Beam position monitors (BPMs) in particle accelerators generate high-rate multivariate time series susceptible to noise, drift, and transient faults. Operational pipelines discard channels that fail quality thresholds, retaining no diagnostic score and no denoised output. We propose the Dual-Decoder Convolutional Autoencoder (DDCAE), an unsupervised model that performs anomaly scoring and nonlinear denoising within a single architecture. A shared 1D-CNN encoder feeds an anomaly decoder, trained on clean inputs, whose reconstruction error yields per-channel fault scores, and a denoising decoder, trained on structured corruptions of the same inputs to restore the underlying waveform. An ablation over three random seeds separates the work of the two branches. Detection does not depend on the denoising decoder, because the anomaly decoder alone flags the faulty channels. The denoising decoder instead restores the accepted signals. It raises their mean signal-to-noise ratio by 4.1 dB, against 1.2 dB for the single-decoder reconstruction, and it halves the residual tune error. This restoration adds 27 393 parameters, a 25 10^-3 on the accepted channels, without modifying the pipeline or requiring labelled anomalies.
High-resolution hydrodynamic simulations are essential for assessing urban pluvial flood hazards, but their computational cost limits rapid scenario analysis and operational use. Simulation-based surrogate models provide efficient approximations, yet standard global predictors often hide the heterogeneous structure of simulator-derived flood responses. We investigate Mixture-of-Experts (MoE) surrogates for discovering model-induced latent response regimes in large-scale environmental simulation data. We train a neural-gated mixture of linear experts, where the gate assigns each cell–event instance a graded membership distribution over specialized experts, and interpret the regimes induced by this specialization rather than individual predictions. The approach is evaluated on a large tabular dataset derived from physics-based hydrodynamic simulations of the urban area of Crotone, Italy. We compare the MoE surrogate with linear, tree-based, linear-tree, and neural models, and characterize regimes through routing, tail enrichment, feature profiles, and spatial coherence. The results show that a few experts capture a disproportionate share of high-response cases, while large-coverage experts describe ordinary low-response conditions. The MoE therefore retains competitive flood-response approximation while exposing hydrologically plausible, response-coherent latent regimes.
Spatial transcriptomics measures gene expression while preserving the spatial organization of transcripts within tissue sections. Combined with complementary modalities such as histological images, these data enable a comprehensive characterization of tissue architecture and disease-associated patterns. Machine learning approaches, particularly deep learning and graph-based models, have proven effective in extracting meaningful representations from these heterogeneous data. However, many existing methods analyze tissue sections independently, mostly relying on the assumption that spatially adjacent spots are functionally similar, which often fails in cancer tissues where sharp tumor-stroma boundaries exist. Moreover, most existing approaches consider each patient separately, limiting generalization across individuals. In this paper, we introduce HG-SC, a graph-based method for spot-level classification that integrates three complementary modalities: gene expression values, tissue morphology from paired hematoxylin and eosin images, and the spatial coordinates of each spot within the tissue. Unlike existing methods, HG-SC constructs a unified graph across different patients and tissue sections, enabling the model to capture shared biological patterns within a patient tissue, taking into account patient-specific variations. We evaluate HG-SC on a colorectal cancer dataset gathered from the HEST-1K repository using a tumor-versus-stroma classification task under a leave-one-patient-out cross-validation setting. Our results show that HG-SC consistently outperforms existing multi-modal and unimodal approaches, also due to its ability to capture pathological patterns across patients.
Personalized learning strategies contribute to more suitable learning and increased student performance. Many educational data mining methods retrieve global patterns: information that fits the majority of students. Instead, we deploy Exceptional Model Mining (EMM) to discover subgroups of students having exceptional learning needs. We model learning behavior as a linear regression, with student performance as the output and selected study efforts as input variables. A third set of variables consisting of student characteristics is kept apart to generate candidate subgroups of students. EMM methods exist to assess the exceptionality of the regression parameter vector for any candidate subgroup. Since many such candidates exist for what essentially boils down to a statistical test, this procedure suffers from the multiple comparison problem; resulting false discoveries can be filtered out by the existing DFD validation method. However, DFD validation only compares the extent to which a difference in estimated regression parameter vectors is likely to arise purely by random effects; it is agnostic to the extent to which the regression models for the subgroup and full dataset fit the data well at all. In this paper, we add that dimension to the statistical testing, partitioning discovered subgroups into five categories: true discoveries, true discoveries with inappropriate model, weak discoveries, false discoveries, and inappropriate models (discoveries where both the subgroup and global model are not better than a simple mean-based baseline). We employ this new data analysis to discover exceptional subgroups on proprietary data of courses Calculus and Data Analytics. We derive stratified learning strategies for all subgroups that are true discoveries, fortified by stakeholder feedback from the responsible lecturer for one of the courses.
False positives remain a persistent challenge in operational remote sensing detection tasks, such as greenhouse gas plume monitoring and oil spill detection. Benchmark performance often masks this problem, as operational settings are characterised by severe class imbalance, limited labelled data, and difficult true negatives that are hard to separate from true positives. We study this problem through the lens of atmospheric methane plume detection, where end-to-end models, which are generally applicable and not task-specific, achieve competitive benchmark results but generate substantially more operational false positives than a machine learning pipeline designed by domain experts. We find that the discriminating features of difficult true negatives form a marginalised discriminant, a feature that is informative only within a subpopulation of the training data, making it hard to learn from the full distribution. We investigate multiple model- and data-centric techniques targeting the marginalised discriminant to address false positives in an end-to-end pipeline and analyse the properties of successful approaches. The most effective solution, introduced in related work, most strongly limits what the model needs to learn by decomposing the problem into two steps to reveal the hidden feature, and by increasing data efficiency through pre-computed features using feature engineering. Our findings suggest that successful operational pipelines do not solely rely on smarter learning algorithms, but also on the degree to which the complexity of the problem to be solved is constrained before machine learning is applied.
