Ersin Yılmaz

Smartphone-based human activity recognition (HAR) uses features collected from smartphone sensors to distinguish between different human activities. The large number of correlated features generated by smartphone sensors can make feature selection and model interpretation challenging. This study evaluates a two-stage feature selection and classification framework for HAR, in which sparse partial least squares discriminant analysis (sPLS-DA) is used primarily as a supervised feature-selection method, while multinomial logistic regression is employed as the subsequent classifier. The approach is compared with LASSO and Elastic Net regularization under consistent training, testing, and cross-validation procedures. The experiment used 30 repetitions. In each repetition, 21 subjects were used for training and nine subjects were used for testing, with five-fold subject-level cross-validation inside the training set. The original variables selected by each method were evaluated using both unpenalized and Ridge-regularized multinomial logistic regression (MLR) classifiers. Native LASSO and Elastic Net showed similar predictive performance, with mean accuracies of approximately 95%. After expanding the tuning grid, sPLS-DA selected almost all available features, indicating limited sparsity under the investigated tuning range. Using unpenalized MLR on the selected features reduced predictive performance, while Ridge-regularized MLR improved the results for all three feature-selection methods. Overall, the results show that predictive performance, feature-set size, tuning behavior, and the choice of downstream classifier should be considered together when evaluating feature-selection methods for smartphone-based HAR.

Smartphone-based human activity recognition (HAR) uses features collected from smartphone sensors to distinguish between different human activities. The large number of correlated features generated by smartphone sensors can make feature selection and model interpretation challenging. This study evaluates a two-stage feature selection and classification framework for HAR, in which sparse partial least squares discriminant analysis (sPLS-DA) is used primarily as a supervised feature selection method, while multinomial logistic regression is employed as the subsequent classifier. The approach is compared with LASSO and Elastic Net regularization under consistent training, testing, and cross-validation procedures. The experiment used 30 repetitions. In each repetition, 21 subjects were used for training and nine subjects were used for testing, with five-fold subject-level cross-validation inside the training set. The original variables selected by each method were also tested with the same unpenalized multinomial logistic regression (MLR) classifier. Native LASSO and Elastic Net gave the highest predictive performance, with mean accuracies of 0.9514 ± 0.0166 and 0.9504 ± 0.0187, respectively, and mean multiclass AUC values of about 0.997. LASSO selected 169.8 ± 25.4 features on average, while Elastic Net selected 255.5 ± 114.4. sPLS-DA selected 353.7 ± 47.5 features on average, while its native max.dist classifier reached a mean accuracy of 0.8831 ± 0.0283 and the selected feature sets had the highest mean pairwise Jaccard overlap (0.6611). Using unpenalized MLR on the selected feature subsets reduced accuracy for all three methods. Overall, LASSO gave the best balance between predictive performance and a smaller feature set, while sPLS-DA showed higher feature-set overlap but kept many more features. These results show that predictive performance, feature-set size, and selection stability should be considered together in smartphone-based HAR.

Individual brains are unique in structure and function. Functional differences are captured by neural fingerprints, which reflect individual differences in behavior and cognition as well as group‐level changes related to neurodegenerative diseases. Most research efforts so far have focused on fingerprints comprising full functional connectomes. However, the high dimensionality of the connectomes can increase computational load and impede performance of machine learning methods in potential applications. A low‐dimensional alternative that retains individual features of the full connectomes would thus be beneficial. The present study employed latent‐noise Bayesian Reduced Rank Regression (lnBRRR) to learn low‐dimensional latent spaces that capture individual features in functional connectivity and power spectral density data derived from MEG recordings. LnBRRR performance was assessed with low training set sizes ( N = 20–44), and against principal component analysis and linear discriminant analysis. Model performance was also assessed with task data, and the solutions were compared across task conditions with cosine similarity to establish whether individual features are altered by different cognitive processes. LnBRRR captured generalizable individual patterns already at N = 20 but N = 30–35 was needed to reach optimal test accuracies and to prevent potential overfitting. The model also achieved comparable performance to the alternative models. Latent fingerprints derived from task data attained comparable performance to resting‐state latent fingerprints, and lnBRRR solutions were shown to generalize across conditions. Additionally, the model solutions for power spectral density data were discovered to be notably similar, yet differently rotated, over task conditions, suggesting that similar patterns of individual features were captured by the model regardless of the task condition. Altogether, the present results highlight lnBRRR as a potential tool for neuroimaging data analysis and demonstrate that individual differences in power spectral density are largely intrinsic and unaffected by varying cognitive processes.

Individual brains are unique in structure and function. Functional differences are captured by neural fingerprints, which reflect individual differences in behavior and cognition as well as group-level changes related to neurodegenerative diseases. Most research efforts so far have focused on fingerprints com-prising full functional connectomes. However, the high dimensionality of the connectomes can increase computational load and impede performance of machine learning methods in potential applications. A low-dimensional alternative that retains individual features of the full connectomes would thus be beneficial. The present study employed latent-noise Bayesian Reduced Rank Regression (lnBRRR) to learn low-dimensional latent spaces that capture individual features in functional connectivity and power spectral density data derived from MEG recordings. LnBRRR performance was assessed with low training set sizes (N=20-44), and against principal component analysis and linear discriminant analysis. Model performance was also assessed with task data, and the solutions were compared across task conditions with cosine similarity to establish whether individual features are altered by different cognitive processes. LnBRRR captured generalizable individual patterns already at N=20 but N=30-35 was needed to reach optimal test accuracies and to prevent potential overfitting. The model also achieved comparable performance to the alternative models. Latent fingerprints derived from task data attained comparable performance to resting-state latent fingerprints, and lnBRRR solutions were shown to generalize across conditions. Additionally, the model solutions for power spectral density data were discovered to be notably similar, yet differently rotated, over task conditions, suggesting that similar patterns of individual features were captured by the model regardless of the task condition. Altogether, the present results highlight lnBRRR as a potential tool for neuroimaging data analysis and demonstrate that individual differences in power spectral density are largely intrinsic and unaffected by varying cognitive processes.

Survival analysis on biomedical data has seen quick methodological growth, but practitioners face limited guidance on which family of methods to prefer when censoring rates are high or the ratio of predictors to events is unsuitable. This paper presents a systematic benchmark of six survival models—the classical Cox model, lasso-penalized Cox, the Bayesian elastic net Cox model (BEN-Cox), random survival forests, DeepSurv, and Cox-Time—across three publicly available datasets (METABRIC, SUPPORT, and TCGA-BRCA) under controlled censoring regimes. The experimental grid scans events-per-variable (EPV) ratios from 0.6 to 7.6 and censoring rates from 33% to 86%. Performance is evaluated using Harrell’s concordance index, time-dependent concordance, the integrated Brier score, and the calibration slope, with statistical comparisons carried out by paired Wilcoxon signed-rank tests with Holm correction. Three main findings are obtained. First, BEN-Cox, lasso-Cox, and random survival forests form a narrow top tier in discrimination, with pooled C-index values within 0.003 of each other, while DeepSurv and Cox-Time do not reach this tier in any configuration. Second, calibration separates models that discrimination does not: BEN-Cox achieves the lowest integrated Brier score in all six configurations, while lasso-Cox produces calibration slopes closest to one on the high-dimensional datasets; deep models and the unpenalized Cox model are systematically overconfident. Third, increasing censoring widens the gap between the regularized top tier and the deep architectures without changing which model family is preferred. An empirical partition of the (c,EPV) plane summarizes these observations. Because the entire grid falls below the classical EPV threshold of ten, the partition documents the low-EPV regime in which regularized models dominate but does not identify the boundary at which deep models may become competitive.

The Artificial Bee Colony (ABC) algorithm is a widely used swarm-based optimizer, and current research continues to develop more effective ABC variants for high-dimensional, constrained, and nonconvex problems. Motivated by this direction, we introduce a global--local ABC framework in which the outer ABC search is combined with an efficient local refinement step based on the alternating direction method of multipliers (ADMM). The main idea is to use ABC to generate and perturb diverse candidate solutions, while using ADMM as an efficient stationary-point solver for each candidate, so that exploration and exploitation are balanced in a controlled way. A key point in our setting is an orthogonality constraint on the decision matrix. To address this constraint more directly, we also introduce a manifold-aware ABC variant that performs candidate perturbations followed by frequent Stiefel projections, aiming to keep feasibility and stabilize the search in strongly sparse settings. We evaluate the introduced ABC variants on an orthogonality-constrained sparse optimization problem arising from one-stage sparse principal component regression (SPCRsvd), which is a challenging nonconvex objective with bilinear coupling and sparsity penalties. In a comprehensive Monte Carlo study that includes low and high-dimensional settings, alignment and misalignment, and data-generating difficulty such as well-conditioned linear vs. correlated and nonlinear, we report prediction error, subspace recovery error, and runtime, and compare against standard ABC, single-start ADMM-type solvers, and a ridge baseline. The results show the situations where ABC--ADMM integration and manifold-aware exploration can provide solutions with lower objective values than pure local search, while remaining substantially cheaper than full population-based search in high dimensions.

High-dimensional survival analyses require calibrated risk and honest uncertainty, but standard elastic-net Cox models yield only point estimates. We develop a fully Bayesian elastic-net Cox (BEN–Cox) modelfor high-dimensional proportional hazards regression that places a hierarchical global–local shrinkage prior on coefficients and performs full Bayesian inference via Hamiltonian Monte Carlo. We represent the elastic–net penalty as a global–local Gaussian scale mixture with hyperpriors that learn the ℓ1/ℓ2 trade-off, enabling adaptive sparsity that preserves correlated gene groups and, using HMC on the Cox partial likelihood, yields full posteriors for hazard ratios and patient-level survival curves. Methodologically, we formalize a Bayesian analogue of the elastic-net grouping effect at the posterior mode and establish posterior contraction under sparsity for the Cox partial likelihood, supporting the stability of the resulting risk scores. On the METABRIC breast-cancer cohort (n = 1 , 903; 440 gene-level features from an Illumina array with ≈ 24,000 gene-level features (probes)), BEN–Cox achieves slightly lower prediction error, higher discrimination, and better global calibration than a tuned ridge Cox baseline on a held-out test set. Posterior summaries provide credible intervals for hazard ratios, identify a compact gene panel that remains biologically plausible. BEN–Cox provides a theory-backed, uncertainty-aware alternative to tuned penalised Cox models, improving calibration and yielding an interpretable sparse signature in correlated, high-dimensional survival data

This paper considers semiparametric estimation strategies for the nonlinear semiparametric regression model (NSRM) under the sparsity assumption by modifying the Gauss–Newton method for both low- and high-dimensional data scenarios. In the low-dimensional case, coefficients are partitioned into two parts that represent nonzero (strong signals) and sparse coefficients. In the high-dimensional case, a weighted-ridge approach is employed, and coefficients are partitioned into three parts, adding weak signals as well. Shrinkage estimators are then obtained in both cases. More importantly, in this paper, we assume that a nonlinear structure is present in the parametric component of the model, which makes the direct application of penalized least squares to the NSRM impossible. To solve this problem, we employ the iterative Gauss–Newton method to obtain the final NSRM estimators. We provide both theoretical and practical details for the suggested estimators. Asymptotic results are derived for both low- and high-dimensional cases. We conduct an extensive simulation study to evaluate the performance of the estimators in a practical setting. Moreover, we substantiate our findings with data examples from two distinct breast cancer datasets: the Breast Cancer in the United States (BCUS) and Wisconsin datasets. By demonstrating the effectiveness of our introduced estimators in these particular biostatistical contexts, our numerical study provides support for the theoretical efficacy of shrinkage estimators, suggesting their potential relevance to breast cancer research and biostatistical methodologies.