Multiclass Classification

Multiclass classification (which includes binary classification) is a classic supervised machine learning task. We study how to design good multiclass learning algorithms for general losses in various settings including the standard setting, learning with a restricted function class, and learning from noisy labels.

[5] Multiclass Learning from Noisy Labels for Non-decomposable Performance Measures.
Mingyuan Zhang, Shivani Agarwal.
In Proceedings of the 27th International Conference on Artificial Intelligence and Statistics (AISTATS), 2024.
Paper

[3] Learning from Noisy Labels with No Change to the Training Process.
Mingyuan Zhang, Jane Lee, Shivani Agarwal.
In Proceedings of the 38th International Conference on Machine Learning (ICML), 2021.
Paper Link

[2] Bayes Consistency vs. H-Consistency: The Interplay between Surrogate Loss Functions and the Scoring Function Class.
Mingyuan Zhang, Shivani Agarwal.
In Advances in Neural Information Processing Systems (NeurIPS), 2020.
Spotlight paper.
Paper Link

Multiclass and multi-label learning with general losses: What is the right output coding and decoding?
Harish G. Ramaswamy, Mingyuan Zhang, Balaji S. Babu, Shivani Agarwal, Ambuj Tewari, Robert C. Williamson.
In preparation.

Learning from Noisy Labels and Weakly Supervised Learning

Noisy labels can occur due to various reasons such as errors in data collection, human error in annotating data, or mislabeling due to subjective or ambiguous definitions. The main challenge in learning from noisy labels is to design algorithms that can learn good classifiers, despite being given noisy training data. We study how to design good algorithms to learn from noisy labels for multiclass and multi-label classification problems.

We are also interested in a more general learning scheme beyond learning from noisy labels, specifically, weakly supervised learning. Our focus is on learning from missing or partial labels, and transfer learning.

Multi‑Label Learning from Noisy Labels.
Mingyuan Zhang, Shivani Agarwal.
Under review.

[5] Multiclass Learning from Noisy Labels for Non-decomposable Performance Measures.
Mingyuan Zhang, Shivani Agarwal.
In Proceedings of the 27th International Conference on Artificial Intelligence and Statistics (AISTATS), 2024.
Paper

[4] Foreseeing the Benefits of Incidental Supervision.
Hangfeng He, Mingyuan Zhang, Qiang Ning, Dan Roth.
In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021.
Oral paper.
Paper Link

[3] Learning from Noisy Labels with No Change to the Training Process.
Mingyuan Zhang, Jane Lee, Shivani Agarwal.
In Proceedings of the 38th International Conference on Machine Learning (ICML), 2021.
Paper Link

Multi-Label Classification and Label Ranking

In multi-label classification, each instance can be associated with multiple labels (or tags) simultaneously. A good example is image tagging, where several tags can be active in the same image. We study how to design good multi-label learning algorithms for general multi-label losses (including Hamming Loss, Precision, Recall and F-measure) in various settings including the standard setting, learning from noisy labels, and learning from partial/missing labels.

Label ranking is a prediction task where the goal is to map instances to rankings over a finite set of predefined labels (or tags). We study the design of effective label ranking learning algorithms for a range of label ranking losses. These include Pairwise Loss, Discounted Cumulative Gain, and Precision, and our study applies to various settings such as the standard and online settings.

Multi‑Label Learning from Noisy Labels.
Mingyuan Zhang, Shivani Agarwal.
Under review.

On the Minimax Regret in Online Ranking with Top-k Feedback.
Mingyuan Zhang, Ambuj Tewari.
Preprint.
Paper Link

[1] Convex Calibrated Surrogates for the Multi-Label F-Measure.
Mingyuan Zhang, Harish G. Ramaswamy, Shivani Agarwal.
In Proceedings of the 37th International Conference on Machine Learning (ICML), 2020.
Paper Link

Multiclass and multi-label learning with general losses: What is the right output coding and decoding?
Harish G. Ramaswamy, Mingyuan Zhang, Balaji S. Babu, Shivani Agarwal, Ambuj Tewari, Robert C. Williamson.
In preparation.

Non‑decomposable Performance Measures

Unlike 0-1 loss (accuracy) or cost-sensitive losses, non-decomposable performance measures cannot be expressed as the expectation or sum of a loss on individual examples; such performance measures are defined by general (usually nonlinear) functions of the confusion matrix of a classifier. Examples include Micro F1 score, Jaccard measure, H-mean, G-mean, Q-mean, AUC-ROC, and AUC-PR. We study how to design learning algorithms to optimize for non-decomposable performance measures.

[5] Multiclass Learning from Noisy Labels for Non-decomposable Performance Measures.
Mingyuan Zhang, Shivani Agarwal.
In Proceedings of the 27th International Conference on Artificial Intelligence and Statistics (AISTATS), 2024.
Paper