A Unified Framework for Task-Driven Data Quality Management

Tianhao Wang, Yi Zeng, Ming Jin, Ruoxi JiaarXiv 2021 · arXiv

Introduction

The "Garbage In, Garbage Out" principle remains a fundamental challenge in machine learning: a model's performance is strictly bounded by the quality of its training data. Real-world datasets are frequently compromised by random noise, human labeling errors, or sophisticated adversarial "poisoning" and "backdoor" attacks.

Existing Data Quality Management (DQM) approaches typically fall into two categories. ML-oblivious methods focus on database-level consistency without regard for downstream model performance. Conversely, data valuation methods like Shapley values attempt to quantify the "price" of individual data points but are often computationally prohibitive and lack robust theoretical guarantees. To bridge this gap, researchers have introduced DATASIFTER, a task-driven, model-agnostic framework designed to identify and remove "bad" data by learning the characteristics of "good" data for specific machine learning tasks.

Methodology: The DATASIFTER Framework

DATASIFTER treats DQM as an optimization problem, operating in two distinct phases: the Learning Phase and the Selection Phase.

The DATASIFTER framework consists of a Sampling Step to generate performance pairs and a Training Step to build a utility model.
  1. The Learning Phase: Rather than calculating the value of every individual point, DATASIFTER learns a "utility model." It begins by randomly sampling subsets of the training data and training a "proxy" model on each. These models are evaluated against a small, clean validation set to generate (subset, performance score) pairs. These pairs are then used to train a DeepSets neural network, which learns to predict the utility of any given collection of data.
  2. The Selection Phase: Once trained, the utility model is used alongside a linear-time stochastic greedy algorithm to sift through the full dataset. The algorithm selects the subset predicted to yield the highest performance for the specific downstream task.

Theoretical Insights: The Shapley Trap

A significant contribution of this work is the Domination Analysis, a theoretical framework for comparing DQM strategies. The authors demonstrate that popular "linear heuristics" like Shapley values can fail in practical scenarios. Specifically, because Shapley values average a point's contribution across all subset sizes, they tend to over-weight performance on very small datasets. In datasets with high redundancy, this can lead to the selection of redundant points, resulting in poor diversity and suboptimal model performance.

Empirical Results

DATASIFTER was evaluated across diverse domains, including image classification (CIFAR-10, MNIST), text analysis (Enron1 SPAM), and medical imaging (COVID-CT).

Performance comparison of DATASIFTER against baselines across various data quality management tasks, including detection of poisoned data and noisy features.

The framework consistently outperformed state-of-the-art techniques across several DQM tasks:

  • Adversarial Detection: In Trojan attack scenarios on CIFAR-10, DATASIFTER achieved a 90% detection rate while inspecting 51.17% fewer data points than the leading baseline.
  • Noise Identification: It identified 90% of corrupted images by examining 67.25% fewer points than competing methods.
  • Data Summarization and Debiasing: On the PubFig83 dataset, DATASIFTER selected representative subsets that maintained higher accuracy than random or Shapley-based selection. It also effectively filtered data to improve model fairness on the Adult Census and COMPAS datasets.

Conclusion and Future Work

While DATASIFTER represents a significant advance in DQM, it relies on the availability of a small, "gold standard" clean validation set. Additionally, while more efficient than exact Shapley calculations, the initial learning phase still involves training numerous proxy models. Future research may focus on more sophisticated set-function learning models beyond the DeepSets architecture and further reducing the computational overhead for massive datasets.

Summary updated

Abstract

High-quality data is critical to train performant Machine Learning (ML) models, highlighting the importance of Data Quality Management (DQM). Existing DQM schemes often cannot satisfactorily improve ML performance because, by design, they are oblivious to downstream ML tasks. Besides, they cannot handle various data quality issues (especially those caused by adversarial attacks) and have limited applications to only certain types of ML models. Recently, data valuation approaches (e.g., based on the Shapley value) have been leveraged to perform DQM; yet, empirical studies have observed that their performance varies considerably based on the underlying data and training process. In this paper, we propose a task-driven, multi-purpose, model-agnostic DQM framework, DataSifter, which is optimized towards a given downstream ML task, capable of effectively removing data points with various defects, and applicable to diverse models. Specifically, we formulate DQM as an optimization problem and devise a scalable algorithm to solve it. Furthermore, we propose a theoretical framework for comparing the worst-case performance of different DQM strategies. Remarkably, our results show that the popular strategy based on the Shapley value may end up choosing the worst data subset in certain practical scenarios. Our evaluation shows that DataSifter achieves and most often significantly improves the state-of-the-art performance over a wide range of DQM tasks, including backdoor, poison, noisy/mislabel data detection, data summarization, and data debiasing.
BibTeX
@misc{wang2021unified,
  author = {Tianhao Wang and Yi Zeng and Ming Jin and Ruoxi Jia},
  title = {{A Unified Framework for Task-Driven Data Quality Management}},
  year = {2021},
  eprint = {2106.05484},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.05484},
}
Figure