CIDR: A Cooperative Integrated Dynamic Refining Method for Minimal Feature Removal Problem

Qian Chen, Taolin Zhang, Dongyang Li, Xiaofeng HeAAAI 2024 · DOI

Abstract

Identifying the Minimal Feature Set (MFS)—the smallest subset of input features whose removal alters a model's prediction—is a critical task for interpreting deep learning models in Natural Language Processing (NLP). Traditional gradient-based explanation methods often fail to account for non-linear word interactions and the non-monotonic nature of prediction probabilities. We present CIDR (Cooperative Integrated Dynamic Refining), a framework that addresses these challenges by modeling feature interactions through Cooperative Integrated Gradients (CIG) and optimizing feature selection via a knapsack-based refinement strategy. Our evaluation across multiple sentiment analysis benchmarks shows that CIDR consistently produces more precise and interpretable explanations than state-of-the-art baselines.

1. Introduction

In the context of "black-box" neural networks, post-hoc explanation methods are essential for understanding model behavior. A prominent approach is identifying the Minimal Feature Set (MFS): the smallest group of input features (e.g., words) that, if removed, would cause the model to flip its prediction.

However, existing methods for finding the MFS face two primary hurdles:

  1. The Monotonicity Trap: Many methods assume that removing features leads to a steady, monotonic decrease in prediction probability. In complex language models, non-linearities often violate this assumption.
  2. Ignoring Synergy: Language is combinatorial; the meaning of a phrase often exceeds the sum of its individual words (e.g., "not bad"). Most methods analyze words in isolation, leading to redundant or misleading explanations.

CIDR (Cooperative Integrated Dynamic Refining) is the first framework to rigorously address the MFS problem specifically for NLP by accounting for these complex interactions.

2. Methodology

The CIDR method is built upon three main technical pillars:

  • Cooperative Integrated Gradients (CIG): While standard Integrated Gradients (IG) measure individual word importance, CIG extends this to calculate the "cooperative contribution" of word pairs. This allows the model to detect whether words work together (cooperative) or against each other (competitive) to influence the output.
  • The Knapsack Transformation: The authors frame the MFS search as a variation of the classic Knapsack Problem. In this formulation, word pairs are treated as "items" with CIG scores acting as "weights." The objective is to maximize the number of non-essential features "packed" away, leaving only the vital minimal set.
  • Minimal Feature Refinement: To handle the unknown "capacity" of the knapsack, CIDR employs a perturbation strategy. By adding Gaussian noise to weights and running the optimization multiple times, the algorithm generates a set of candidates. A "frequency filter" then retains only those features that consistently appear across runs, effectively discarding false positives.
The CIDR framework architecture, detailing the stages of feature interaction detection and refinement.

3. Experimental Setup

The performance of CIDR was evaluated using two transformer architectures—BERT and DistilBERT—across three benchmark sentiment analysis datasets:

  • SST2: Short movie reviews (binary classification).
  • IMDB: Long-form movie reviews, testing long-distance dependency handling.
  • Rotten Tomatoes: Short-text sentiment data.

The researchers compared CIDR against six established baselines, including Integrated Gradients (IG), DeepLift, and Sequential Integrated Gradients (SIG).

4. Results and Discussion

CIDR demonstrated superior performance across all metrics, including Comprehensiveness (Comp), Log-odds (LO), and a newly introduced Feature Minimality Score (FMS).

Quantitative results comparing CIDR against various baseline methods on BERT and DistilBERT.

Key findings include:

  • Superior Precision: CIDR consistently outperformed all baselines. On the SST2/BERT setup, CIDR achieved a +19.1% improvement in the FMS metric compared to the previous state-of-the-art.
  • Interpretability: Qualitative analysis showed that CIDR identifies cohesive, semantically meaningful phrases (e.g., "good-natured and never dull") rather than isolated adjectives.
  • Ablation Studies: Removing either the Cooperative (CIG) component or the Refinement (R) component resulted in significant performance drops, validating the necessity of both interaction modeling and frequency filtering.

5. Conclusion and Future Work

CIDR represents a significant advancement in generating robust, minimal explanations for NLP models. By treating word pairs as units of analysis and utilizing a knapsack-based optimization, it captures the semantic "glue" of language more effectively than previous methods.

Future research directions include improving scalability for extremely long documents, extending the framework to multi-class and generative tasks, and utilizing CIDR as a diagnostic tool to uncover and mitigate spurious correlations in deep learning models.

Summary updated

BibTeX
@inproceedings{chen2024cidr,
  author = {Qian Chen and Taolin Zhang and Dongyang Li and Xiaofeng He},
  title = {{CIDR: A Cooperative Integrated Dynamic Refining Method for Minimal Feature Removal Problem}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i16.29729},
}
Figure