PnPXAI: A Universal XAI Framework Providing Automatic Explanations Across Diverse Modalities and Models

TL;DR

PnPXAI introduces a universal plug-and-play XAI framework that operates across diverse data modalities and neural architectures. It combines architecture detection, an explanation recommender, evaluation metrics, and hyperparameter optimization to deliver AutoExplanation with minimal user effort. Validation through a user survey and use cases in liver tumor imaging, AKI, and fraud detection demonstrates improved explanation relevance, reliability, and efficiency. This framework enables practitioners to readily select and optimize explainers, accelerating the adoption of XAI in real-world, high-stakes domains.

Abstract

Recently, post hoc explanation methods have emerged to enhance model transparency by attributing model outputs to input features. However, these methods face challenges due to their specificity to certain neural network architectures and data modalities. Existing explainable artificial intelligence (XAI) frameworks have attempted to address these challenges but suffer from several limitations. These include limited flexibility to diverse model architectures and data modalities due to hard-coded implementations, a restricted number of supported XAI methods because of the requirements for layer-specific operations of attribution methods, and sub-optimal recommendations of explanations due to the lack of evaluation and optimization phases. Consequently, these limitations impede the adoption of XAI technology in real-world applications, making it difficult for practitioners to select the optimal explanation method for their domain. To address these limitations, we introduce \textbf{PnPXAI}, a universal XAI framework that supports diverse data modalities and neural network models in a Plug-and-Play (PnP) manner. PnPXAI automatically detects model architectures, recommends applicable explanation methods, and optimizes hyperparameters for optimal explanations. We validate the framework's effectiveness through user surveys and showcase its versatility across various domains, including medicine and finance.

Figures (5)

• Figure 1: Overview of the proposed framework, PnPXAI. The detector module automatically identifies the provided neural network model architecture, which the recommender module uses to filter applicable explanation methods. The evaluator module then optimizes the explanation results through hyperparameter optimization before presenting them to end users.
• Figure 2: Illustration of AutoExplanation for liver tumor detection. PnPXAI recommends 14 applicable explanation methods and optimizes the selection of hyperparameters for each method on the pre-defined objective, ABPC han2023abpc. This optimization improves relevance accuracy when evaluated against the ground truth segmentation mask. The attribution heatmaps at the bottom rows, where higher attribution scores are indicated by more intense red colors, demonstrate that PnPXAI enables to identify whether the model attributes the segments of the liver.
• Figure 3: Evaluation of PnPXAI in AKI detection. The top 5 features identified by the selected explanation methods are compared against known AKI biomarkers (eGFR, creatinine, and BUN). The line graph illustrates the alignment of the ABPC metric with the share of expected features among the most attributed ones.
• Figure 4: Illustration of the interface of a web application for a bank account fraud detection task. The interface allows users to choose specific explainers and evaluators for given data points. It demonstrates the importance of various features in the model's decision-making process and provides evaluation scores for each explanation algorithm.
• Figure : Example code snippet for running AutoExplanation in a plug-and-play manner. This code demonstrates how to initialize and execute the AutoExplanation, showcasing its ease of integration and use within the PnPXAI framework.