🤖 machine learning

AGOP-IxG: A Gradient Covariance Filter for Local Feature Attribution on Tabular Data, with a Controlled Benchmark

This paper introduces AGOP-IxG, a fast per-sample feature attribution method for tabular data that leverages a gradient covariance filter to outperform widely used baselines like SHAP and LIME in accuracy and speed on synthetic benchmarks, while demonstrating comparable global faithfulness on real-world datasets.

Original authors: Raj Kiran Gupta Katakam

Published 2026-05-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Raj Kiran Gupta Katakam

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very smart, but mysterious, robot that makes decisions about things like loan approvals or medical risks. You know what it decided, but you don't know why. To fix this, we use "explanation tools" that try to point a finger at the specific clues the robot used to make its choice.

The paper you provided introduces a new tool called AGOP-IxG and tests it against the four most popular tools currently in use (SHAP, Integrated Gradients, LIME, and InputXGradient).

Here is the breakdown of their findings, using simple analogies:

1. The Problem: The "Blind" Evaluation

Currently, when people choose an explanation tool, they often pick one just because it's famous or because it looks good to experts. The problem is that on real-world data, no one knows the "true" answer. It's like trying to grade a student's essay without an answer key; you can only guess if they are right.

The authors built a controlled laboratory (synthetic datasets) where they did create the answer key. They designed three specific puzzles:

The Linear Puzzle: A simple, straight-line relationship (easy).
The Sparse Nonlinear Puzzle: A complex relationship where only a few clues matter (medium).
The Interaction Puzzle: A puzzle where clues only matter if they are combined with other clues (hard).

Because they built these puzzles, they knew exactly which clues were important and which were just "noise" (fake clues).

2. The New Tool: AGOP-IxG

The authors created a new method called AGOP-IxG.

How it works: Imagine the robot's brain is a giant map of directions. Most explanation tools just look at the direction the robot is pointing right now. AGOP-IxG is smarter: it looks at the entire history of how the robot learned. It builds a "noise filter" based on that history.
The Analogy: Think of the robot's decision as a radio signal. The "true" clues are the music, and the "noise" is static. AGOP-IxG is like a high-end radio tuner that knows exactly which frequencies the music usually occupies. It amplifies the music and completely silences the static. Other tools often amplify the static along with the music.

3. The Results: Speed and Accuracy

The authors ran a race between AGOP-IxG and the four other tools on their puzzles.

The Accuracy Race:

Noise Filtering: AGOP-IxG was the best at ignoring "noise." While other tools sometimes pointed at fake clues, AGOP-IxG almost always ignored them.
Ranking: It was the most consistent at ranking the clues in the correct order of importance.
The Exception: On the simplest puzzle (Linear), one old tool (LIME) was slightly better at picking the top clue, but AGOP-IxG was still very close and much better at ignoring the fake clues.

The Speed Race:

This is where AGOP-IxG wins by a landslide.
The Analogy: If the other tools are like a team of 100 accountants manually checking every single number in a spreadsheet, AGOP-IxG is like a super-fast calculator that does the same math in a split second.
The Numbers: AGOP-IxG was 350 to 1,650 times faster.
- AGOP-IxG: Takes about 0.02 to 0.1 seconds per decision.
- SHAP (the most popular competitor): Takes about 33 seconds per decision.
- Real-world impact: If you needed to explain one million decisions, AGOP-IxG would take less than 2 minutes. SHAP would take about 9 hours.

4. The Reality Check: Real Data

The authors also tested these tools on real-world data (Adult Income and Credit Card Default datasets).

The Finding: On real data, where there is no "answer key," all five tools performed almost identically. They all clustered together.
The Lesson: This suggests that for global questions (like "What are the top 5 features for the whole bank?"), it doesn't matter much which tool you pick. However, for local questions (like "Why was this specific person denied a loan?"), AGOP-IxG is the superior choice because it is faster and more accurate at the individual level.

Summary

The paper argues that while we often pick explanation tools by habit, we should pick them by performance.

AGOP-IxG is a new tool that acts like a noise-canceling headphone for AI explanations.
It is much faster than current tools, making it possible to explain decisions in real-time (like instantly when you apply for a loan).
It is more accurate at finding the true reasons for a decision and ignoring fake ones, especially on complex problems.

The authors conclude that AutoML systems (systems that build AI automatically) should use tools like AGOP-IxG when they need to explain individual decisions quickly and accurately, rather than relying on slower, older methods.

Technical Summary: AGOP-IxG for Local Feature Attribution on Tabular Data

Problem Statement

Automated Machine Learning (AutoML) pipelines increasingly deploy black-box models in high-stakes domains (e.g., credit scoring, medical risk), necessitating per-prediction explanations for regulators and end-users. While feature attribution methods like SHAP, Integrated Gradients (IG), and LIME are standard, their selection is often driven by convention rather than measured fidelity. A fundamental evaluation gap exists: real-world tabular data lacks ground-truth attribution, forcing reliance on circular proxies (e.g., inter-method agreement) or global metrics (e.g., ROAR) that do not assess the quality of individual per-sample explanations. Consequently, practitioners lack a principled basis for selecting an attribution method that balances accuracy, noise robustness, and computational cost.

Methodology

The Proposed Method: AGOP-IxG

The authors propose AGOP-IxG, a fast per-sample attribution method for tabular classifiers that combines InputXGradient with a gradient filter derived from the Average Gradient Outer Product (AGOP) matrix.

Training Stage (Fit):
- For a training set of $n$ samples, the gradient of the predicted class score ( $g_i = \nabla_x \hat{s}_{c_i}(x_i)$ ) is computed for each sample.
- The AGOP matrix $M$ is constructed as the empirical gradient outer-product: $M = \frac{1}{n} \sum_{i=1}^n g_i g_i^\top$ .
- $M$ is eigendecomposed ( $M = V\Lambda V^\top$ ). A rank- $K$ truncation is applied, retaining only eigenvectors where eigenvalues exceed 1% of the maximum eigenvalue ( $\lambda_{max}$ ). This creates a low-rank reconstruction $M_K = V_K \Lambda_K V_K^\top$ .
- This process acts as a noise filter, amplifying gradient components aligned with high-variance training-distribution directions while zeroing out low-eigenvalue noise directions.
Attribution Stage (Test):
- For a test sample $x$ with baseline $x'$ (per-feature training mean), the gradient $g$ is computed.
- The gradient is pre-multiplied by the truncated AGOP matrix: $g_{attr} = g M_K$ .
- The final attribution $e_j$ is calculated as the product of the feature difference and the filtered gradient: $e_j = (x_j - x'_j) \cdot (g_{attr})_j$ .
- Attributions are normalized to sum to 1.

Computational Efficiency: By utilizing the low-rank factorization ( $M_K = V_K \Lambda_K V_K^\top$ ), the method avoids forming the full $d \times d$ matrix, requiring only two matrix-vector products. This results in a runtime approximately 350× to 1,650× faster than SHAP.

Evaluation Framework

The paper introduces a two-part benchmark:

Part 1: Controlled Synthetic Benchmarks: Three multi-class tabular tasks (Linear, Sparse Nonlinear, and Feature Interaction) are generated where ground-truth attribution is analytically or numerically derivable. Metrics include Spearman rank correlation, Top- $k$ precision, and Noise Feature Mass (attribution assigned to irrelevant features).
Part 2: Real-World Global Faithfulness: The ROAR (Remove And Retrain) protocol is applied to the Adult Income and Credit Card Default datasets to measure global feature ranking fidelity.

Key Contributions

AGOP-IxG Algorithm: A novel attribution scheme that composes InputXGradient with a top- $K$ rank-truncated AGOP gradient filter. It is designed specifically for per-sample local attribution fidelity.
Controlled Evaluation Framework: A parametric synthetic benchmark suite (Linear, Sparse Nonlinear, Interaction) with known ground-truth attribution, enabling direct evaluation of per-sample quality without proxy measures. The framework is model-agnostic and designed for reuse in AutoML pipelines.
Comprehensive Empirical Analysis: A comparison of AGOP-IxG against four baselines (SHAP, Integrated Gradients, InputXGradient, LIME) across synthetic and real datasets, highlighting the trade-off between local fidelity and global ranking.

Results

Part 1: Synthetic Benchmarks

Noise Robustness: AGOP-IxG achieved the lowest noise feature mass on all three synthetic datasets, significantly outperforming SHAP, IG, and LIME. This validates the efficacy of the AGOP truncation in filtering out noise.
Rank Correlation: AGOP-IxG led in Spearman rank correlation on all three datasets.
Top- $k$ Precision:
- Linear Task: LIME slightly outperformed AGOP-IxG (0.915 vs. 0.895), consistent with LIME's suitability for linear processes.
- Sparse Nonlinear: AGOP-IxG and SHAP were nearly tied (0.514 vs. 0.515), though AGOP-IxG led in Spearman correlation.
- Interaction: AGOP-IxG led SHAP on all metrics (Spearman 0.534 vs. 0.466).
Runtime: AGOP-IxG was orders of magnitude faster (0.02–0.10s per dataset) compared to SHAP (~33s).

Part 2: Real-World Data (ROAR)

On the Adult Income and Credit Card Default datasets, all five methods produced closely clustered results (AUC spread < 0.01).
No single method dominated global feature ranking. AGOP-IxG ranked last on Adult and third on Credit, but the differences were marginal (~1.7% relative change).
The authors note this is consistent with AGOP-IxG's design: it optimizes for local per-sample fidelity, whereas ROAR evaluates global rankings derived from averaging normalized vectors.

Significance and Claims

The paper claims that AGOP-IxG provides a strong inductive bias for local per-prediction explainability without sacrificing runtime, making it operationally valuable for AutoML pipelines with strict time budgets (e.g., real-time inference).

Method Selection Guidance: The results suggest that the choice of attribution method should depend on the downstream consumer. For per-decision explanations (e.g., regulatory reports on specific denials), AGOP-IxG offers superior local fidelity and noise robustness. For global feature selection within an AutoML pipeline, the choice of method matters less, as methods cluster closely in global ranking performance.
Deployability: The sub-millisecond per-sample cost of AGOP-IxG enables real-time deployment (e.g., attaching reason codes to every denied credit decision), whereas SHAP's latency makes it prohibitive at scale (e.g., explaining one million decisions would take ~9 hours with SHAP vs. <2 minutes with AGOP-IxG).
Benchmark Utility: The proposed synthetic generators serve as a method-agnostic tool for future AutoML systems to auto-select or auto-tune attribution methods based on measured fidelity rather than convention.

The authors remain modest, acknowledging that results are based on MLPs and synthetic tasks, and that ground-truth attribution is unavailable on real data. They do not claim AGOP-IxG is universally superior but rather that it excels in the specific regime of local, noise-robust, real-time attribution.