Random Patches Research Articles

K-Surfaces: Bézier-Splines Interpolating at Gaussian Curvature Extrema

K-surfaces are an interactive modeling technique for Bézier-spline surfaces. Inspired by k -curves by [Yan et al. 2017], each patch provides a single control point that is being interpolated at a local extremum of Gaussian curvature. The challenge is to solve the inverse problem of finding the center control point of a Bézier patch given the boundary control points and the handle. Unlike the situation in 2D, bi-quadratic Bézier patches may exhibit none, one, or several extrema, and finding them is non-trivial. We solve the difficult inverse problem, including the possible selection among several extrema, by learning the desired function from samples, generated by computing Gaussian curvature of random patches. This approximation provides a stable solution to the ill-defined inverse problem and is much more efficient than direct numerical optimization, facilitating the interactive modeling framework. The local solution is used in an iterative optimization incorporating continuity constraints across patches. We demonstrate that the surface varies smoothly with the handle location and that the resulting modeling system provides local and generally intuitive control. The idea of learning the inverse mapping from handles to patches may be applicable to other parametric surfaces.

Convolutional Tsetlin Machine-based Training and Inference Accelerator for 2-D Pattern Classification

The Tsetlin Machine (TM) is a machine learning algorithm based on an ensemble of Tsetlin Automata (TAs) that learns propositional logic expressions from Boolean input features. In this paper, the design and implementation of a Field Programmable Gate Array (FPGA) accelerator based on the Convolutional Tsetlin Machine (CTM) is presented. The accelerator performs classification of two pattern classes in 4 × 4 Boolean images with a 2 × 2 convolution window. Specifically, there are two separate TMs, one per class. Each TM comprises 40 propositional logic formulas, denoted as clauses, which are conjunctions of literals. Include/exclude actions from the TAs determine which literals are included in each clause. The accelerator supports full training, including random patch selection during convolution based on parallel reservoir sampling across all clauses. The design is implemented on a Xilinx Zynq XC7Z020 FPGA platform. With an operating clock speed of 40 MHz, the accelerator achieves a classification rate of 4.4 million images per second with an energy per classification of 0.6 μJ. The mean test accuracy is 99.9% when trained on the 2-dimensional Noisy XOR dataset with 40% noise in the training labels. To achieve this performance, which is on par with the original software implementation, Linear Feedback Shift Register (LFSR) random number generators of minimum 16 bits are required. The solution demonstrates the core principles of a CTM and can be scaled to operate on multi-class systems for larger images.

Evaluation of the effects of grass and shrub cover on overland flow resistance and its attributes under simulated rainfall

Vegetation plays a crucial role in controlling soil erosion and regulating overland flow connectivity. However, the effects of vegetation coverage on the overland flow resistance mechanisms remain unclear. This study aimed to examine the impact of grass and shrub cover on the resistance to overland flow concerning a random patch distribution pattern. This was achieved through a series of simulated rainfall experiments. Thirty slope surfaces with different vegetation coverages, consisting of complete combinations of five grass and six shrub coverages, were tested under five rainfall intensities (60–120 mm h−1) and three slope gradients (8.75%–26.79%). The results indicated that the Darcy–Weisbach friction factor (f) values on unvegetated slopes ranged from 0.455 to 1.385, whereas those on vegetated slopes ranged from 0.500 to 5.624. Vegetation coverage increased flow resistance on hillslopes, and the relationship between f and the Reynolds number (Re) was influenced by vegetation cover and the slope gradient; thus, it showed a positive correlation when vegetation form resistance was dominant. In contrast, a negative correlation was observed. According to the t-test results, the linear superposition method had limitations in calculating overland flow resistance and was only applicable when the grass coverage was 10%. Moreover, as the coverage increased, significant differences (p < 0.05) emerged between the total resistance and the linear summation of the individual resistance components. Additional vegetation resistance plays a significant role in impeding the water flow, contributing up to 50% of the total resistance. Finally, an equation for calculating additional vegetation resistance was developed using dimensional analysis and multiple nonlinear regression. The accuracy of the equation was validated by comparing it with experimental results (NSE = 0.918, RMSE = 0.144). These findings emphasize the effectiveness of constructing grass-shrub communities and increasing vegetation coverage as measures for soil and water conservation.

PatchCensor: Patch Robustness Certification for Transformers via Exhaustive Testing

In the past few years, Transformer has been widely adopted in many domains and applications because of its impressive performance. Vision Transformer (ViT), a successful and well-known variant, attracts considerable attention from both industry and academia thanks to its record-breaking performance in various vision tasks. However, ViT is also highly nonlinear like other classical neural networks and could be easily fooled by both natural and adversarial perturbations. This limitation could pose a threat to the deployment of ViT in the real industrial environment, especially in safety-critical scenarios. How to improve the robustness of ViT is thus an urgent issue that needs to be addressed. Among all kinds of robustness, patch robustness is defined as giving a reliable output when a random patch in the input domain is perturbed. The perturbation could be natural corruption, such as part of the camera lens being blurred. It could also be a distribution shift, such as an object that does not exist in the training data suddenly appearing in the camera. And in the worst case, there could be a malicious adversarial patch attack that aims to fool the prediction of a machine learning model by arbitrarily modifying pixels within a restricted region of an input image. This kind of attack is also called physical attack, as it is believed to be more real than digital attack. Although there has been some work on patch robustness improvement of Convolutional Neural Network, related studies on its counterpart ViT are still at an early stage as ViT is usually much more complex with far more parameters. It is harder to assess and improve its robustness, not to mention to provide a provable guarantee. In this work, we propose PatchCensor, aiming to certify the patch robustness of ViT by applying exhaustive testing. We try to provide a provable guarantee by considering the worst patch attack scenarios. Unlike empirical defenses against adversarial patches that may be adaptively breached, certified robust approaches can provide a certified accuracy against arbitrary attacks under certain conditions. However, existing robustness certifications are mostly based on robust training, which often requires substantial training efforts and the sacrifice of model performance on normal samples. To bridge the gap, PatchCensor seeks to improve the robustness of the whole system by detecting abnormal inputs instead of training a robust model and asking it to give reliable results for every input, which may inevitably compromise accuracy. Specifically, each input is tested by voting over multiple inferences with different mutated attention masks, where at least one inference is guaranteed to exclude the abnormal patch. This can be seen as complete-coverage testing, which could provide a statistical guarantee on inference at the test time. Our comprehensive evaluation demonstrates that PatchCensor is able to achieve high certified accuracy (e.g., 67.1% on ImageNet for 2%-pixel adversarial patches), significantly outperforming state-of-the-art techniques while achieving similar clean accuracy (81.8% on ImageNet). The clean accuracy is the same as vanilla ViT models. Meanwhile, our technique also supports flexible configurations to handle different adversarial patch sizes by simply changing the masking strategy.

Few-shot biomedical image segmentation using diffusion models: Beyond image generation

BackgroundMedical image analysis pipelines often involve segmentation, which requires a large amount of annotated training data, which is time-consuming and costly. To address this issue, we proposed leveraging generative models to achieve few-shot image segmentation. MethodsWe trained a denoising diffusion probabilistic model (DDPM) on 480,407 pelvis radiographs to generate 256 ✕ 256 px synthetic images. The DDPM was conditioned on demographic and radiologic characteristics and was rigorously validated by domain experts and objective image quality metrics (Frechet inception distance [FID] and inception score [IS]). For the next step, three landmarks (greater trochanter [GT], lesser trochanter [LT], and obturator foramen [OF]) were annotated on 45 real-patient radiographs; 25 for training and 20 for testing. To extract features, each image was passed through the pre-trained DDPM at three timesteps and for each pass, features from specific blocks were extracted. The features were concatenated with the real image to form an image with 4225 channels. The feature-set was broken into random patches, which were fed to a U-Net. Dice Similarity Coefficient (DSC) was used to compare the performance with a vanilla U-Net trained on radiographs. ResultsExpert accuracy was 57.5 % in determining real versus generated images, while the model reached an FID = 7.2 and IS = 210. The segmentation UNet trained on the 20 feature-sets achieved a DSC of 0.90, 0.84, and 0.61 for OF, GT, and LT segmentation, respectively, which was at least 0.30 points higher than the naively trained model. ConclusionWe demonstrated the applicability of DDPMs as feature extractors, facilitating medical image segmentation with few annotated samples.

Simulating the LOcal Web (SLOW)

Context. Several observations of the Local Universe point toward the existence of very prominent structures: massive galaxy clusters and local superclusters on the one hand, but also large local voids and underdensities on the other. However, it is highly nontrivial to connect such different observational selected tracers to the underlying dark matter (DM) distribution. Aims. Therefore, constructing mock catalogs of such observable tracers using cosmological hydrodynamics simulations is needed. These simulations have to follow galaxy formation physics and also have to be constrained to reproduce the Local Universe. Such constraints should be based on observables that directly probe the full underlying gravitational field, such as the observed peculiar velocity field, to provide an independent test on the robustness of these distinctive structures. Methods. We used a 500 h−1 Mpc constrained simulation of the Local Universe to investigate the anomalies in the local density field, as found in observations. Constructing the initial conditions based on peculiar velocities derived from the CosmicFlows-2 catalog makes the predictions of the simulations completely independent from the distribution of the observed tracer population, and following galaxy formation physics directly in the hydrodynamics simulations also allows the comparison to be based directly on the stellar masses of galaxies or X-ray luminosity of clusters. We also used the 2668 h−1 Mpc large cosmological box from the Magneticum simulations to evaluate the frequency of finding such anomalies in random patches within simulations. Results. We demonstrate that halos and galaxies in our constrained simulation trace the local dark matter density field very differently. Thus, this simulation reproduces the observed 50% underdensity of galaxy clusters and groups within the sphere of ≈100 Mpc when applying the same mass or X-ray luminosity limit used in the observed cluster sample (CLASSIX), which is consistent with a ≈1.5σ feature. At the same time, the simulation reproduces the observed overdensity of massive galaxy clusters within the same sphere, which on its own also corresponds to a ≈1.5σ feature. Interestingly, we find that only 44 out of 15 635 random realizations (i.e., 0.28%) match both anomalies, thus making the Local Universe a ≈3σ environment. We finally compared a mock galaxy catalog with the observed distribution of galaxies in the Local Universe, finding a match to the observed factor of 2 overdensity at ∼16 Mpc as well as the observed 15% underdensity at ∼40 Mpc. Conclusions. Constrained simulations of the Local Universe which reproduce the main features of the local density field open a new window for local field cosmology, where the imprint of the specific density field and the impact on the bias through the observational specific tracers can be investigated in detail.

Balancing Performance and Energy Consumption of Bagging Ensembles for the Classification of Data Streams in Edge Computing

In recent years, the Edge Computing (EC) paradigm has emerged as an enabling factor for developing technologies like the Internet of Things (IoT) and 5G networks, bridging the gap between Cloud Computing services and end-users, supporting low latency, mobility, and location awareness to delay-sensitive applications. An increasing number of solutions in EC have employed machine learning (ML) methods to perform data classification and other information processing tasks on continuous and evolving data streams. Usually, such solutions have to cope with vast amounts of data that come as data streams while balancing energy consumption, latency, and the predictive performance of the algorithms. Ensemble methods achieve remarkable predictive performance when applied to evolving data streams due to several models and the possibility of selective resets. This work investigates a strategy that introduces short intervals to defer the processing of mini-batches. Well balanced, our strategy can improve the performance (i.e., delay, throughput) and reduce the energy consumption of bagging ensembles to classify data streams. The experimental evaluation involved six state-of-art ensemble algorithms (OzaBag, OzaBag Adaptive Size Hoeffding Tree, Online Bagging ADWIN, Leveraging Bagging, Adaptive RandomForest, and Streaming Random Patches) applying five widely used machine learning benchmark datasets with varied characteristics on three computer platforms. As a result, our strategy can significantly reduce energy consumption in 96% of the experimental scenarios evaluated. Despite the trade-offs, it is possible to balance them to avoid significant loss in predictive performance.

Unsteady evolution of distributed roughness-induced vortices under re-entry conditions

AbstractThis study investigates a re-entry scenario of an Apollo-like space capsule with Direct Numerical Simulations (DNS). The simulation includes the chemical equilibrium gas model. Cross-flow-like vortices are induced through random distributed roughness patches on the capsule surface. Two different machine learning methods are used to predict the maximum vorticity magnitude downstream of a pseudo-random roughness patch, the wall-normal location of the vortex core and spanwise and wall-normal gradient maxima of u. A large DNS database is formed for training and testing of the neural networks. In order to understand the influence of the vorticity magnitude on the transition process, local one-dimensional inviscid (LODI) relations are used to describe perturbations at the inflow. The disturbance evolution in the streamwise direction is analysed with a two-dimensional Fourier transformation in time and space. We show how the vorticity magnitudes of the cross-flow-like vortices, spanwise and wall-normal derivatives of the streamwise velocity influence the transition location.

Hyperspectral Image Classification Based on Fusing S3-PCA, 2D-SSA and Random Patch Network

Recently, the rapid development of deep learning has greatly improved the performance of image classification. However, a central problem in hyperspectral image (HSI) classification is spectral uncertainty, where spectral features alone cannot accurately and robustly identify a pixel point in a hyperspectral image. This paper presents a novel HSI classification network called MS-RPNet, i.e., multiscale superpixelwise RPNet, which combines superpixel-based S3-PCA with two-dimensional singular spectrum analysis (2D-SSA) based on the Random Patches Network (RPNet). The proposed frame can not only take advantage of the data-driven method, but can also apply S3-PCA to efficiently consider more global and local spectral knowledge at the super-pixel level. Meanwhile, 2D-SSA is used for noise removal and spatial feature extraction. Then, the final features are obtained by random patch convolution and other steps according to the cascade structure of RPNet. The layered extraction superimposes the different sparial information into multi-scale spatial features, which complements the features of various land covers. Finally, the final fusion features are classified by SVM to obtain the final classification results. The experimental results in several HSI datasets demonstrate the effectiveness and efficiency of MS-RPNet, which outperforms several current state-of-the-art methods.

Multi-Task Distributed Learning using Vision Transformer with Random Patch Permutation.

The widespread application of artificial intelligence in health research is currently hampered by limitations in data availability. Distributed learning methods such as federated learning (FL) and split learning (SL) are introduced to solve this problem as well as data management and ownership issues with their different strengths and weaknesses. The recent proposal of federated split task-agnostic (FESTA) learning tries to reconcile the distinct merits of FL and SL by enabling the multi-task collaboration between participants through Vision Transformer (ViT) architecture, but they suffer from higher communication overhead. To address this, here we present a multi-task distributed learning using ViT with random patch permutation, dubbed p-FESTA. Instead of using a CNN-based head as in FESTA, p-FESTA adopts a simple patch embedder with random permutation, improving the multi-task learning performance without sacrificing privacy. Experimental results confirm that the proposed method significantly enhances the benefit of multi-task collaboration, communication efficiency, and privacy preservation, shedding light on practical multi-task distributed learning in the field of medical imaging.

M3AE: Multimodal Representation Learning for Brain Tumor Segmentation with Missing Modalities

Multimodal magnetic resonance imaging (MRI) provides complementary information for sub-region analysis of brain tumors. Plenty of methods have been proposed for automatic brain tumor segmentation using four common MRI modalities and achieved remarkable performance. In practice, however, it is common to have one or more modalities missing due to image corruption, artifacts, acquisition protocols, allergy to contrast agents, or simply cost. In this work, we propose a novel two-stage framework for brain tumor segmentation with missing modalities. In the first stage, a multimodal masked autoencoder (M3AE) is proposed, where both random modalities (i.e., modality dropout) and random patches of the remaining modalities are masked for a reconstruction task, for self-supervised learning of robust multimodal representations against missing modalities. To this end, we name our framework M3AE. Meanwhile, we employ model inversion to optimize a representative full-modal image at marginal extra cost, which will be used to substitute for the missing modalities and boost performance during inference. Then in the second stage, a memory-efficient self distillation is proposed to distill knowledge between heterogenous missing-modal situations while fine-tuning the model for supervised segmentation. Our M3AE belongs to the ‘catch-all’ genre where a single model can be applied to all possible subsets of modalities, thus is economic for both training and deployment. Extensive experiments on BraTS 2018 and 2020 datasets demonstrate its superior performance to existing state-of-the-art methods with missing modalities, as well as the efficacy of its components. Our code is available at: https://github.com/ccarliu/m3ae.

Using a Local Binary Pattern and Random Patch Convolution for A Multi-Kernel Mode using classification of hyper-spectral Image

Using a Local Binary Pattern and Random Patch Convolution for A Multi-Kernel Mode using classification of hyper-spectral Image

Small Sample Hyperspectral Image Classification Based on the Random Patches Network and Recursive Filtering

In recent years, different deep learning frameworks were introduced for hyperspectral image (HSI) classification. However, the proposed network models have a higher model complexity, and do not provide high classification accuracy if few-shot learning is used. This paper presents an HSI classification method that combines random patches network (RPNet) and recursive filtering (RF) to obtain informative deep features. The proposed method first convolves image bands with random patches to extract multi-level deep RPNet features. Thereafter, the RPNet feature set is subjected to dimension reduction through principal component analysis (PCA), and the extracted components are filtered using the RF procedure. Finally, the HSI spectral features and the obtained RPNet-RF features are combined to classify the HSI using a support vector machine (SVM) classifier. In order to test the performance of the proposed RPNet-RF method, some experiments were performed on three widely known datasets using a few training samples for each class, and classification results were compared with those obtained by other advanced HSI classification methods adopted for small training samples. The comparison showed that the RPNet-RF classification is characterized by higher values of such evaluation metrics as overall accuracy and Kappa coefficient.

John Stewart: A personal appreciation

John Stewart: A personal appreciation

Grading of HCC Biopsy Images using Nucleus and Texture Features.

Hepatocellular carcinoma (HCC) is one of the most critical health problems in the world. For proper treatment, it is important to identify the grade of cancer morbidity from HCC biopsy image. The diagnostic work is not only time-consuming but also subjective. The same biopsy image may be diagnosed as of different grades by different doctors, due to lack of experience or difference in opinion. In this work, we proposed an automatic grading system with classification accuracy matching to an experienced doctor, to help augment the diagnosis process. First, we proposed a segmentation method to isolate all nucleus-like objects present in a biopsy image. Non-target objects (here the target is a single HCC nucleus) present in the biopsy image are isolated too in the segmentation process. To eliminate such non-target objects, we proposed clustering of segmented images and a novel method to filter out target objects. Next, we proposed a two track neural network, where input consists of 2 different images. It combines a single segmented nucleus and a random cropped texture patch of the biopsy image to which the nucleus belongs. At this classifier output, we grade the single nucleus. Finally, a majority voting method is used to identify the grade of the whole biopsy image. We achieved an accuracy of 99.03% for nucleus image grading and 99.67% accuracy for grading biopsy images.

Analyzing the role of tofacitinib in treatment of alopecia areata: A retrospective analysis from a tertiary care center of North India

Background: Alopecia areata is a disease, in which patients’ loss hairs as random patches due to autoimmune complications. Tofacitinib acts as an inhibiter in the Janus kinase signal transducer and activator of transcription pathway and is an effective drug approved for the of rheumatoid arthritis treatment. Limited evidence is available on the role of tofacitinib in treatment of alopecia areata. Aim and Objective: The present study aims to evaluate the efficacy of tofacitinib in treatment of alopecia areata patients visiting the tertiary care center of North India. Materials and Methods: Total 17 patients with alopecia areata were recruited in this study and prescribed with the tofacitinib 5 mg twice daily for 6 months. Severity of alopecia tool (SALT) was used to analyze the severity of hair loss. The SALT score measures the total percent of hair loss in all area of scalp. A higher SALT score indicates higher hair loss. Results: The mean age of the patients was 27.88 ± 16.30 years with male-to-female sex ratio of 2.4:1. The mean duration of disease was 32.29 ± 22.14 months. There was a significant decrease in the SALT score (P = 0.0001) following the 6 months tofacitinib treatment. Conclusion: Tofacitinib was found to be effective in the treatment of alopecia areata. Further multicentric studies with ample sample size are required to establish tofacitinib as a standard treatment for the alopecia areata.

Self-supervised learning-based Multi-Scale feature Fusion Network for survival analysis from whole slide images

Understanding prognosis and mortality is critical for evaluating the treatment plan of patients. Advances in digital pathology and deep learning techniques have made it practical to perform survival analysis in whole slide images (WSIs). Current methods are usually based on a multi-stage framework which includes patch sampling, feature extraction and prediction. However, the random patch sampling strategy is highly unstable and prone to sampling non-ROI. Feature extraction typically relies on hand-crafted features or convolutional neural networks (CNNs) pre-trained on ImageNet, while the artificial error or domain gaps may affect the survival prediction performance. Besides, the limited information representation of local sampling patches will create a bottleneck limitation on the effectiveness of prediction. To address the above challenges, we propose a novel patch sampling strategy based on image information entropy and construct a Multi-Scale feature Fusion Network (MSFN) based on self-supervised feature extractor. Specifically, we adopt image information entropy as a criterion to select representative sampling patches, thereby avoiding the noise interference caused by random to blank regions. Meanwhile, we pretrain the feature extractor utilizing self-supervised learning mechanism to improve the efficiency of feature extraction. Furthermore, a global–local feature fusion prediction network based on the attention mechanism is constructed to improve the survival prediction effect of WSIs with comprehensive multi-scale information representation. The proposed method is validated by adequate experiments and achieves competitive results on both of the most popular WSIs survival analysis datasets, TCGA-GBM and TCGA-LUSC. Code and trained models are made available at: https://github.com/Mercuriiio/MSFN.

Multi-Subject Image Retrieval by Fusing Object and Scene-Level Feature Embeddings

Most existing image retrieval methods separately retrieve single images, such as a scene, content, or object, from a single database. However, for general purposes, target databases for image retrieval can include multiple subjects because it is not easy to predict which subject is entered. In this paper, we propose that image retrieval can be performed in practical applications by combining multiple databases. To deal with multi-subject image retrieval (MSIR), image embedding is generated through the fusion of scene- and object-level features, which are based on Detection Transformer (DETR) and a random patch generator with a deep-learning network, respectively. To utilize these feature vectors for image retrieval, two bags-of-visual-words (BoVWs) were used as feature embeddings because they are simply integrated with preservation of the characteristics of both features. A fusion strategy between the two BoVWs was proposed in three stages. Experiments were conducted to compare the proposed method with previous methods on conventional single-subject datasets and multi-subject datasets. The results validated that the proposed fused feature embeddings are effective for MSIR.

Multi-Scenario Simulation to Predict Ecological Risk Posed by Urban Sprawl with Spontaneous Growth: A Case Study of Quanzhou.

The rapid expansion of different types of urban land continues to erode natural and semi-natural ecological space and causes irreversible ecological damage to rapidly industrialized and urbanized areas. This work considers Quanzhou, a typical industrial and trade city in southeastern China as the research area and uses a Markov chain integrated into the patch-generating land use simulation (PLUS) model to simulate the urban expansion of Quanzhou from 2005 to 2018. The PLUS model uses the random forest algorithm to determine the contribution of driving factors and simulate the organic and spontaneous growth process based on the seed generation mechanism of multi-class random patches. Next, leveraging the importance of ecosystem services and ecological sensitivity as indicators of evaluation endpoints, we explore the temporal and spatial evolution of ecological risks from 2018 to 2031 under the scenarios of business as usual (BAU), industrial priority, and urban transformation scenarios. The evaluation endpoints cover water conservation service, soil conservation service, biodiversity maintenance service, soil erosion sensitivity, riverside sensitivity, and soil fertility. The ecological risk studied in this work involves the way in which different types of construction land expansion can possibly affect the ecosystem. The ecological risk index is divided into five levels. The results show that during the calibration simulation period from 2005 to 2018 the overall accuracy and Kappa coefficient reached 91.77% and 0.878, respectively. When the percent-of-seeds (PoS) parameter of random patch seeds equals 0.0001, the figure of merit of the simulated urban construction land improves by 3.9% compared with the logistic-based cellular automata model (Logistic-CA) considering organic growth. When PoS = 0.02, the figure of merit of the simulated industrial and mining land is 6.5% higher than that of the Logistic-CA model. The spatial reconstruction of multiple types of construction land under different urban development goals shows significant spatial differentiation on the district and county scale. In the industrial-priority scenario, the area of industrial and mining land is increased by 20% compared with the BAU scenario, but the high-level risk area is 42.5% larger than in the BAU scenario. Comparing the spatial distribution of risks under the BAU scenario, the urban transition scenario is mainly manifested as the expansion of medium-level risk areas around Quanzhou Bay and the southern region. In the future, the study area should appropriately reduce the agglomeration scale of urban development and increase the policy efforts to guide the development of industrial land to the southeast.

Evaluating the performance of ensemble classifiers in stock returns prediction using effective features

Stock market prediction is considered as an important yet challenging aspect of financial analysis. The difficulty of forecasting arises from volatile and non-linear nature of stock market, which is affected by varied uncertain factors, ranging from financial ratios to macroeconomic indicators. Recent advances in machine learning, particularly ensembles, have made it possible for academic researchers and financial practitioners to forecast the stock market more efficiently. The novelty of this work is to evaluate how stock return in an oil-dependent country (i.e., Iran), which has been facing stagflation for a long time due to economic and political issues, is affected by fundamental and macroeconomic indicators. Our main objectives are to (1) find the most important fundamental and macroeconomic indicators that control the stock returns of companies listed on the Tehran Stock Exchange (TSE); (2) compare the performance of newly developed bagging- and boosting-based ensembles in predicting annual real stock returns of the TSE; and (3) develop multiclass classification models to forecast stock returns. Prior studies mainly focused on developing binary classification models, which simply predict whether stock returns will be positive or negative in the future. We, however, design multiclass classification models to provide more information for the investors and reduce the uncertainties associated with the prediction. To this end, we first provide a comprehensive list of 57 potential features affecting the stock returns. Next, the data are carefully preprocessed and fed to 14 different bagging- and boosting-based ensembles (e.g., Random Forest, LightGBM, XGBoost, Extra-Trees, AdaBoost, CatBoost) to predict the stock returns. The performance of ensembles is evaluated through different measures (e.g., accuracy, F-score, G-mean). We then propose a novel feature selection method to identify the most contributing features to the stock returns. Our proposed model identifies nearly 65% of 57 original features as redundancy, resulting in 20 most significant features. The selected features are fed to the mentioned ensembles to re-predict the stock returns. Finally, the performance of stock returns forecasts with and without selected features is compared. To design the ensembles, we employ the data from listed companies on the TSE for a 15-year period, spanning between 2005 and 2020. Results suggest that boosting ensembles, in general, outperform bagging-based methods. Among the boosting ensembles, XGBoost and AdaBoost provide the best and worst predictive performance, respectively. Among the bagging-like ensembles, Rotation Forest is the most accurate one, whereas Random Patches performs the worst. Further, our proposed feature selection approach effectively identifies the most representative features for stock returns prediction and can be used as a reliable framework for future investment decisions.