Umeå University's logo

Explainable machine learning aims to strike a balance between prediction accuracy and model transparency, particularly in settings where black-box predictive models, such as deep neural networks or kernel-based methods, achieve strong empirical performance but remain difficult to interpret. This work introduces a mixture of generalized additive models (GAMs) in which random Fourier feature (RFF) representations are leveraged to uncover locally adaptive structure in the data. In the proposed method, an RFF-based embedding is first learned and then compressed via principal component analysis. The resulting low-dimensional representations are used to perform soft clustering of the data through a Gaussian mixture model. These cluster assignments are then applied to construct a mixture-of-GAMs framework, where each local GAM captures nonlinear effects through interpretable univariate smooth functions. Numerical experiments on real-world regression benchmarks, including the California Housing, NASA Airfoil Self-Noise, and Bike Sharing datasets, demonstrate improved predictive performance relative to classical interpretable models. Overall, this construction provides a principled approach for integrating representation learning with transparent statistical modeling.

The low-temperature embrittlement of duplex stainless steels significantly restricts their broader application in critical environments. Despite extensive research, a reliable predictive model for this phenomenon remains unavailable. In this study, we present a data-driven framework that integrates two machine learning (ML) models to predict the evolution of ferrite micro-hardness and steel toughness during thermal ageing. We assembled two experimental datasets from both the open literature and in-house experiments. The ferrite hardness dataset includes chemical composition and ageing conditions, while the steel toughness dataset incorporates all features from the ferrite dataset, along with ferrite grain size and ferrite fraction.

A systematic selection of input features and ML algorithms was conducted to optimize model performance. The integrated framework is based on random forest regression, where ML1 predicts changes in ferrite hardness, and ML2 estimates variations in steel toughness using the predicted ferrite hardness from ML1 as an input feature. This linkage reflects the metallurgical understanding that ferrite hardness serves as a key indicator of low-temperature embrittlement in duplex stainless steel. The trained models achieved high predictive accuracy, with R2 values exceeding 0.97 for both ferrite hardness and steel toughness across multiple stainless steel grades. Furthermore, the models demonstrated strong generalizability when applied to unseen alloys and new ageing conditions. To assess model interpretability, feature importance analysis was performed to evaluate the influence of individual input variables, and the results were interpreted through the lens of physical metallurgy, offering insights into the underlying mechanisms of embrittlement.

Understanding the timely regulation of plant growth and phenology is crucial for assessing a terrestrial ecosystem's productivity and carbon budget. The circadian clock, a system of genetic oscillators, acts as 'Master of Ceremony' during plant physiological processes. The mechanism is particularly elusive in trees despite its relevance. The primary and secondary tree growth, leaf senescence, bud set, and bud burst timing were investigated in 68 constructs transformed into Populus hybrids and compared with untransformed or transformed controls grown in natural or controlled conditions. The results were analyzed using generalized additive models with ordered-factor-smooth interaction smoothers. This meta-analysis shows that several genetic components are associated with the clock. Especially core clock-regulated genes affected tree growth and phenology in both controlled and field conditions. Our results highlight the importance of field trials and the potential of using the clock to generate trees with improved characteristics for sustainable silviculture (e.g., reprogrammed to new photoperiodic regimes and increased growth).

• Understanding isotope fractionation mechanisms is fundamental for analyses of plant ecophysiology and paleoclimate based on tree-ring isotope data.
• To gain new insights into isotope fractionation, we analysed intramolecular ¹³C discrimination in tree-ring glucose (Δ_i', i = C-1 to C-6) and metabolic deuterium fractionation at H¹ and H² (ε_met) combinedly. This dual-isotope approach was used for isotope-signal deconvolution.
• We found evidence for metabolic processes affecting Δ₁' and Δ₃', which respond to air vapour pressure deficit (VPD), and processes affecting Δ₁' and Δ₃', and ε_met, which respond to precipitation but not VPD. These relationships exhibit change points dividing a period of homeostasis (1961–1980) from a period of metabolic adjustment (1983–1995). Homeostasis may result from sufficient groundwater availability. Additionally, we found Δ₅' and Δ₆' relationships with radiation and temperature, which are temporally stable and consistent with previously proposed isotope fractionation mechanisms.
• Based on the multitude of climate covariables, intramolecular carbon isotope analysis has a remarkable potential for climate reconstruction. While isotope fractionation beyond leaves is currently considered to be constant, we propose significant parts of the carbon and hydrogen isotope variation in tree-ring glucose originate in stems (precipitation-dependent signals). As basis for follow-up studies, we propose mechanisms introducing Δ₁', Δ₂', Δ₃', and ε_met variability.

This paper introduces a novel ridgelet transform-based method for Poisson image denoising. Our work focuses on harnessing the Poisson noise's unique non-additive and signal-dependent properties, distinguishing it from Gaussian noise. The core of our approach is a new thresholding scheme informed by theoretical insights into the ridgelet coefficients of Poisson-distributed images and adaptive thresholding guided by Stein's method. We verify our theoretical model through numerical experiments and demonstrate the potential of ridgelet thresholding across assorted scenarios. Our findings represent a significant step in enhancing the understanding of Poisson noise and offer an effective denoising method for images corrupted with it.

Low-count positron emission tomography (PET) data suffer from high noise levels, leading topoor image quality and reduced diagnostic accuracy. Compressed sensing (CS) based denoisingmethods have shown potential in medical imaging. This study investigates the performance ofCS-based denoising methods on PET sinograms.Three simulated datasets were used in this study, including circular phantom, patient pelvisphantom, and patient brain phantom. Ten sampling levels were employed to investigate the effect of data reduction on diagnostic accuracy. CS-based denoising methods were applied prereconstruction, and a conventional Gaussian post-filter was used for comparison. Performancemeasures included rRMSE, SSIM, SNR, line profiles, and FWHM.Overall, the proposed CS-based denoising methods performed similarly to the benchmark interms of lesion contrast, spatial resolution, and noise texture. The proposed methods outperformed the benchmark in low-count situations by suppressing background noise and preservingcontrast better.The results of this study demonstrate that CS-based denoising methods in the sinogram domain can improve the quality of low-count PET images, particularly in suppressing backgroundnoise and preserving contrast. These findings suggest that CS-based denoising could be apromising solution for improving the diagnostic accuracy of low-count PET data.

This study investigated the climate effect under consecutive winters on the arrival delay of high-speed passenger trains. Inhomogeneous Markov chain model and stratified Cox model were adopted to account for the time-varying risks of train delays. The inhomogeneous Markov chain modelling used covariates weather variables, train operational direction, and findings from the primary delay analysis through stratified Cox model. The results showed that temperature, snow depth, ice/snow precipitation, and train operational direction significantly impacted the arrival delay. Further, by partitioning the train line into three segments as per transition intensity, the model identified that the middle segment had the highest chance of a transfer from punctuality to delay, and the last segment had the lowest probability of recovering from delayed state. The performance of the fitted inhomogeneous Markov chain model was evaluated by the walk-forward validation method, which indicated that approximately 9% of trains may be misclassified as having arrival delays by the fitted model at a measuring point on the train line. With the model performance, the fitted model could be beneficial for both travellers to plan their trips reasonably and railway operators to design more efficient and wiser train schedules as per weather condition.

Objective: In this study, the aim was to compare the performance of four spatiotemporal decomposition algorithms (stICA, stJADE, stSOBI, and sPCA) and parameters for identifying single motor units in human skeletal muscle under voluntary isometric contractions in ultrafast ultrasound image sequences as an extension of a previous study. The performance was quantifed using two measures: (1) the similarity of components’ temporal characteristics against gold standard needle electromyography recordings and (2) the agreement of detected sets of components between the diferent algorithms.

Results: We found that out of these four algorithms, no algorithm signifcantly improved the motor unit identifcation success compared to stICA using spatial information, which was the best together with stSOBI using either spatialor temporal information. Moreover, there was a strong agreement of detected sets of components between the different algorithms. However, stJADE (using temporal information) provided with complementary successful detections. These results suggest that the choice of decomposition algorithm is not critical, but there may be a methodological improvement potential to detect more motor units

Explicitly using the block structure of the unknown signal can achieve better reconstruction performance in compressive sensing. An unknown signal with block structure can be accurately recovered from under-determined linear measurements provided that it is sufficiently block sparse. However, in practice, the block sparsity level is typically unknown. In this paper, we propose a soft measure of block sparsity k_α(x) = (||x||_2,α/||x||_2,1) ^α/(1−α) with α ∈ [0,∞], and present a procedure to estimate it by using multivariate centered isotropic symmetric α-stable random projections. The limiting distribution of the estimator is given. Simulations are conducted to illustrate our theoretical results.

• Stable isotope abundances convey valuable information about plant physiological processes and underlying environmental controls. Central gaps in our mechanistic understanding of hydrogen isotope abundances impede their widespread application within the plant and biogeosciences.
• To address these gaps, we analysed intramolecular deuterium abundances in glucose of Pinus nigra extracted from an annually resolved tree-ring series (1961–1995).
• We found fractionation signals (i.e. temporal variability in deuterium abundance) at glucose H¹ and H² introduced by closely related metabolic processes. Regression analysis indicates that these signals (and thus metabolism) respond to drought and atmospheric CO₂ concentration beyond a response change point. They explain ≈ 60% of the whole-molecule deuterium variability. Altered metabolism is associated with below-average yet not exceptionally low growth.
• We propose the signals are introduced at the leaf level by changes in sucrose-to-starch carbon partitioning and anaplerotic carbon flux into the Calvin–Benson cycle. In conclusion, metabolism can be the main driver of hydrogen isotope variation in plant glucose.