Abstract
The interpretability of machine learning models has been an essential area of research for the safe deployment of machine learning systems. One particular approach is to attribute model decisions to high-level concepts that humans can understand. However, such concept-based explainability for Deep Neural Networks (DNNs) has been studied mostly on the image domain. In this paper, we extend TCAV, the concept attribution approach, to tabular learning, by providing an idea on how to define concepts over tabular data. On a synthetic dataset with ground-truth concept explanations and a real-world dataset, we show the validity of our method in generating interpretability results that match human-level intuitions. On top of this, we propose a notion of fairness based on TCAV that quantifies what layer of a DNN has learned representations that lead to biased predictions of the model. Also, we empirically demonstrate the relation of TCAV-based fairness to a group fairness notion, Demographic Parity.
Introduction
Deep learning has achieved significant success in computer vision and natural language processing. However, for tabular data—the structured information prevalent in financial and administrative databases—"black-box" models like TabNet are often overlooked in favor of simpler, more interpretable tree-based models. Standard interpretability tools for tabular data typically focus on feature importance, identifying which column influenced a decision without explaining the underlying high-level concepts. For example, while a model might weigh "Age" and "Education," stakeholders often need to know if it is conceptualizing "Seniority" or "Youth." Furthermore, diagnosing fairness is difficult; while bias may be detectable in outputs, identifying the specific neural network layers responsible for learning that bias remains a challenge.
Methodology
This research adapts the Testing with Concept Activation Vectors (TCAV) framework to the tabular domain. The methodology involves four primary components:
- Defining Concepts: Unlike the visual domain where concepts are patterns like "stripes," tabular concepts are defined using boolean predicates (e.g.,
if age < 25: return True). - Generating Examples with CTGAN: Training a Concept Activation Vector (CAV) requires numerous positive and negative examples. To overcome data scarcity in real-world datasets, the authors use CTGAN (a generative adversarial network for tables) to synthesize diverse examples for any given predicate.
- Conceptual Sensitivity: The framework measures how a model’s prediction changes when internal activations are shifted in the direction of a specific CAV.
- TCAV-Based Fairness: A new fairness metric is proposed based on the geometric relationship between model gradients and protected concept vectors. An angle of 90° (orthogonality) indicates that the model is "blind" to a sensitive attribute, representing a fair representation.
Experimental Setup
The method was evaluated using two distinct datasets:
- Synthetic XOR Dataset: A dataset featuring non-linear XOR relationships used to verify if TCAV correctly identifies relevant versus irrelevant concepts against a known ground truth.
- Adult Income Dataset: A standard benchmark for binary classification and fairness, used to predict whether an individual's income exceeds $50k based on census data such as age, education, and gender.
Key Results and Discussion
The experiments demonstrate that TCAV effectively translates high-level concepts into model explanations for tabular data.
Validation and Bias Detection
On the synthetic dataset, TCAV correctly assigned high scores to relevant features and near-zero scores to irrelevant ones. When applied to the Adult Income dataset, TCAV scores mirrored inherent biases in the raw data, showing high sensitivity to concepts like gender == Female or age < 25 in ways that aligned with the dataset's imbalances.
Layer-Wise Diagnosis and Fairness
A significant finding is the ability to perform layer-wise bias diagnosis. TCAV allows researchers to pinpoint exactly where in the network's architecture a model begins to learn biased representations. Furthermore, the proposed TCAV-based fairness metric shows a strong correlation with Demographic Parity. As training data bias increases, the TCAV angles for protected groups drift further from the ideal 90° mark, providing a geometric quantification of model unfairness.
Conclusion
This work extends concept-based interpretability to tabular data, offering a path toward more transparent and fair machine learning systems. While the approach relies on the quality of generative models like CTGAN and currently requires manual concept definition, it provides a robust framework for diagnosing model behavior. Future research will focus on automated concept discovery and the detection of hidden proxy features that may circumvent fairness constraints.
Summary updated