Assessing ML Model Robustness with Item Response Theory

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Aug 13 2024

In a paper published in the journal Machine Learning Science and Technology, researchers presented a framework for assessing the robustness of machine learning (ML) models using item response theory to estimate instance difficulty in supervised tasks.

By introducing perturbation methods that mimic real-world noise, they evaluated performance deviations and developed a taxonomy of ML techniques based on model robustness and instance difficulty. This study highlighted the vulnerabilities of specific ML model families and provided insights into their strengths and limitations.

Related Work

Past work revisited key concepts of model robustness, such as difficulty using item response theory (IRT) and behavioral taxonomies of ML techniques. Robustness was examined through noise simulation, focusing on erroneous values and their impact on performance.

IRT was applied to assess instance difficulty, focusing on binary response models and their associated item characteristic curves. Additionally, using Cohen's kappa statistic to understand their performance in dense and sparse data regions, ML techniques were categorized based on behavioral agreement.

Robustness Evaluation Framework

This section outlines a methodology for assessing the robustness of ML models to noise and instance difficulty. The approach involves estimating instance difficulty using a framework informed by IRT and visualizing it through system characteristic curves (SCCs).

Datasets are perturbed in a controlled manner to reflect real-world disturbances, enabling the construction of robustness taxonomies that cluster models based on their consistency in performance amidst noise and varying instance difficulties. This comprehensive approach offers a detailed understanding of model behavior, aiding in deploying resilient artificial intelligence (AI) systems.

Estimating instance difficulty begins with ensuring each benchmark dataset has sufficient model responses from diverse architectures. A response matrix, denoted U, is formed from these responses, which allows for calculating item characteristic curves (ICCs) to graphically represent the probability of correct predictions as a function of instance difficulty and model ability. The IRT framework focuses on one-parameter logistic (1PL) models where only difficulty and model ability parameters are inferred. This method helps in understanding model behavior relative to instance complexity.

Model robustness is evaluated by introducing random noise—Gaussian for numerical attributes and recalculated probabilities for nominal attributes—across benchmarks, with prediction consistency measured using Cohen's kappa statistic. To create robustness taxonomies, models are clustered based on their performance consistency under noise and difficulty.

Hierarchical clustering analysis helps refine these categories, revealing models that exhibit robust performance across a range of instance difficulties and noise levels. Additionally, average kappa loss analysis provides insight into the robustness of model groups against noise and instance difficulty.

Model Robustness Analysis

The team conducted experiments using the R language and the caret package, with models trained from scratch and the IRT 1PL models estimated using the R package—predictions involving up to 2000 evaluations per dataset. The team used five difficulty bins to maintain clarity and balance in visual representations.

Test datasets were generated with a predefined noise level of ν = 0.2, and the proportion of instances altered by noise was varied—the process for estimating SCCs employed a 5-fold cross-validation framework. Datasets required for IRT difficulty estimation were sourced from OpenML, including 23 benchmarks with varying instances, attributes, and classes, covering diverse domains like handwriting recognition and spam detection.

The analysis of model robustness about noise and instance difficulty revealed various behaviors. The kappa metric compared model predictions on original versus noisy test sets. Different models exhibited different responses to increasing noise levels and varying instance difficulties.

Some models showed stability across difficulty bins but were sensitive to noise, while others were more affected by instance difficulty but less by noise. Notably, models with consistent performance across different problems were generally more robust to noise. In contrast, models with sensitivity to instance difficulty tended to perform poorly with increased noise, especially in challenging cases.

The investigation into dataset complexity revealed that models generally performed more robustly with complex datasets than with simpler ones. This resilience was attributed to the richer feature sets in complex datasets, which buffer against the impacts of noise and difficulty.

In contrast, simpler datasets exhibited more pronounced noise and instance difficulty effects, likely due to class specialization and fewer distinguishing features. This pattern underscores the importance of considering dataset complexity when assessing model robustness, as the complexity of a dataset can significantly influence how models respond to noise and instance challenges.

Conclusion

To sum up, the evaluation framework and taxonomy provided a comprehensive approach for examining ML model robustness against noisy instances, offering insights into model strengths and weaknesses. SCCs helped identify suitable models based on robustness to varying difficulties, and methods were proposed for estimating instance difficulties when unknown.

Future work will extend the framework to additional domains and perturbation functions, exploring model robustness in object detection and text perturbation scenarios. These efforts will deepen understanding and enhance the applicability of the framework.