stemming from confounding. Estimating CaCE is difficult in situations where we cannot easily simulate the do-operator. To mitigate this problem, we use a generative model, specifically a Variational AutoEncoder (VAE), to measure VAE-CaCE. In an extensive experimental analysis, we show that the VAE-CaCE is able to estimate the true concept causal effect, compared to baselines for a number of datasets including high dimensional images.

Figures (5)

• Figure 1: Causal graph relating high-level concepts (such as objects, background, lighting conditions, etc.), images and classifier $f$ output. The dashed edge indicates possible confounding of the two concepts, by other concepts (not shown in the graph). The thick arrow from Image to the output of $f$ indicates that this relation is mechanistic and we have direct access to it through our knowledge of $f$. This is different from the edges connecting the concepts to the Image, which correspond to the natural generation process of images. In Section \ref{['sec:meth']}, we propose using a conditional-VAE conditioned on concepts to approximate this relation.
• Figure 2: Our Dec-CaCE approach generates pairs of images from the decoder with the same latent sample vector $z$ but with different values of the concept $C$, resulting in pairs of counterfactual images which differ from each other in terms of the concept only.
• Figure 3: Our EncDec-CaCE approach first infers the latent distribution for a given image and its corresponding concept label using the encoder and then generates a counterfactual image from the decoder using a sample from the inferred latent distribution and the flipped value of the concept $C$.
• Figure 4: Colored-MNIST images from the test set (leftmost column) alongside their counterfactuals generated from DC-VAE. Each of the 2-14 columns correspond to one possible value for the color concept. Each row corresponds to one image from each digit class.
• Figure 5: CelebA test images (with eyeglasses) along with their counterfactuals (without eyeglasses) generated from the DC-VAE.

Theorems & Definitions (1)

• Definition 1: Causal Concept Effect, CaCE