We study the statistical properties of learning to defer (L2D) to multiple experts. In particular, we address the open problems of deriving a consistent surrogate loss, confidence calibration, and principled ensembling of experts. Firstly, we derive two consistent surrogates -- one based on a softmax parameterization, the other on a one-vs-all (OvA) parameterization -- that are analogous to the single expert losses proposed by Mozannar and Sontag (2020) and Verma and Nalisnick (2022), respectively. We then study the frameworks' ability to estimate P( m_j = y | x ), the probability that the jth expert will correctly predict the label for x. Theory shows the softmax-based loss causes mis-calibration to propagate between the estimates while the OvA-based loss does not (though in practice, we find there are trade offs). Lastly, we propose a conformal inference technique that chooses a subset of experts to query when the system defers. We perform empirical validation on tasks for galaxy, skin lesion, and hate speech classification.

This paper studies algorithmic fairness when the protected attribute is location. To handle protected attributes that are continuous, such as age or income, the standard approach is to discretize the domain into predefined groups, and compare algorithmic outcomes across groups. However, applying this idea to location raises concerns of gerrymandering and may introduce statistical bias. Prior work addresses these concerns but only for regularly spaced locations, while raising other issues, most notably its inability to discern regions that are likely to exhibit spatial unfairness. Similar to established notions of algorithmic fairness, we define spatial fairness as the statistical independence of outcomes from location. This translates into requiring that for each region of space, the distribution of outcomes is identical inside and outside the region. To allow for localized discrepancies in the distribution of outcomes, we compare how well two competing hypotheses explain the observed outcomes. The null hypothesis assumes spatial fairness, while the alternate allows different distributions inside and outside regions. Their goodness of fit is then assessed by a likelihood ratio test. If there is no significant difference in how well the two hypotheses explain the observed outcomes, we conclude that the algorithm is spatially fair.

Accurately estimating the shape of objects in dense clutters makes important contribution to robotic packing, because the optimal object arrangement requires the robot planner to acquire shape information of all existed objects. However, the objects for packing are usually piled in dense clutters with severe occlusion, and the object shape varies significantly across different instances for the same category. They respectively cause large object segmentation errors and inaccurate shape recovery on unseen instances, which both degrade the performance of shape estimation during deployment. In this paper, we propose a category-level shape estimation method for densely cluttered objects. Our framework partitions each object in the clutter via the multi-view visual information fusion to achieve high segmentation accuracy, and the instance shape is recovered by deforming the category templates with diverse geometric transformations to obtain strengthened generalization ability. Specifically, we first collect the multi-view RGB-D images of the object clutters for point cloud reconstruction. Then we fuse the feature maps representing the visual information of multi-view RGB images and the pixel affinity learned from the clutter point cloud, where the acquired instance segmentation masks of multi-view RGB images are projected to partition the clutter point cloud. Finally, the instance geometry information is obtained from the partially observed instance point cloud and the corresponding category template, and the deformation parameters regarding the template are predicted for shape estimation. Experiments in the simulated environment and real world show that our method achieves high shape estimation accuracy for densely cluttered everyday objects with various shapes.

We study the problem of online binary classification where strategic agents can manipulate their observable features in predefined ways, modeled by a manipulation graph, in order to receive a positive classification. We show this setting differs in fundamental ways from non-strategic online classification. For instance, whereas in the non-strategic case, a mistake bound of $\ln|H|$ is achievable via the halving algorithm when the target function belongs to a known class $H$, we show that no deterministic algorithm can achieve a mistake bound $o(\Delta)$ in the strategic setting, where $\Delta$ is the maximum degree of the manipulation graph (even when $|H|=O(\Delta)$). We obtain an algorithm achieving mistake bound $O(\Delta\ln|H|)$. We also extend this to the agnostic setting and obtain an algorithm with a $\Delta$ multiplicative regret, and we show no deterministic algorithm can achieve $o(\Delta)$ multiplicative regret. Next, we study two randomized models based on whether the random choices are made before or after agents respond, and show they exhibit fundamental differences. In the first model, at each round the learner deterministically chooses a probability distribution over classifiers inducing expected values on each vertex (probabilities of being classified as positive), which the strategic agents respond to. We show that any learner in this model has to suffer linear regret. On the other hand, in the second model, while the adversary who selects the next agent must respond to the learner's probability distribution over classifiers, the agent then responds to the actual hypothesis classifier drawn from this distribution. Surprisingly, we show this model is more advantageous to the learner, and we design randomized algorithms that achieve sublinear regret bounds against both oblivious and adaptive adversaries.

In autonomous driving, detection of abnormal driving behaviors is essential to ensure the safety of vehicle controllers. Prior works in vehicle anomaly detection have shown that modeling interactions between agents improves detection accuracy, but certain abnormal behaviors where structured road information is paramount are poorly identified, such as wrong-way and off-road driving. We propose a novel unsupervised framework for highway anomaly detection named Structural Attention-Based Recurrent VAE (SABeR-VAE), which explicitly uses the structure of the environment to aid anomaly identification. Specifically, we use a vehicle self-attention module to learn the relations among vehicles on a road, and a separate lane-vehicle attention module to model the importance of permissible lanes to aid in trajectory prediction. Conditioned on the attention modules' outputs, a recurrent encoder-decoder architecture with a stochastic Koopman operator-propagated latent space predicts the next states of vehicles. Our model is trained end-to-end to minimize prediction loss on normal vehicle behaviors, and is deployed to detect anomalies in (ab)normal scenarios. By combining the heterogeneous vehicle and lane information, SABeR-VAE and its deterministic variant, SABeR-AE, improve abnormal AUPR by 18% and 25% respectively on the simulated MAAD highway dataset over STGAE-KDE. Furthermore, we show that the learned Koopman operator in SABeR-VAE enforces interpretable structure in the variational latent space. The results of our method indeed show that modeling environmental factors is essential to detecting a diverse set of anomalies in deployment. For code implementation, please visit https://sites.google.com/illinois.edu/saber-vae.

The US federal government spends more than a trillion dollars per year on health care, largely provided by private third parties and reimbursed by the government. A major concern in this system is overbilling, waste and fraud by providers, who face incentives to misreport on their claims in order to receive higher payments. In this paper, we develop novel machine learning tools to identify providers that overbill Medicare, the US federal health insurance program for elderly adults and the disabled. Using large-scale Medicare claims data, we identify patterns consistent with fraud or overbilling among inpatient hospitalizations. Our proposed approach for Medicare fraud detection is fully unsupervised, not relying on any labeled training data, and is explainable to end users, providing reasoning and interpretable insights into the potentially suspicious behavior of the flagged providers. Data from the Department of Justice on providers facing anti-fraud lawsuits and several case studies validate our approach and findings both quantitatively and qualitatively.

The complex heterogeneity of brain tumours is increasingly recognized to demand data of magnitudes and richness only fully-inclusive, large-scale collections drawn from routine clinical care could plausibly offer. This is a task contemporary machine learning could facilitate, especially in neuroimaging, but its ability to deal with incomplete data common in real world clinical practice remains unknown. Here we apply state-of-the-art methods to large scale, multi-site MRI data to quantify the comparative fidelity of automated tumour segmentation models replicating the various levels of sequence availability observed in the clinical reality. We compare deep learning (nnU-Net-derived) segmentation models with all possible combinations of T1, contrast-enhanced T1, T2, and FLAIR sequences, trained and validated with five-fold cross-validation on the 2021 BraTS-RSNA glioma population of 1251 patients, with further testing on a real-world 50 patient sample diverse in not only MRI scanner and field strength, but a random selection of pre- and post-operative imaging also. Models trained on incomplete imaging data segmented lesions well, often equivalently to those trained on complete data, exhibiting Dice coefficients of 0.907 (single sequence) to 0.945 (full datasets) for whole tumours, and 0.701 (single sequence) to 0.891 (full datasets) for component tissue types. Incomplete data segmentation models could accurately detect enhancing tumour in the absence of contrast imaging, quantifying its volume with an R2 between 0.95-0.97, and were invariant to lesion morphometry. Deep learning segmentation models characterize tumours well when missing data and can even detect enhancing tissue without the use of contrast. This suggests translation to clinical practice, where incomplete data is common, may be easier than hitherto believed, and may be of value in reducing dependence on contrast use.

Selective classification (or classification with a reject option) pairs a classifier with a selection function to determine whether or not a prediction should be accepted. This framework trades off coverage (probability of accepting a prediction) with predictive performance, typically measured by distributive loss functions. In many application scenarios, such as credit scoring, performance is instead measured by ranking metrics, such as the Area Under the ROC Curve (AUC). We propose a model-agnostic approach to associate a selection function to a given probabilistic binary classifier. The approach is specifically targeted at optimizing the AUC. We provide both theoretical justifications and a novel algorithm, called AUCROSS, to achieve such a goal. Experiments show that our method succeeds in trading-off coverage for AUC, improving over existing selective classification methods targeted at optimizing accuracy.

To date, much of the algorithmic fairness literature has focused on the fairness of classification systems which are used, for example, to decide whether a person should be granted a loan or be released from prison on bail. Even in cases where such classification decisions are based on risk score models - such as in the highly influential COMPAS case [5, 11, 16] - their fairness is typically considered a function of the decisions, or classifications, made by the system. Of course, any risk score model can be turned into a classifier by selecting a probability threshold (in binary classification) or predicting the most likely outcome (in multi-class classification). Nevertheless, we argue here that it is worthwhile to distinguish between these two settings and consider the fairness of risk models independent of their downstream use, be it as the basis for a classifier or otherwise. We discuss notions of fairness for risk scores as well as their relationship to classical, classification-level notions of fairness, and we develop robust tools to empirically quantify risk score fairness. We illustrate our methodology in two case studies, one situated in the criminal justice system and one in healthcare. Why distinguish between fair models and fair decisions? In the statistical literature, it is generally considered desirable to distinguish between inference (e.g., identifying a risk score model) and subsequent decision-making (e.g., deriving a classification from a risk score model): while the former represents a purely statistical task, the latter depends on subjective

Traditional machine learning follows a close-set assumption that the training and test set share the same label space. While in many practical scenarios, it is inevitable that some test samples belong to unknown classes (open-set). To fix this issue, Open-Set Recognition (OSR), whose goal is to make correct predictions on both close-set samples and open-set samples, has attracted rising attention. In this direction, the vast majority of literature focuses on the pattern of open-set samples. However, how to evaluate model performance in this challenging task is still unsolved. In this paper, a systematic analysis reveals that most existing metrics are essentially inconsistent with the aforementioned goal of OSR: (1) For metrics extended from close-set classification, such as Open-set F-score, Youden's index, and Normalized Accuracy, a poor open-set prediction can escape from a low performance score with a superior close-set prediction. (2) Novelty detection AUC, which measures the ranking performance between close-set and open-set samples, ignores the close-set performance. To fix these issues, we propose a novel metric named OpenAUC. Compared with existing metrics, OpenAUC enjoys a concise pairwise formulation that evaluates open-set performance and close-set performance in a coupling manner. Further analysis shows that OpenAUC is free from the aforementioned inconsistency properties. Finally, an end-to-end learning method is proposed to minimize the OpenAUC risk, and the experimental results on popular benchmark datasets speak to its effectiveness. Project Page: https://github.com/wang22ti/OpenAUC.