Conditional Local Importance by Quantile Expectations

This paper introduces CLIQUE, a novel model-agnostic method for calculating local variable importance that overcomes the limitations of existing techniques like LIME and SHAP by capturing locally dependent relationships and interaction behaviors while being natively adapted for multi-class classification problems.

The Gist

Imagine you have a very smart, but mysterious, robot (a machine learning model) that makes predictions. Maybe it predicts how strong a piece of concrete will be, or whether a picture is a cat or a dog. You want to know why the robot made a specific decision for a specific situation.

This is where "Variable Importance" comes in. It's like asking, "Which ingredient mattered most in this specific cake?" or "Which clue solved this specific mystery?"

For a long time, scientists have had tools to answer this, like LIME, SHAP, and ICI. But the authors of this paper argue that these old tools have a big flaw: they often look at the "average" behavior of a variable and miss the specific context. They are like a detective who only looks at the suspect's criminal record from the last 10 years, ignoring what they were doing right now.

The authors introduce a new tool called CLIQUE (Conditional Local Importance by QUantile Expectations). Here is how it works, explained simply:

The Problem: The "Average" Trap

Let's use an analogy of a Traffic Light.

• The Rule: The light turns red only if both the "Pedestrian Button" is pressed AND the "Car Sensor" detects a vehicle.
• The Flaw in Old Tools: If you ask an old tool (like LIME or SHAP) about the "Pedestrian Button" on a day when no cars are around, it might say, "This button is important!" because, in general, buttons are important. It misses the fact that today, without a car, the button does absolutely nothing. It gives you a "false alarm" of importance.

The Solution: CLIQUE

CLIQUE is like a detective who checks the specific context before making a judgment.

1. It Tests "What If?" Scenarios: Instead of just looking at the data, CLIQUE plays a game of "What if?" for a specific moment.

   • Scenario: "Okay, the car sensor is off. What happens if I change the pedestrian button?"
   • Result: The light stays green. The button didn't matter.
   • CLIQUE's Verdict: "Importance = 0."
   • Old Tool's Verdict: "Importance = High!" (Because buttons are usually important).

2. It Measures "Mistakes," Not Just "Guesses":
   Most tools look at how much the robot's prediction changes. CLIQUE looks at how much the robot's mistake (error) changes.

   • Imagine you are guessing the weight of a watermelon.
   • If you change the color of the rind, and your guess stays the same, you didn't make a mistake.
   • If you change the size, and your guess becomes wildly wrong, that feature was crucial.
   • CLIQUE focuses on where the model gets confused. If changing a variable doesn't make the model mess up more, that variable isn't important in this specific situation.

Why CLIQUE is a Game-Changer

• It's a "Zero-Importance" Machine: If a variable truly doesn't matter for a specific data point, CLIQUE gives it a score of zero. The old tools often give it a tiny, confusing number, making you think it matters when it doesn't.
• It Handles Complex Relationships: In the "Traffic Light" example, the button only matters if the car is there. CLIQUE understands this "If/Then" relationship perfectly. The old tools struggle with this and often get it wrong.
• It Works for Many Classes: Imagine a robot sorting fruits. Old tools get confused when there are 100 types of fruit (multi-class). They have to be tricked into looking at one fruit at a time. CLIQUE can look at the whole basket of fruit and explain the decision for any specific fruit naturally.

Real-World Examples from the Paper

• Concrete Strength: The paper looked at how concrete hardens. They found that "Cement" is super important when the concrete is young, but less important once it's old. Old tools saw "Cement is important" and stopped there. CLIQUE saw, "Cement is critical only in the first 75 days."
• Lichen Survival: Lichen (a type of plant) needs specific temperatures to survive. The paper found that "Tree Age" only matters if the temperature is warm enough. If it's freezing, tree age doesn't matter at all. CLIQUE correctly said, "Tree Age = 0 importance" in the freezing zones. Old tools kept saying tree age was important everywhere.
• Handwritten Digits (MNIST): When identifying the number "5," the importance of a specific pixel depends on the shape of the other pixels. CLIQUE figured out that a specific pixel is only a clue for a "5" if the surrounding pixels look a certain way.

The Bottom Line

Think of LIME and SHAP as a weather forecaster who says, "It's usually rainy in April," so they tell you to bring an umbrella every day in April, even if it's sunny today.

CLIQUE is the smart forecaster who looks out the window right now. If it's sunny, it says, "Leave the umbrella at home." If it's raining, it says, "Take the umbrella."

CLIQUE is a new, smarter way to explain AI decisions that respects the specific context of every single situation, avoiding false alarms and giving us a clearer, more honest picture of how these models really think.

Technical Summary

Here is a detailed technical summary of the paper "Conditional Local Importance by Quantile Expectations" (CLIQUE).

1. Problem Statement

Local variable importance (LVI) methods aim to explain how specific features contribute to the prediction of individual observations. While popular methods like LIME (Local Interpretable Model-agnostic Explanations) and SHAP (Shapley Additive Explanations) are widely used, they suffer from critical limitations:

• Failure to Capture Conditional Dependencies: These methods often focus on marginal importance, failing to accurately reflect relationships where a variable's importance depends on the value of another variable (interaction effects).
• False Positives: They frequently assign non-zero importance to variables that have no effect on the response in specific regions of the feature space (locally invariant features).
• Multi-class Complexity: Extending prediction-based decomposition methods (like SHAP) to multi-class classification is complex and often requires ad-hoc adaptations (e.g., one-vs-all).
• High Variance: Permutation-based local methods (like Individual Conditional Importance, ICI) can exhibit high variance and require computationally expensive repeated permutations.

2. Methodology: CLIQUE

The authors propose CLIQUE (Conditional Local Importance by QUantile Expectations), a model-agnostic framework designed to assess local variable importance based on model error rather than prediction changes.

Core Definition

For a given observation $x_i$ and feature $j$, CLIQUE calculates the importance $V_{ij}$ by measuring the change in model loss when feature $j$ is systematically altered while all other features remain fixed.

The formula is defined as:
$$V_{ij} = \frac{1}{M} \sum_{m=1}^{M} \left[ L(\hat{f}(\tilde{x}_i(j, m)), y_i) - L(\hat{f}(x_i), y_i) \right]$$

Where:

• $\hat{f}$ is a prediction function trained via Cross-Validation (CV) (specifically, a model trained on a fold excluding $x_i$ to prevent overfitting).
• $L$ is a loss function (e.g., squared error, absolute error).
• $\tilde{x}_i(j, m)$ is the altered observation where feature $j$ is replaced by the $m$-th value from a quantile grid constructed from the training data distribution.
• $M$ is the number of quantile points (e.g., 25).

Key Mechanisms

1. Quantile Grid vs. Random Permutation: Instead of random permutations (which introduce noise and high variance), CLIQUE replaces the feature value with deterministic quantile points from the feature's distribution. This stabilizes the estimate and reduces the number of perturbations needed.
2. Error-Based Metric: By measuring the change in loss (error) rather than the change in prediction, CLIQUE naturally handles multi-class classification without needing separate calculations for each class.
3. Cross-Validation: Importance is calculated using models trained on data excluding the target observation, ensuring the metric reflects generalizable behavior rather than training artifacts.

3. Key Properties

The authors establish eight theoretical and practical properties of CLIQUE:

• P1 (Feature Invariance): If a variable has no effect on the model output for a specific observation, CLIQUE assigns it zero importance. This is a critical theoretical guarantee not met by LIME, SHAP, or ICI.
• P2 (Stability): Quantile replacements yield lower variance compared to random permutations.
• P3 (Model-Agnostic): Works with any predictive model (Random Forests, Neural Networks, etc.).
• P4 (Multi-class Native): Directly applicable to multi-class problems without decomposition.
• P5 (Aggregability): Local values can be aggregated (mean, median, dispersion) to assess subgroup or global behavior.
• P6 (Computational Efficiency): Competitive with existing methods; scales linearly with observations.
• P7 (Error-Centric): Focuses on predictive performance degradation rather than prediction shifts.
• P8 (Out-of-Sample): Uses CV errors to avoid overfitting bias.

4. Experimental Results

Simulated Experiments

The authors tested CLIQUE against LIME, SHAP, and ICI on three synthetic datasets with known ground-truth interactions:

1. AND Gate Data: A scenario where feature $v_1$ is only important if $v_2 > -1/3$.
   • Result: CLIQUE correctly assigned zero importance to $v_1$ when $v_2 < -1/3$. LIME, SHAP, and ICI all assigned false-positive importance in this region.
2. Corners Data: A non-symmetric interaction where importance depends on specific thresholds of other variables.
   • Result: CLIQUE captured the conditional zero-importance regions perfectly. Competing methods failed to distinguish between relevant and irrelevant regions.
3. Regression Interaction: A location-dependent regression where $v_1$ is only relevant if $v_3 > 0$.
   • Result: CLIQUE again showed near-zero false positives, while others assigned significant importance to irrelevant features.

Quantitative Metric: The authors introduced False-Positive Mean Absolute Error (FP-MAE). CLIQUE consistently achieved FP-MAE values an order of magnitude lower than competitors (e.g., 0.005 vs. 0.405 for LIME in the AND Gate scenario).

Real-World Applications

1. Concrete Compressive Strength (Regression):
   • CLIQUE revealed that the importance of "Cement" is significantly higher for younger concrete (Age < 75 days) compared to older concrete. LIME and ICI showed little difference between these groups.
2. Lichen Classification (Binary):
   • In predicting lichen presence, CLIQUE showed that the importance of "Conifer Age" (ACONIF) is near zero when minimum temperatures are low (where lichen cannot survive). SHAP showed some variation, but LIME and ICI failed to detect this conditional null effect.
3. MNIST Digit Classification (Multi-class):
   • CLIQUE successfully identified pixel interactions. For example, pixel $x_4y_6$ was more important when $x_3y_6$ was low.
   • It also distinguished subgroups within digit classes (e.g., different ways of writing the number "5") and identified specific pixels critical for distinguishing digit "6" from others, demonstrating its utility in high-dimensional multi-class settings.

5. Significance and Conclusion

The paper presents CLIQUE as a robust, theoretically grounded alternative to current local importance methods. Its primary contributions are:

• Elimination of False Positives: By strictly adhering to the principle that invariant features should have zero importance, CLIQUE provides more trustworthy local explanations.
• Superior Interaction Detection: It effectively captures conditional dependencies that marginal-based methods (LIME) or biased conditional methods (SHAP) miss.
• Unified Multi-class Framework: It offers a native solution for multi-class problems without complex decompositions.
• Stability: The use of quantile grids reduces the variance inherent in permutation-based approaches.

The authors conclude that CLIQUE offers a principled shift from "prediction change" to "error reduction," providing a more accurate lens for understanding how machine learning models behave locally, particularly in complex, non-linear, and multi-class scenarios.