BACKGROUND: Single-cell RNA sequencing (scRNA-seq) enables the identification of cell types within complex biological systems, yet accurately classifying both known and novel cell types remains a significant challenge. Supervised learning methods perform well when all cell types are labeled in the training data, but struggle with unseen cell types, while rejection-based approaches can mitigate misclassification but fail to leverage unlabeled data for learning. Deep learning-based methods, such as MARS, offer promising solutions but often suffer from poor generalization to novel cell populations. RESULTS: We propose Semi-supervised learning for Robust Novel Cell-type identification (SRNC), a novel semi-supervised framework that enhances classification accuracy while effectively identifying unknown cell types. By integrating self-supervised feature learning with semi-supervised classification, SRNC leverages both labeled and unlabeled data to improve generalization. Evaluated across six benchmark scRNA-seq datasets, SRNC consistently outperforms state-of-the-art methods, achieving higher ARI, F1-score, and precision than both the rejection-based approach and deep-learning-based MARS. Moreover, SRNC demonstrates robustness across datasets from different laboratories and excels in imbalanced classification scenarios, accurately identifying rare cell populations that other methods often misclassify. CONCLUSIONS: Our results demonstrate that SRNC is a powerful and adaptable tool for cell-type classification in scRNA-seq analysis. By leveraging semi-supervised learning, SRNC effectively identifies both known and novel cell types, surpassing competing methods in multiple performance metrics. Its ability to generalize across datasets and handle class imbalances makes it a valuable approach for discovering new cell types, advancing precision medicine, and improving our understanding of cellular heterogeneity.

Abstract In this paper, we present a formal framework to (1) aggregate probabilistic ensemble members into either a representative classifier or a credal classifier, and (2) perform various decision tasks based on this uncertainty quantification. We first elaborate on the aggregation problem under a class of distances between distributions. We then propose generic methods to robustify uncertainty quantification and decisions, based on the obtained ensemble and representative probability. To facilitate the scalability of the proposed framework, for all the problems and applications covered, we elaborate on their computational complexities from the theoretical aspects and leverage theoretical results to derive efficient algorithmic solutions. Finally, relevant sets of experiments are conducted to assess the usefulness of the proposed framework in uncertainty sampling, classification with a reject option, and set-valued prediction-making.

Multi-dimensional classification (MDC) can be employed in a range of applications where one needs to predict multiple class variables for each given instance. Many existing MDC methods suffer from at least one of inaccuracy, scalability, limited use to certain types of data, hardness of interpretation or lack of probabilistic (uncertainty) estimations. This paper is an attempt to address all these disadvantages simultaneously. We propose a formal framework for probabilistic MDC in which learning an optimal multi-dimensional classifier can be decomposed, without loss of generality, into learning a set of (smaller) single-variable multi-class probabilistic classifiers and a directed acyclic graph. Current and future developments of both probabilistic classification and graphical model learning can directly enhance our framework, which is flexible and provably optimal. A collection of experiments is conducted to highlight the usefulness of this MDC framework.

Abstract Various strategies for active learning have been proposed in the machine learning literature. In uncertainty sampling, which is among the most popular approaches, the active learner sequentially queries the label of those instances for which its current prediction is maximally uncertain. The predictions as well as the measures used to quantify the degree of uncertainty, such as entropy, are traditionally of a probabilistic nature. Yet, alternative approaches to capturing uncertainty in machine learning, alongside with corresponding uncertainty measures, have been proposed in recent years. In particular, some of these measures seek to distinguish different sources and to separate different types of uncertainty, such as the reducible (epistemic) and the irreducible (aleatoric) part of the total uncertainty in a prediction. The goal of this paper is to elaborate on the usefulness of such measures for uncertainty sampling, and to compare their performance in active learning. To this end, we instantiate uncertainty sampling with different measures, analyze the properties of the sampling strategies thus obtained, and compare them in an experimental study.

In contrast to conventional (single-label) classification, the setting of multilabel classification (MLC) allows an instance to belong to several classes simultaneously. Thus, instead of selecting a single class label, predictions take the form of a subset of all labels. In this paper, we study an extension of the setting of MLC, in which the learner is allowed to partially abstain from a prediction, that is, to deliver predictions on some but not necessarily all class labels. This option is useful in cases of uncertainty, where the learner does not feel confident enough on the entire label set. Adopting a decision-theoretic perspective, we propose a formal framework of MLC with partial abstention, which builds on two main building blocks: First, the extension of underlying MLC loss functions so as to accommodate abstention in a proper way, and second the problem of optimal prediction, that is, finding the Bayes-optimal prediction minimizing this generalized loss in expectation. It is well known that different (generalized) loss functions may have different risk-minimizing predictions, and finding the Bayes predictor typically comes down to solving a computationally complexity optimization problem. In the most general case, given a prediction of the (conditional) joint distribution of possible labelings, the minimizer of the expected loss needs to be found over a number of candidates which is exponential in the number of class labels. We elaborate on properties of risk minimizers for several commonly used (generalized) MLC loss functions, show them to have a specific structure, and leverage this structure to devise efficient methods for computing Bayes predictors. Experimentally, we show MLC with partial abstention to be effective in the sense of reducing loss when being allowed to abstain.

In contrast to conventional (single-label) classification, the setting of multilabel classification (MLC) allows an instance to belong to several classes simultaneously. Thus, instead of selecting a single class label, predictions take the form of a subset of all labels. In this paper, we study an extension of the setting of MLC, in which the learner is allowed to partially abstain from a prediction, that is, to deliver predictions on some but not necessarily all class labels. We propose a formalization of MLC with abstention in terms of a generalized loss minimization problem and present first results for the case of the Hamming loss, rank loss, and F-measure, both theoretical and experimental.

We propose a method for reliable prediction in multi-class classification, where reliability refers to the possibility of partial abstention in cases of uncertainty. More specifically, we allow for predictions in the form of preorder relations on the set of classes, thereby generalizing the idea of set-valued predictions. Our approach relies on combining learning by pairwise comparison with a recent proposal for modeling uncertainty in classification, in which a distinction is made between reducible (a.k.a. epistemic) uncertainty caused by a lack of information and irreducible (a.k.a. aleatoric) uncertainty due to intrinsic randomness. The problem of combining uncertain pairwise predictions into a most plausible preorder is then formalized as an integer programming problem. Experimentally, we show that our method is able to appropriately balance reliability and precision of predictions.