Multi-Label Online Continual Learning (MOCL) requires models to learn continuously from endless multi-label data streams, facing complex challenges including persistent catastrophic forgetting, potential missing labels, and uncontrollable imbalanced class distributions. While existing MOCL methods attempt to address these challenges through various techniques, \textit{they all overlook label-specific region identifying and feature learning} - a fundamental solution rooted in multi-label learning but challenging to achieve in the online setting with incremental and partial supervision. To this end, we first leverage the inherent structural information of input data to evaluate and verify the innate localization capability of different pre-trained models. Then, we propose CUTER (CUT-out-and-Experience-Replay), a simple yet versatile strategy that provides fine-grained supervision signals by further identifying, strengthening and cutting out label-specific regions for efficient experience replay. It not only enables models to simultaneously address catastrophic forgetting, missing labels, and class imbalance challenges, but also serves as an orthogonal solution that seamlessly integrates with existing approaches. Extensive experiments on multiple multi-label image benchmarks demonstrate the superiority of our proposed method. The code is available at \href{https://github.com/wxr99/Cut-Replay}{https://github.com/wxr99/Cut-Replay}
We introduce a framework for robust uncertainty quantification in situations where labeled training data are corrupted, through noisy or missing labels. We build on conformal prediction, a statistical tool for generating prediction sets that cover the test label with a pre-specified probability. The validity of conformal prediction, however, holds under the i.i.d assumption, which does not hold in our setting due to the corruptions in the data. To account for this distribution shift, the privileged conformal prediction (PCP) method proposed leveraging privileged information (PI) -- additional features available only during training -- to re-weight the data distribution, yielding valid prediction sets under the assumption that the weights are accurate. In this work, we analyze the robustness of PCP to inaccuracies in the weights. Our analysis indicates that PCP can still yield valid uncertainty estimates even when the weights are poorly estimated. Furthermore, we introduce uncertain imputation (UI), a new conformal method that does not rely on weight estimation. Instead, we impute corrupted labels in a way that preserves their uncertainty. Our approach is supported by theoretical guarantees and validated empirically on both synthetic and real benchmarks. Finally, we show that these techniques can be integrated into a triply robust framework, ensuring statistically valid predictions as long as at least one underlying method is valid.
Extreme Classification (XC) aims to map a query to the most relevant documents from a very large document set. XC algorithms used in real-world applications learn this mapping from datasets curated from implicit feedback, such as user clicks. However, these datasets inevitably suffer from missing labels. In this work, we observe that systematic missing labels lead to missing knowledge, which is critical for accurately modelling relevance between queries and documents. We formally show that this absence of knowledge cannot be recovered using existing methods such as propensity weighting and data imputation strategies that solely rely on the training dataset. While LLMs provide an attractive solution to augment the missing knowledge, leveraging them in applications with low latency requirements and large document sets is challenging. To incorporate missing knowledge at scale, we propose SKIM (Scalable Knowledge Infusion for Missing Labels), an algorithm that leverages a combination of small LM and abundant unstructured meta-data to effectively mitigate the missing label problem. We show the efficacy of our method on large-scale public datasets through exhaustive unbiased evaluation ranging from human annotations to simulations inspired from industrial settings. SKIM outperforms existing methods on Recall@100 by more than 10 absolute points. Additionally, SKIM scales to proprietary query-ad retrieval datasets containing 10 million documents, outperforming contemporary methods by 12% in offline evaluation and increased ad click-yield by 1.23% in an online A/B test conducted on a popular search engine. We release our code, prompts, trained XC models and finetuned SLMs at: https://github.com/bicycleman15/skim
This technical report describes the CP-JKU team's submission for Task 4 Sound Event Detection with Heterogeneous Training Datasets and Potentially Missing Labels of the DCASE 24 Challenge. We fine-tune three large Audio Spectrogram Transformers, PaSST, BEATs, and ATST, on the joint DESED and MAESTRO datasets in a two-stage training procedure. The first stage closely matches the baseline system setup and trains a CRNN model while keeping the large pre-trained transformer model frozen. In the second stage, both CRNN and transformer are fine-tuned using heavily weighted self-supervised losses. After the second stage, we compute strong pseudo-labels for all audio clips in the training set using an ensemble of all three fine-tuned transformers. Then, in a second iteration, we repeat the two-stage training process and include a distillation loss based on the pseudo-labels, boosting single-model performance substantially. Additionally, we pre-train PaSST and ATST on the subset of AudioSet that comes with strong temporal labels, before fine-tuning them on the Task 4 datasets.
Recently, large-scale visual language pre-trained (VLP) models have demonstrated impressive performance across various downstream tasks. Motivated by these advancements, pioneering efforts have emerged in multi-label image recognition with missing labels, leveraging VLP prompt-tuning technology. However, they usually cannot match text and vision features well, due to complicated semantics gaps and missing labels in a multi-label image. To tackle this challenge, we propose $\textbf{T}$ext-$\textbf{R}$egion $\textbf{M}$atching for optimizing $\textbf{M}$ulti-$\textbf{L}$abel prompt tuning, namely TRM-ML, a novel method for enhancing meaningful cross-modal matching. Compared to existing methods, we advocate exploring the information of category-aware regions rather than the entire image or pixels, which contributes to bridging the semantic gap between textual and visual representations in a one-to-one matching manner. Concurrently, we further introduce multimodal contrastive learning to narrow the semantic gap between textual and visual modalities and establish intra-class and inter-class relationships. Additionally, to deal with missing labels, we propose a multimodal category prototype that leverages intra- and inter-category semantic relationships to estimate unknown labels, facilitating pseudo-label generation. Extensive experiments on the MS-COCO, PASCAL VOC, Visual Genome, NUS-WIDE, and CUB-200-211 benchmark datasets demonstrate that our proposed framework outperforms the state-of-the-art methods by a significant margin. Our code is available here: https://github.com/yu-gi-oh-leilei/TRM-ML.
Cross-silo federated learning (FL) enables decentralized organizations to collaboratively train models while preserving data privacy and has made significant progress in medical image classification. One common assumption is task homogeneity where each client has access to all classes during training. However, in clinical practice, given a multi-label classification task, constrained by the level of medical knowledge and the prevalence of diseases, each institution may diagnose only partial categories, resulting in task heterogeneity. How to pursue effective multi-label medical image classification under task heterogeneity is under-explored. In this paper, we first formulate such a realistic label missing setting in the multi-label FL domain and propose a two-stage method FedMLP to combat class missing from two aspects: pseudo label tagging and global knowledge learning. The former utilizes a warmed-up model to generate class prototypes and select samples with high confidence to supplement missing labels, while the latter uses a global model as a teacher for consistency regularization to prevent forgetting missing class knowledge. Experiments on two publicly-available medical datasets validate the superiority of FedMLP against the state-of-the-art both federated semi-supervised and noisy label learning approaches under task heterogeneity. Code is available at https://github.com/szbonaldo/FedMLP.
Keyword mnemonics are memorable explanations that link new terms to simpler keywords. Prior work generates mnemonics for students, but they do not train models using mnemonics students prefer and aid learning. We build SMART, a mnemonic generator trained on feedback from real students learning new terms. To train SMART, we first fine-tune LLaMA-2 on a curated set of user-written mnemonics. We then use LLM alignment to enhance SMART: we deploy mnemonics generated by SMART in a flashcard app to find preferences on mnemonics students favor. We gather 2684 preferences from 45 students across two types: expressed (inferred from ratings) and observed (inferred from student learning), yielding three key findings. First, expressed and observed preferences disagree; what students think is helpful does not always capture what is truly helpful. Second, Bayesian models can synthesize complementary data from multiple preference types into a single effectiveness signal. SMART is tuned via Direct Preference Optimization on this signal, which resolves ties and missing labels in the typical method of pairwise comparisons, augmenting data for LLM output quality gains. Third, mnemonic experts assess SMART as matching GPT-4 at much lower deployment costs, showing the utility of capturing diverse student feedback to align LLMs in education.
Multi-label learning (MLL) requires comprehensive multi-semantic annotations that is hard to fully obtain, thus often resulting in missing labels scenarios. In this paper, we investigate Single Positive Multi-label Learning (SPML), where each image is associated with merely one positive label. Existing SPML methods only focus on designing losses using mechanisms such as hard pseudo-labeling and robust losses, mostly leading to unacceptable false negatives. To address this issue, we first propose a generalized loss framework based on expected risk minimization to provide soft pseudo labels, and point out that the former losses can be seamlessly converted into our framework. In particular, we design a novel robust loss based on our framework, which enjoys flexible coordination between false positives and false negatives, and can additionally deal with the imbalance between positive and negative samples. Extensive experiments show that our approach can significantly improve SPML performance and outperform the vast majority of state-of-the-art methods on all the four benchmarks.
Advances in mobile and wearable technologies have enabled the potential to passively monitor a person's mental, behavioral, and affective health. These approaches typically rely on longitudinal collection of self-reported outcomes, e.g., depression, stress, and anxiety, to train machine learning (ML) models. However, the need to continuously self-report adds a significant burden on the participants, often resulting in attrition, missing labels, or insincere responses. In this work, we introduce the Scale Scores Simulation using Mental Models (SeSaMe) framework to alleviate participants' burden in digital mental health studies. By leveraging pre-trained large language models (LLMs), SeSaMe enables the simulation of participants' responses on psychological scales. In SeSaMe, researchers can prompt LLMs with information on participants' internal behavioral dispositions, enabling LLMs to construct mental models of participants to simulate their responses on psychological scales. We demonstrate an application of SeSaMe, where we use GPT-4 to simulate responses on one scale using responses from another as behavioral information. We also evaluate the alignment between human and SeSaMe-simulated responses to psychological scales. Then, we present experiments to inspect the utility of SeSaMe-simulated responses as ground truth in training ML models by replicating established depression and anxiety screening tasks from a previous study. Our results indicate SeSaMe to be a promising approach, but its alignment may vary across scales and specific prediction objectives. We also observed that model performance with simulated data was on par with using the real data for training in most evaluation scenarios. We conclude by discussing the potential implications of SeSaMe in addressing some challenges researchers face with ground-truth collection in passive sensing studies.
Missing data is a common problem in practical data science settings. Various imputation methods have been developed to deal with missing data. However, even though the labels are available in the training data in many situations, the common practice of imputation usually only relies on the input and ignores the label. We propose Classification Based on MissForest Imputation (CBMI), a classification strategy that initializes the predicted test label with missing values and stacks the label with the input for imputation, allowing the label and the input to be imputed simultaneously. In addition, we propose the imputation using labels (IUL) algorithm, an imputation strategy that stacks the label into the input and illustrates how it can significantly improve the imputation quality. Experiments show that CBMI has classification accuracy when the test set contains missing data, especially for imbalanced data and categorical data. Moreover, for both the regression and classification, IUL consistently shows significantly better results than imputation based on only the input data.
Extreme multi-label classification (XMLC) is the task of selecting a small subset of relevant labels from a very large set of possible labels. As such, it is characterized by long-tail labels, i.e., most labels have very few positive instances. With standard performance measures such as precision@k, a classifier can ignore tail labels and still report good performance. However, it is often argued that correct predictions in the tail are more "interesting" or "rewarding," but the community has not yet settled on a metric capturing this intuitive concept. The existing propensity-scored metrics fall short on this goal by confounding the problems of long-tail and missing labels. In this paper, we analyze generalized metrics budgeted "at k" as an alternative solution. To tackle the challenging problem of optimizing these metrics, we formulate it in the expected test utility (ETU) framework, which aims to optimize the expected performance on a fixed test set. We derive optimal prediction rules and construct computationally efficient approximations with provable regret guarantees and robustness against model misspecification. Our algorithm, based on block coordinate ascent, scales effortlessly to XMLC problems and obtains promising results in terms of long-tail performance.
Developing generalizable ML-based solutions for disparate learning problems in network security is highly desired. However, despite a rich history of applying ML to network security, most existing solutions lack generalizability. This lack of progress can be attributed to an overreliance on supervised learning techniques and the associated challenges of curating well-specified labeled training data. This paper addresses a fundamental gap by introducing a novel transformer-based network foundation model, netFound. We employ self-supervised learning techniques on abundant, unlabeled network telemetry data for pre-training. This pretrained model can subsequently be fine-tuned to create generalizable learning artifacts for disparate learning tasks, even when using commonly available but challenging labeled datasets that are sparse, noisy, and skewed. To realize this goal, netFound leverages various domain-specific attributes and constraints unique to network data (packet traces) by developing multi-modal embeddings, protocol-aware tokenization, data-driven token composition, and hierarchical transformers. Our results demonstrate that netFound's domain-specific design choices ensure that it (1) effectively captures the hidden networking context in production settings, (2) outperforms four different SOTA methods on five different learning tasks, and (3) is robust to both noisy labels and learning shortcuts -- critical for developing generalizable ML models in practical settings.
Available works addressing multi-label classification in a data stream environment focus on proposing accurate models; however, these models often exhibit inefficiency and cannot balance effectiveness and efficiency. In this work, we propose a neural network-based approach that tackles this issue and is suitable for high-dimensional multi-label classification. Our model uses a selective concept drift adaptation mechanism that makes it suitable for a non-stationary environment. Additionally, we adapt our model to an environment with missing labels using a simple yet effective imputation strategy and demonstrate that it outperforms a vast majority of the state-of-the-art supervised models. To achieve our purposes, we introduce a weighted binary relevance-based approach named ML-BELS using the Broad Ensemble Learning System (BELS) as its base classifier. Instead of a chain of stacked classifiers, our model employs independent weighted ensembles, with the weights generated by the predictions of a BELS classifier. We show that using the weighting strategy on datasets with low label cardinality negatively impacts the accuracy of the model; with this in mind, we use the label cardinality as a trigger for applying the weights. We present an extensive assessment of our model using 11 state-of-the-art baselines, five synthetics, and 13 real-world datasets, all with different characteristics. Our results demonstrate that the proposed approach ML-BELS is successful in balancing effectiveness and efficiency, and is robust to missing labels and concept drift.
While reliable data-driven decision-making hinges on high-quality labeled data, the acquisition of quality labels often involves laborious human annotations or slow and expensive scientific measurements. Machine learning is becoming an appealing alternative as sophisticated predictive techniques are being used to quickly and cheaply produce large amounts of predicted labels; e.g., predicted protein structures are used to supplement experimentally derived structures, predictions of socioeconomic indicators from satellite imagery are used to supplement accurate survey data, and so on. Since predictions are imperfect and potentially biased, this practice brings into question the validity of downstream inferences. We introduce cross-prediction: a method for valid inference powered by machine learning. With a small labeled dataset and a large unlabeled dataset, cross-prediction imputes the missing labels via machine learning and applies a form of debiasing to remedy the prediction inaccuracies. The resulting inferences achieve the desired error probability and are more powerful than those that only leverage the labeled data. Closely related is the recent proposal of prediction-powered inference, which assumes that a good pre-trained model is already available. We show that cross-prediction is consistently more powerful than an adaptation of prediction-powered inference in which a fraction of the labeled data is split off and used to train the model. Finally, we observe that cross-prediction gives more stable conclusions than its competitors; its confidence intervals typically have significantly lower variability.
Federated learning (FL) has been introduced to the healthcare domain as a decentralized learning paradigm that allows multiple parties to train a model collaboratively without privacy leakage. However, most previous studies have assumed that every client holds an identical label set. In reality, medical specialists tend to annotate only diseases within their area of expertise or interest. This implies that label sets in each client can be different and even disjoint. In this paper, we propose the framework FedLSM to solve the problem of Label Set Mismatch. FedLSM adopts different training strategies on data with different uncertainty levels to efficiently utilize unlabeled or partially labeled data as well as class-wise adaptive aggregation in the classification layer to avoid inaccurate aggregation when clients have missing labels. We evaluated FedLSM on two public real-world medical image datasets, including chest X-ray (CXR) diagnosis with 112,120 CXR images and skin lesion diagnosis with 10,015 dermoscopy images, and showed that it significantly outperformed other state-of-the-art FL algorithms. The code can be found at https://github.com/dzp2095/FedLSM.
In recent years, multi-view multi-label learning has aroused extensive research enthusiasm. However, multi-view multi-label data in the real world is commonly incomplete due to the uncertain factors of data collection and manual annotation, which means that not only multi-view features are often missing, and label completeness is also difficult to be satisfied. To deal with the double incomplete multi-view multi-label classification problem, we propose a deep instance-level contrastive network, namely DICNet. Different from conventional methods, our DICNet focuses on leveraging deep neural network to exploit the high-level semantic representations of samples rather than shallow-level features. First, we utilize the stacked autoencoders to build an end-to-end multi-view feature extraction framework to learn the view-specific representations of samples. Furthermore, in order to improve the consensus representation ability, we introduce an incomplete instance-level contrastive learning scheme to guide the encoders to better extract the consensus information of multiple views and use a multi-view weighted fusion module to enhance the discrimination of semantic features. Overall, our DICNet is adept in capturing consistent discriminative representations of multi-view multi-label data and avoiding the negative effects of missing views and missing labels. Extensive experiments performed on five datasets validate that our method outperforms other state-of-the-art methods.
Although originally developed to evaluate sets of items, recall is often used to evaluate rankings of items, including those produced by recommender, retrieval, and other machine learning systems. The application of recall without a formal evaluative motivation has led to criticism of recall as a vague or inappropriate measure. In light of this debate, we reflect on the measurement of recall in rankings from a formal perspective. Our analysis is composed of three tenets: recall, robustness, and lexicographic evaluation. First, we formally define `recall-orientation' as the sensitivity of a metric to a user interested in finding every relevant item. Second, we analyze recall-orientation from the perspective of robustness with respect to possible content consumers and providers, connecting recall to recent conversations about fair ranking. Finally, we extend this conceptual and theoretical treatment of recall by developing a practical preference-based evaluation method based on lexicographic comparison. Through extensive empirical analysis across three recommendation tasks and 17 information retrieval tasks, we establish that our new evaluation method, lexirecall, has convergent validity (i.e., it is correlated with existing recall metrics) and exhibits substantially higher sensitivity in terms of discriminative power and stability in the presence of missing labels. Our conceptual, theoretical, and empirical analysis substantially deepens our understanding of recall and motivates its adoption through connections to robustness and fairness.
We develop and analyze a principled approach to kernel ridge regression under covariate shift. The goal is to learn a regression function with small mean squared error over a target distribution, based on unlabeled data from there and labeled data that may have a different feature distribution. We propose to split the labeled data into two subsets, and conduct kernel ridge regression on them separately to obtain a collection of candidate models and an imputation model. We use the latter to fill the missing labels and then select the best candidate accordingly. Our non-asymptotic excess risk bounds demonstrate that our estimator adapts effectively to both the structure of the target distribution and the covariate shift. This adaptation is quantified through a notion of effective sample size that reflects the value of labeled source data for the target regression task. Our estimator achieves the minimax optimal error rate up to a polylogarithmic factor, and we find that using pseudo-labels for model selection does not significantly hinder performance.
Contrastive learning (CL) has shown impressive advances in image representation learning in whichever supervised multi-class classification or unsupervised learning. However, these CL methods fail to be directly adapted to multi-label image classification due to the difficulty in defining the positive and negative instances to contrast a given anchor image in multi-label scenario, let the label missing one alone, implying that borrowing a commonly-used way from contrastive multi-class learning to define them will incur a lot of false negative instances unfavorable for learning. In this paper, with the introduction of a label correction mechanism to identify missing labels, we first elegantly generate positives and negatives for individual semantic labels of an anchor image, then define a unique contrastive loss for multi-label image classification with missing labels (CLML), the loss is able to accurately bring images close to their true positive images and false negative images, far away from their true negative images. Different from existing multi-label CL losses, CLML also preserves low-rank global and local label dependencies in the latent representation space where such dependencies have been shown to be helpful in dealing with missing labels. To the best of our knowledge, this is the first general multi-label CL loss in the missing-label scenario and thus can seamlessly be paired with those losses of any existing multi-label learning methods just via a single hyperparameter. The proposed strategy has been shown to improve the classification performance of the Resnet101 model by margins of 1.2%, 1.6%, and 1.3% respectively on three standard datasets, MSCOCO, VOC, and NUS-WIDE. Code is available at https://github.com/chuangua/ContrastiveLossMLML.
Albeit the Dice loss is one of the dominant loss functions in medical image segmentation, most research omits a closer look at its derivative, i.e. the real motor of the optimization when using gradient descent. In this paper, we highlight the peculiar action of the Dice loss in the presence of missing or empty labels. First, we formulate a theoretical basis that gives a general description of the Dice loss and its derivative. It turns out that the choice of the reduction dimensions $\Phi$ and the smoothing term $\epsilon$ is non-trivial and greatly influences its behavior. We find and propose heuristic combinations of $\Phi$ and $\epsilon$ that work in a segmentation setting with either missing or empty labels. Second, we empirically validate these findings in a binary and multiclass segmentation setting using two publicly available datasets. We confirm that the choice of $\Phi$ and $\epsilon$ is indeed pivotal. With $\Phi$ chosen such that the reductions happen over a single batch (and class) element and with a negligible $\epsilon$, the Dice loss deals with missing labels naturally and performs similarly compared to recent adaptations specific for missing labels. With $\Phi$ chosen such that the reductions happen over multiple batch elements or with a heuristic value for $\epsilon$, the Dice loss handles empty labels correctly. We believe that this work highlights some essential perspectives and hope that it encourages researchers to better describe their exact implementation of the Dice loss in future work.