PCEvE: Part Contribution Evaluation Based Model Explanation for Human Figure Drawing Assessment and Beyond

Jongseo Lee, Geo Ahn, Jinwoo Choi, Seongtae KimarXiv 2024 · OpenReview

Abstract

For automatic human figure drawing (HFD) assessment tasks, such as diagnosing autism spectrum disorder (ASD) using HFD images, the clarity and explainability of a model decision are crucial. Existing pixel-level attribution-based explainable AI (XAI) approaches demand considerable effort from users to interpret the semantic information of a region in an image, which can be often time-consuming and impractical. To overcome this challenge, we propose a part contribution evaluation based model explanation (PCEvE) framework. On top of the part detection, we measure the Shapley Value of each individual part to evaluate the contribution to a model decision. Unlike existing attribution-based XAI approaches, the PCEvE provides a straightforward explanation of a model decision, i.e., a part contribution histogram. Furthermore, the PCEvE expands the scope of explanations beyond the conventional sample-level to include class-level and task-level insights, offering a richer, more comprehensive understanding of model behavior. We rigorously validate the PCEvE via extensive experiments on multiple HFD assessment datasets. Also, we sanity-check the proposed method with a set of controlled experiments. Additionally, we demonstrate the versatility and applicability of our method to other domains by applying it to a photo-realistic dataset, the Stanford Cars.

Introduction

Deep learning models have achieved high accuracy in specialized tasks like assessing Human Figure Drawings (HFD) for Autism Spectrum Disorder (ASD) screening. However, their "black-box" nature remains a barrier to clinical adoption. Existing Explainable AI (XAI) methods, such as Grad-CAM or LIME, typically produce pixel-level heatmaps. While visually striking, these "glowing blobs" often fail to provide actionable insights for clinicians who reason using high-level semantic concepts like "eyes," "hands," or "hair." This "semantic gap" makes interpretation time-consuming and subjective.

Comparison of existing attribution-based XAI versus the proposed multi-level PCEvE framework.

To address this, we propose the Part Contribution Evaluation based model explanation (PCEvE) framework. PCEvE bridges the gap by quantifying the contribution of specific semantic parts to a model's decision, providing intuitive histograms that align with human expert reasoning.

The PCEvE Framework

PCEvE treats model explanation as a cooperative game where image parts are "players" collaborating to produce a prediction. The framework operates in two primary stages:

1. Image Partition and Set Generation

The system first identifies $K$ semantic parts (e.g., head, torso, limbs) using either ground-truth labels or automated detectors like YOLO. It then generates a power set of $2^K$ image variations by masking different combinations of these parts. Masked regions are filled with uninformative values (e.g., the image's average pixel value) to simulate the absence of that feature.

2. Shapley Value Calculation

Using cooperative game theory, PCEvE calculates the Shapley Value for each part. This metric represents the average marginal contribution of a part across all possible coalitions, ensuring a mathematically fair distribution of "credit" for the model's output.

The PCEvE framework stages: from image partitioning to multi-level Shapley value calculation.

Multi-Level Explanations

Unlike standard XAI tools that focus solely on individual samples, PCEvE offers three levels of insight:

  • Sample-level (S-PCEvE): Explains a single prediction via a part-contribution histogram.
  • Class-level (C-PCEvE): Aggregates sample-level data to identify which parts characterize a specific category (e.g., identifying "Hair" as the primary feature for "Female" drawings).
  • Task-level (T-PCEvE): Provides a global overview of feature importance across the entire dataset.

Experimental Validation

We evaluated PCEvE across three diverse datasets: ASD Screening (100 HFD sketches), SCAT (28,000 drawings for gender classification), and Stanford Cars (16,185 photo-realistic images).

Key Results

  • Clinical Alignment: In ASD screening, PCEvE identified the "Face" area (excluding specific features like eyes/mouth) as the most significant contributor. Removing this "important" part led to a 25% drop in model accuracy, whereas removing "unimportant" parts had negligible impact.
  • Robustness: Explanations derived from automated YOLOv8 detections showed a 0.99 cosine similarity to those using ground-truth annotations, proving the framework's practical utility.
  • Model Agnostic: PCEvE provided consistent explanations across various architectures, including ResNet, EfficientNet, and Vision Transformers (ViT).

Conclusion

PCEvE transforms cryptic pixel-level heatmaps into intuitive, semantic explanations. By leveraging Shapley Values at the part level, it provides a transparent and multi-faceted view of model behavior that mirrors human expert logic. Future work will explore unsupervised part discovery and the integration of Large Language Models to generate natural language diagnostic reports.

Summary updated

Abstract

For automatic human figure drawing (HFD) assessment tasks, such as diagnosing autism spectrum disorder (ASD) using HFD images, the clarity and explainability of a model decision are crucial. Existing pixel-level attribution-based explainable AI (XAI) approaches demand considerable effort from users to interpret the semantic information of a region in an image, which can be often time-consuming and impractical. To overcome this challenge, we propose a part contribution evaluation based model explanation (PCEvE) framework. On top of the part detection, we measure the Shapley Value of each individual part to evaluate the contribution to a model decision. Unlike existing attribution-based XAI approaches, the PCEvE provides a straightforward explanation of a model decision, i.e., a part contribution histogram. Furthermore, the PCEvE expands the scope of explanations beyond the conventional sample-level to include class-level and task-level insights, offering a richer, more comprehensive understanding of model behavior. We rigorously validate the PCEvE via extensive experiments on multiple HFD assessment datasets. Also, we sanity-check the proposed method with a set of controlled experiments. Additionally, we demonstrate the versatility and applicability of our method to other domains by applying it to a photo-realistic dataset, the Stanford Cars.
BibTeX
@misc{lee2024pceve,
  author = {Jongseo Lee and Geo Ahn and Jinwoo Choi and Seongtae Kim},
  title = {{PCEvE: Part Contribution Evaluation Based Model Explanation for Human Figure Drawing Assessment and Beyond}},
  year = {2024},
  eprint = {2409.18260},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2409.18260},
}
Figure