It is shown that a usable signal is available after convergence, when loss no longer distinguishes the two populations, and applying a fixed perturbation to a converged model's inputs flips the predictions of the latter far more often than the former.
Abstract
Models trained by empirical risk minimization on data containing spurious correlations achieve high average accuracy while failing on subpopulations where the correlation does not hold. Existing methods for identifying the affected samples without group annotations rely on signals from early training, which requires locating the epoch at which to intervene, a hyperparameter typically selected using group-labeled validation data. We show that a usable signal is available after convergence, when loss no longer distinguishes the two populations. Samples consistent with the spurious correlation are classified by a shared rule, while the remaining samples are fit through configurations specific to individual inputs and are correspondingly more fragile. Applying a fixed perturbation to a converged model's inputs flips the predictions of the latter far more often than the former. The resulting procedure requires two forward passes per training sample, no group annotations at any stage, and no early-stopping epoch. Using the detected samples to rebalance training raises worst-group accuracy on Waterbirds from 57.3% to 80.8%, against 85.8% with ground-truth group labels.
Early stopping is a fundamental technique for mitigating overfitting in machine learning and is especially critical in many weakly supervised learning scenarios. Conventional early stopping, however, needs a sufficiently sized and representative validation set. This creates a trade-off: dedicating more data for validation can diminish the training set and limit model performance, while an insufficient validation set may lead to sub-optimal model selection. This challenge is particularly serious in many weakly supervised learning scenarios where validation data may be entirely unavailable or unreliable. In this paper, we propose Label Wave, which does not require validation data for selecting the desired model across various weakly supervised learning paradigms, including learning with noisy labels (LNL), positive-unlabeled learning, and unlabeled-unlabeled learning. It works by tracking the changes in the model's predictions on the training set during the training process, aiming to halt training before the model excessively fits misleading patterns. This method is empirically supported by our observation that minimum fluctuations in predictions typically occur at the training epoch before the model transitions from fitting reliable patterns to misleading ones. Extensive experiments further provide large-scale evaluation for LNL across diverse noise conditions, model families, optimizers, and data modalities. Label Wave selects near-oracle checkpoints (test-accuracy oracle gaps: 0.15-0.93 percentage points in controlled sweeps and 1.90 percentage points on 71 tabular tasks) and boosts 7 LNL methods by 1.98/2.86 percentage points over 10%/20% hold-out validation.
Suqin Yuan, Muyang Li, Lei Feng et al.· IEEE Transactions on Pattern...· 0 citations
Concept-Residual eXpansion (CRX), a concept-augmented framework that improves robustness by expanding the set of candidate predictive features by improving robustness to spurious correlations, is proposed.
Eric Xie, Guangzhi Xiong, Wenqian Ye et al.· Proceedings of the 32nd ACM...· 0 citations
Benchmark contamination is diagnosed with n-gram overlap, likelihood-based membership inference, or canary strings, and each needs something usually unavailable: the training corpus, a well-chosen test statistic, or foresight at release. A recent alternative reads it off a linear probe on internal activations. We show the natural way to do this does not work, specify one that survives measurement, then find that the correction making it work carries more variance than the null it is tested against. The protocol reports a zero-sum contrast on the depth profile of probe accuracy, recentred on a level-matched placebo baseline, tested against a label-permutation null, with the reference set twice the size of the suspect set. Each choice replaces a simpler alternative we rejected on measurement. Reporting the level of excess separability rather than its shape makes the false positive rate track the size of the analyst's own control set, 0.03 to 0.99 under a true null. Contrasting against a flat depth profile rejects a true null 0.72 of the time when surface decodability rises with depth, and loses all power when it falls. On real transformers the protocol fails a test the simulations did not pose. The recentring subtracts an estimate, and the permutation null holds it fixed. Re-estimated across split seeds on four audits of contaminated checkpoints, its standard deviation is 1.30 to 1.56 times the null's own in every arm: what is subtracted to remove a bias is more variable than what it corrects. The one nominally significant result, p = 0.0075, becomes 0.0745 once that variance is propagated, and no verdict is issued. The simulations missed this because their surface key is the covariate driving item variation; on real text it is a proxy, and degrading key quality in simulation reproduces it. We add a companion measurement and a widened null. No arm shows contamination.
The Normalised Sensitivity Ratio (NSR) is introduced, a post-hoc, model-agnostic diagnostic for this question under a structured-shift regime: environments differ primarily in the mean of spurious features while the causal mechanism and causal marginals remain stable, as in multi-site clinical data or multi-batch genomics.
Athanasios Vlontzos, G. Papanastasiou, Bernhard Kainz et al.· 0 citations
Longitudinal sensing studies routinely collect thousands of windows from a few dozen participants. The records are numerous; the independent scientific units are not. When the outcome is defined per participant, this mismatch makes apparently precise findings vulnerable to pseudo-replication, to partition choice, and to the ordinary analytic flexibility of comparing several pipelines before reporting one. Splitting on participants prevents a person's records from straddling a split, but it does not calibrate the label-dependent workflow fold construction, preprocessing, tuning, calibration, and candidate selection that produced the reported number. We define a participant-level estimand and obtain an analysis-matched null by permuting participant labels and rerunning that entire workflow. In controlled simulation, window-level inference rejects in 70-80% of replicates when no effect exists and a window bootstrap rejects at the same rate; a participant bootstrap still rejects at 10-17%; the analysis-matched test holds 0.025-0.100 across cohorts of 20 to 80 participants. Freezing the selected pipeline instead of repeating the search inflates Type-I error to 0.240 with eight candidates, where repeating it holds 0.040. Applied to two public cohorts, wrist actigraphy (n=55) yields participant AUROC 0.928 with p=0.0050, a conclusion that persists under a scale-robust rank-pooled statistic and under a matched permutation null computed after excluding hospitalized participants (p=0.0089). Smartphone sensing (n=38, 7 positives) yields 0.636 and does not reject (p=0.1724) despite sufficient resolution, with sensitivity 0.143. The practical rule is narrow: every step that reads labels belongs inside the permuted analysis, and repeated records do not create additional independent participants.
Prediction-based methods, including Large Language Models (LLMs) and other machine learning techniques, are often used to construct measures of political phenomena that are difficult to quantify directly, such as policy positions in manifestos or emotions expressed on social media. In many applications, these prediction-generated measures are used as explanatory variables in regression models, even though they are measured with error. This leads to biased estimates. In this paper, we propose a simple solution to these biases: instrumental variables constructed from multiple measures created on independent splits of the original data. This approach is theoretically valid, easy to implement, and does not require new data. Through simulations, we show that this approach recovers estimates close to the true values, even in relatively small samples, while the standard approach can produce substantial bias in practice. We illustrate the method by revisiting two applications: whether gendered speech affects legislative outcomes in the German Parliament, and whether political risk influences poverty alleviation programs in China.