Learning to Abstain in the Presence of Uninformative Data

Yikai Zhang, Songzhu Zheng, Pengxiang Wu, Yuriy Nevmyvaka, Chao ChenICLR 2022 · OpenReview

TL;DR

This paper introduces a novel machine learning approach designed to handle datasets where a significant portion of the data is essentially random noise and uninformative. By developing a "selector" that learns to identify and abstain from this noisy data, the method can make more reliable predictions on the truly informative parts, even in highly imbalanced scenarios. This is crucial for applications in fields like finance and healthcare where data is often noisy.

What Problem Does This Paper Address?

Many real-world machine learning applications, particularly in domains like finance and public health, suffer from a high noise-to-signal ratio. This means that a substantial fraction of the available data might not contain any useful information for learning a predictive model. Standard supervised learning methods struggle in such environments. During training, this noisy data can heavily bias the model or even completely overwhelm the true signal, leading to unreliable predictions. Deep learning models, with their strong memorization capabilities, are particularly susceptible to overfitting this noise, making overly confident predictions on data that lacks any underlying structure. The core problem is how to build robust models that can effectively ignore or abstain from uninformative data during both training and inference, rather than being misled by it.

Methods

The authors propose a novel approach that combines a predictor (a standard classifier) with a selector. The selector's job is to learn to distinguish between informative and uninformative data points. Instead of forcing a prediction for every sample, the method learns to abstain from making predictions when the selector deems the data uninformative.

The key innovation lies in a novel selector loss function. This loss is inspired by selective learning theory and is designed to train the selector without direct supervision. It penalizes incorrect predictions on informative data and also discourages the selector from incorrectly abstaining from informative data. The theoretical analysis shows that minimizing this loss, in conjunction with minimizing the predictor's risk, can lead to a near-optimal decision-maker that can reliably identify uninformative samples.

The paper also presents a heuristic iterative algorithm that jointly optimizes the predictor and the selector. This algorithm addresses the practical challenge of optimizing complex models and non-convex losses. It iteratively updates the predictor using data selected by the selector and then updates the selector based on the predictor's performance. A "soft abstention" scheme is used to gradually down-weight uninformative data during predictor updates.

Datasets

The paper uses synthetic datasets for experiments, created by mixing datasets from different domains. Specifically:

  • MNIST and Fashion-MNIST: These are used to create a scenario where one dataset (shuffled MNIST) serves as uninformative data, and the other (Fashion-MNIST) as informative data.
  • SVHN: This dataset is used to create a more challenging scenario where both informative and uninformative data come from similar domains, making it harder to distinguish them.

Key Insights and Ideas

  • Abstention as a Solution: The core idea that abstaining from prediction on uninformative data is a more robust strategy than trying to force a prediction is a key insight. This is particularly relevant in domains with inherent noise.
  • Learning to Abstain: The paper demonstrates that it's possible to learn to abstain effectively, not just rely on heuristics. The proposed selector loss function is central to this.
  • Theoretical Guarantees: The theoretical analysis provides strong guarantees for the proposed method, showing it can recover the ground truth selector with high probability, even when uninformative data is the majority. This is a significant theoretical contribution.
  • Robustness to Imbalance: The method is shown to be effective even in highly imbalanced settings, where informative data is scarce.
  • Practical Algorithm: The development of an iterative heuristic algorithm makes the theoretically sound approach practical for real-world applications, especially with deep learning models.

Key Results

  • Outperforms Baselines: The proposed method consistently outperforms existing selective learning algorithms (SelectiveNet, DeepGambler, and a confidence-based baseline) across various synthetic datasets and noise levels.
  • Low Selective Risk: The method achieves significantly lower selective risk (SR) compared to baselines, indicating more accurate predictions on the selected (informative) data. For instance, in Table 1, the proposed method achieves SRs around 11-20% compared to baselines in the 20-45% range.
  • High Precision and Recall: The selector achieves high precision (proportion of true informative data among selected data) and high recall (proportion of true informative data selected out of all informative data), demonstrating its ability to accurately identify and utilize informative samples.
  • Stability: The empirical study shows that the proposed method's performance is stable across a range of hyperparameter choices, unlike some baselines that exhibit significant fluctuations.

Limitations and Future Work

  • Assumption of Separability: The theoretical analysis relies on the assumption that informative and uninformative data have distinguishable supports (H-Separable condition). While the paper explores a scenario where domains are similar (SVHN), the theoretical guarantees are strongest when separation is clearer.
  • Continuous Transitions: The current method assumes a binary distinction between informative and uninformative data. Extending it to handle continuous transitions between informative and uninformative data is noted as a non-trivial future work.
  • Computational Complexity: While the heuristic algorithm makes the approach practical, the theoretical analysis relies on empirical minimizers, which can be computationally expensive for complex models.
  • Hyperparameter Tuning: While the method is shown to be more stable than baselines, hyperparameter tuning is still necessary. The paper explores the impact of hyperparameters like $\beta$ and $\lambda$ in ablation studies.

Summary updated

Abstract

Learning and decision making in domains with naturally high noise-to-signal ratios – such as Finance or Public Health – can be challenging and yet extremely important. In this paper, we study a problem of learning on datasets in which a significant proportion of samples does not contain useful information. To analyze this setting, we introduce a noisy generative process with a clear distinction between uninformative/not learnable/purely random data and a structured/informative component. This dichotomy is present both during the training and in the inference phase. We propose a novel approach to learn under these conditions via a loss inspired by the selective learning theory. By minimizing the loss, our method is guaranteed to make a near-optimal decision by simultaneously distinguishing structured data from the non-learnable and making predictions, even in a highly imbalanced setting. We build upon the strength of our theoretical guarantees by describing an iterative algorithm, which jointly optimizes both a predictor and a selector, and evaluate its empirical performance under a variety of conditions.
BibTeX
@inproceedings{zhang2022learning,
  author = {Yikai Zhang and Songzhu Zheng and Pengxiang Wu and Yuriy Nevmyvaka and Chao Chen},
  title = {{Learning to Abstain in the Presence of Uninformative Data}},
  booktitle = {International Conference on Learning Representations},
  year = {2022},
  url = {https://openreview.net/forum?id=i4qKmHdq6y8},
}
Figure