Combination Of Classifiers Research Articles

A hybrid classification approach for automatically recognizing COVID-19 using deep transfer learning using chest radiographs

Coronavirus 2019 causes COVID-19, a worldwide epidemic. It endangers millions globally. Early illness detection improves recovery and control. X-ray image processing is used to categorise and identify COVID-19 in the present study. Preprocessing, feature extraction using local binary pattern (LBP) and edge orient histogram (EOH), and classification utilising K-nearest neighbour (KNN), Navie Bayes, support vector machine (SVM), and transfer learning convolution neural networks (CNNs) are some of the stages that are implemented in the process. Other phases in the process include preprocessing, feature extraction, and preprocessing. LBP+KNN, EOH+KNN, LBP+SVM, EOH +SVM, CNN+LBP, and CNN+EOH are the outputs derived from the combinations of feature extraction operators and classifiers. Other possible outcomes are CNN+EOH and CNN+LBP. A total of 4,000 pictures were used as the basis for conducting an analysis of the performance of six different models. In order to train the models, 10-fold cross-validation was used, and their accuracy was measured accordingly. The evaluation results indicate a high level of accuracy in diagnosis, ranging from 90.2% to 97.56%. The CNN+LBP and CNN+EOH models have demonstrated superior performance compared to other models, achieving average accuracies ranging from 96.66% and 98.54%.

A comparative analysis of PlanetScope 4-band and 8-band imageries for land use land cover classification

Earth-observing satellites have become essential in comprehending human impacts on the landscape. Satellite-based imagery is indispensable for mapping Earth's features, managing resources, and studying environmental changes. Readily available remote sensing data with improved radiometric, spectral, spatial, and temporal resolution presents opportunities for advanced data analysis. Precise and accurate land use land cover (LULC) information is essential for the surveillance of environmental conditions and the effective management of natural resources. This research assesses the performance of PlanetScope product SuperDove sensor (PSB.SD), having two different band combinations, including 4-band (Red, Blue, Green and Near-Infrared (NIR)) and 8-band (Blue, Green II, Red, NIR, Coastal Blue, Green, Yellow, Red-Edge) imagery in ArcGIS Pro for the month of July 2021. Four different supervised classifiers, including support vector machine (SVM), k-nearest neighbours (KNN), random forest (RF), and maximum likelihood (ML) classifiers. This study was carried out for the three major areas, i.e., City of Summerside, City of Charlottetown, and Town of Three Rivers in Prince Edward Island (PEI), Canada and LULC classification scheme consist of six major classes which includes Agriculture, Forest, Vegetation, Bare Land, Urban and Water bodies. For accuracy assessment, overall accuracy as well as kappa coefficient were estimated to identify the most accurate combination of LULC classifier and different band combination imagery from PlanetScope. Results show that the highest overall accuracy of 0.94 for Town of Three Rivers and 0.93 for City of Summerside and City of Charlottetown were observed using 8-band imagery with SVM classifier. The lowest overall accuracy of 0.78 for Town of Three Rivers, 0.83 for City of Charlottetown, and 0.82 for City of Summerside was observed using 4-band imagery using ML classifier. Further, the SVM classifier performs well in accuracy with 8-band imagery of PlanetScope, showcasing its potential in LULC classification compared to previous PlanetScope 4-band imagery.

Intelligent fault diagnosis for tribo-mechanical systems by machine learning: Multi-feature extraction and ensemble voting methods

Timely fault detection is crucial for preventing issues like worn clutch plates and excessive friction material degradation, enhancing fuel efficiency, and prolonging clutch lifespan. This study focuses on early fault diagnosis in dry friction clutch systems using machine learning (ML) techniques. Vibration data is analyzed under different load and fault conditions, extracting statistical, histogram, and auto-regressive moving average (ARMA) features. Feature selection employs the J48 decision tree algorithm, evaluated with eight ML classifiers: support vector machines (SVM), k-nearest neighbor (kNN), linear model tree (LMT), random forest (RF), multilayer perceptron (MLP), logistic regression (LR), J48, and Naive Bayes. The evaluation revealed that individual classifiers achieved the highest testing accuracies with statistical feature selection as 83% for both MLP and LR at no load, 90% for MLP at 5 kg, and 93% for KNN at 10 kg. For histogram feature selection, KNN and MLP both reached 85% at no load, MLP achieved 91% at 5 kg, and RF attained 97% at 10 kg. With ARMA feature selection, KNN reached 93% at no load, LR achieved 94% at 5 kg, and RF reached 86% at 10 kg. The voting strategy notably improved these results, with the RF-KNN-J48 ensemble reaching 98% for histogram features at 10 kg, the KNN-LMT-RF ensemble achieving 94% for ARMA features at no load, and the SVM-MLP-LMT ensemble attaining 95% for ARMA features at 5 kg. Hence, a combination of three classifiers using the majority voting rule consistently outperforms standalone classifiers, striking a balance between diversity and complexity, facilitating robust decision-making. In practical applications, selecting the optimal combination of feature selection method and classifier is vital for accurate fault classification. This study provides valuable guidance for engineers and practitioners implementing robust load classification systems in industrial settings.

Hybrid Intelligent Anomaly Detection System Using Attention based Deep Learning Approach for Cyber Attacks Prevention

Objectives: Network Intrusion Detection System (NIDS) plays an important role in finding and preventing cyber-attacks, which helps to improve the entire security posture of an organization’s network infrastructure. The development of Deep Learning (DL) techniques possess the ability of IDS to detect attacks without delay and protects from intrusions even in real-time environment. Methods: The present study proposes an improved IDS framework called Enhanced Gated Recurrent Unit Hyper-Model combined Attention Bidirectional Long-Short Term Memory (EGHAB) approach, to effectively address the detection of attacks with maximum accuracy and minimal error rate. The proposed model is enhanced by using methods like numericalization and normalisation for pre-processing the input data, Genetic Algorithm (GA) for extracting intricate features from the input data, and the EGHAB classifier, which works on the principles of both Gated Recurrent Unit (GRU) and Bidirectional Long-Short Term Memory (BiLSTM) models along with attention mechanism. Findings: The EGHAB model efficiently learns the abnormal behaviour of network and classified the attack and non-attack data using the publicly available NSL-KDD dataset and a generated real time data set, scapy. And it achieved 99.94% accuracy over NSL KDD dataset during multiclass classification and 93.3% on real time data set with reduced Error rate. Additionally, to examine the efficacy, the proposed approach is compared with other existing methods and proved its improved performance. Novelty: A combined ensemble classifier with GRU, BiLSTM and attention mechanism is designed to predict the attacks in earlier stage rather than due over. The model achieved better accuracy and reduced false assumption using anomaly prediction mechanism. Keywords: Cyber Attack, Anomaly Detection, Deep learning, Gated Recurrent Unit, Bi­ Directional Long­-Short Term Memory, Attention mechanism

A Lightweight Method for Breast Cancer Detection Using Thermography Images with Optimized CNN Feature and Efficient Classification.

Breast cancer is a prominent cause of death among women worldwide. Infrared thermography, due to its cost-effectiveness and non-ionizing radiation, has emerged as a promising tool for early breast cancer diagnosis. This article presents a hybrid model approach for breast cancer detection using thermography images, designed to process and classify these images into healthy or cancerous categories, thus supporting disease diagnosis. Multiple pre-trained convolutional neural networks are employed for image feature extraction, and feature filter methods are proposed for feature selection, with diverse classifiers utilized for image classification. Evaluating the DRM-IR test set revealed that the combination of ResNet34, Chi-square ( ) filter, and SVM classifier demonstrated superior performance, achieving the highest accuracy at . Furthermore, the highest accuracy improvement obtained was when using the SVM classifier and Chi-square filter compared to regular convolutional neural networks. The results confirmed that the proposed method, with its high accuracy and lightweight model, outperforms state-of-the-art breast cancer detection from thermography image methods, making it a good choice for computer-aided diagnosis.

Handling imbalance dataset issue in insider threat detection using machine learning methods

Insider threats, characterized by their baleful impact and substantial costs, arise from internal factors within organizations. These threats are rare and usually unnoticed, as the malicious actions are often submerged in numerous normal activities, causing dataset imbalance and making detection hard. To address these challenges, in this paper we propose a Two-Step Insider Threat Detection (TSITD) approach. First, it preprocesses the CERT r4.2 and r5.2 datasets into day-long sequences. Second, it handles the dataset imbalance and detects threats by forming various combinations of sampling techniques and classifiers, referred to as TSITD models. When we compare these TSITD models to baseline models, we observe a significant improvement in anomaly detection rate and balanced accuracy. The TSITD models also achieve higher rankings when evaluated using the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) method.

Preoperative prediction of CNS WHO grade and tumour aggressiveness in intracranial meningioma based on radiomics and structured semantics

Preoperative identification of intracranial meningiomas with aggressive behaviour may help in choosing the optimal treatment strategy. Radiomics is emerging as a powerful diagnostic tool with potential applications in patient risk stratification. In this study, we aimed to compare the predictive value of conventional, semantic based and radiomic analyses to determine CNS WHO grade and early tumour relapse in intracranial meningiomas. We performed a single-centre retrospective analysis of intracranial meningiomas operated between 2007 and 2018. Recurrence within 5 years after Simpson Grade I-III resection was considered as early. Preoperative T1 CE MRI sequences were analysed conventionally by two radiologists. Additionally a semantic feature score based on systematic analysis of morphological characteristics was developed and a radiomic analysis were performed. For the radiomic model, tumour volume was extracted manually, 791 radiomic features were extracted. Eight feature selection algorithms and eight machine learning methods were used. Models were analysed using test and training datasets. In total, 226 patients were included. There were 21% CNS WHO grade 2 tumours, no CNS WHO grade 3 tumour, and 25 (11%) tumour recurrences were detected in total. In ROC analysis the best radiomic models demonstrated superior performance for determination of CNS WHO grade (AUC 0.930) and early recurrence (AUC 0.892) in comparison to the semantic feature score (AUC 0.74 and AUC 0.65) and conventional radiological analysis (AUC 0.65 and 0.54). The combination of human classifiers, semantic score and radiomic analysis did not markedly increase the model performance. Radiomic analysis is a promising tool for preoperative identification of aggressive and atypical intracranial meningiomas and could become a useful tool in the future.

A Framework for Detecting Distributed Denial of Services Attack in Cloud Enviorment using Machine Learning Techniques

Distributed Denial of Service (DDoS) persists in Online Applications as One of those significant threats. Attackers can execute DDoS by the more natural steps. Then with the high productivity to slow the consumer access services down. To detect an attack on DDoS and using machine learning algorithms. The Overseen to detect and mitigate the attack, machine learning algorithms such as Naive Bayes, decision tree, k-nearest neighbours (k-NN) and random forest are used. There are three steps: gathering information, preprocessing and feature Extraction in "Normal or DDoS" classification algorithm for detection Attack use Dataset NSL-KDD. Similar algorithms have different functions Conduct that is dependent on the features selected. DDOS-attack performance Detection is compared, and it indicates the best algorithm. Attempts at Distributed Denial of Service ( DDoS) Were the most powerful attacks of the last period. A Virtual Network. The intrusion detection system (NIDS) should be designed seamlessly to Fight the latest strategies and trends of those attackers NIDS on DDoS. In this paper, we propose an NIDS capable of detecting Current DDoS attacks, as well as new forms. The main characteristic of Our NIDS is the combination of various classifiers using an ensemble Models, with the concept of each classifier being able to target different Aspects/types of intrusions, and more effective in doing so Mechanism for protecting against new intrusions. Additionally, we perform a detailed study of and based on, DDoS attacks check the reduced set of functions [27, 28] to be essential to Enhance accuracy. We are playing with and analyzing NSL-KDD Dataset with a feature set reduced, and our proposed NIDS will Detect 99.1 per cent of active DDoS attacks. Let's compare our Tests with other methods which already exist. Our approach to NIDS has Able to learn to keep up with existing and evolving DDoS Attack trends.

Attention-aware with stacked embedding for sentiment analysis of student feedback through deep learning techniques.

Automatic polarity prediction is a challenging assessment issue. Even though polarity assessment is a critical topic with many existing applications, it is probably not an easy challenge and faces several difficulties in natural language processing (NLP). Public polling data can give useful information, and polarity assessment or classification of comments on Twitter and Facebook may be an effective approach for gaining a better understanding of user sentiments. Text embedding techniques and models related to the artificial intelligence field and sub-fields with differing and almost accurate parameters are among the approaches available for assessing student comments. Existing state-of-the-art methodologies for sentiment analysis to analyze student responses were discussed in this study endeavor. An innovative hybrid model is proposed that uses ensemble learning-based text embedding, a multi-head attention mechanism, and a combination of deep learning classifiers. The proposed model outperforms the existing state-of-the-art deep learning-based techniques. The proposed model achieves 95% accuracy, 97% recall, having a precision of 95% with an F1-score of 96% demonstrating its effectiveness in sentiment analysis of student feedback.

Exhaustive Exploitation of Nature-Inspired Computation for Cancer Screening in an Ensemble Manner.

Accurate screening of cancer types is crucial for effective cancer detection and precise treatment selection. However, the association between gene expression profiles and tumors is often limited to a small number of biomarker genes. While computational methods using nature-inspired algorithms have shown promise in selecting predictive genes, existing techniques are limited by inefficient search and poor generalization across diverse datasets. This study presents a framework termed Evolutionary Optimized Diverse Ensemble Learning (EODE) to improve ensemble learning for cancer classification from gene expression data. The EODE methodology combines an intelligent grey wolf optimization algorithm for selective feature space reduction, guided random injection modeling for ensemble diversity enhancement, and subset model optimization for synergistic classifier combinations. Extensive experiments were conducted across 35 gene expression benchmark datasets encompassing varied cancer types. Results demonstrated that EODE obtained significantly improved screening accuracy over individual and conventionally aggregated models. The integrated optimization of advanced feature selection, directed specialized modeling, and cooperative classifier ensembles helps address key challenges in current nature-inspired approaches. This provides an effective framework for robust and generalized ensemble learning with gene expression biomarkers.

Gastric Cancer Image Classification: A Comparative Analysis and Feature Fusion Strategies.

Gastric cancer is the fifth most common and fourth deadliest cancer worldwide, with a bleak 5-year survival rate of about 20%. Despite significant research into its pathobiology, prognostic predictability remains insufficient due to pathologists' heavy workloads and the potential for diagnostic errors. Consequently, there is a pressing need for automated and precise histopathological diagnostic tools. This study leverages Machine Learning and Deep Learning techniques to classify histopathological images into healthy and cancerous categories. By utilizing both handcrafted and deep features and shallow learning classifiers on the GasHisSDB dataset, we conduct a comparative analysis to identify the most effective combinations of features and classifiers for differentiating normal from abnormal histopathological images without employing fine-tuning strategies. Our methodology achieves an accuracy of 95% with the SVM classifier, underscoring the effectiveness of feature fusion strategies. Additionally, cross-magnification experiments produced promising results with accuracies close to 80% and 90% when testing the models on unseen testing images with different resolutions.

VVV catalog of ab-type RR Lyrae in the inner Galactic bulge

Observational evidence has accumulated in recent years, showing that the Galactic bulge includes two populations, a metal-poor one and a metal-rich one, which in addition to having different metallicities show different alpha over iron abundances, spatial distribution, and kinematics. While the metal-rich, barred component has been fairly well characterized, the metal-poor, spheroidal component has been more elusive and harder to describe. RR Lyrae variables are clean tracers of the old bulge component, and they are, on average, more metal-poor than red clump stars. In the present paper, we provide a new catalog of 16488 ab-type RR Lyrae variables in the bulge region within |l|$ and |b|$ extracted from multi-epoch Point Spread Function photometry performed on VISTA Variable in the V\' a L\'actea survey data. We used the catalog to constrain the shape of the old, metal-poor, bulge stellar population. The identification of ab-type RR Lyrae among a large sample of candidate variables of different types has been performed via a combination of a Random Forest classifier and visual inspection. We optimized this process in such a way to extract a clean catalog with high purity, although for this reason its completeness, close to the midplane, is lower compared to a few other near-infrared catalogs covering the same region of the sky. We used the present catalog to derive the shape of their distribution around the Galactic center, resulting in an elongated spheroid with projected axis ratio of b/asim 0.7 and an inclination angle of phi sim 20 degrees. We discuss how observational biases, such as errors on the distances and a nonuniform sampling in longitude, affect both the present measurements and previous ones, especially those based on red clump stars. Because the latter have not been taken into account before, we refrain from a quantitative comparison between these shape parameters and those derived for the main Galactic bar. Nonetheless, qualitatively, taking into account observational biases would lower the estimated ellipticity of the bar derived from red clump stars, and hence reduce the difference with the present results. We publish a high-purity RRab sample for future studies of the oldest Galactic bulge population, close to the midplane. We explore different choices for the period-luminosity-metallicity relation, highlighting how some of them introduce spurious trends of distance with either the period or the metallicity, or both. We provide evidence that they trace a structure that is less elongated than the main bar, though we also highlight some biases of these kind of studies not discussed before.

Classifier Combination Supported by the Sleep-Wake Cycle Improves EEG Seizure Prediction Performance.

Seizure prediction is a promising solution to improve the quality of life for drug-resistant patients, which concerns nearly 30% of patients with epilepsy. The present study aimed to ascertain the impact of incorporating sleep-wake information in seizure prediction. We developed five patient-specific prediction approaches that use vigilance state information differently: i) using it as an input feature, ii) building a pool of two classifiers, each with different weights to sleep/wake training samples, iii) building a pool of two classifiers, each with only sleep/wake samples, iv) changing the alarm-threshold concerning each sleep/wake state, and v) adjusting the alarm-threshold after a sleep-wake transition. We compared these approaches with a control method that did not integrate sleep-wake information. Our models were tested with data (43 seizures and 482 hours) acquired during presurgical monitoring of 17 patients from the EPILEPSIAE database. As EPILEPSIAE does not contain vigilance state annotations, we developed a sleep-wake classifier using 33 patients diagnosed with nocturnal frontal lobe epilepsy from the CAP Sleep database. Although different patients may require different strategies, our best approach, the pool of weighted predictors, obtained 65% of patients performing above chance level with a surrogate analysis (against 41% in the control method). The inclusion of vigilance state information improves seizure prediction. Higher results and testing with long-term recordings from daily-life conditions are necessary to ensure clinical acceptance. As automated sleep-wake detection is possible, it would be feasible to incorporate these algorithms into future devices for seizure prediction.

Towards automatic phytolith classification using feature extraction and combination strategies

Phytolith analysis is now an essential technique, both for the reconstruction of past environmental and climatic changes and for the study of anthropic and faunal plant use, in such disciplines as archaeology, paleoecology, paleonthology, and palynology. Currently, phytolith identification and categorisation involves time-consuming and tedious manual classification tasks that are not always error free. Automated phytolith classification will be key to the standardisation of phytolith identification processes, circumventing human error in the phytolith identification process. In this paper, a comparative analysis is presented of different types of feature sets, feature combinations, and classifier combinations (through stacking), and their use for automatic phytolith classification, including state-of-the-art vision transformers and convolutional neural networks, techniques which have shown remarkable progress within different areas, including computer vision. In this research, twenty-two different sets of features (three based on shape, sixteen on appearance, and three on texture) and six classifier strategies (single and combined via stacking) were compared. The experimental results revealed that texture-related features offered no valuable information for classification purposes. However, classification tasks were efficiently performed with strategies based on shape and appearance features (extracted using deep neural networks). More specifically, the use of those features combined with a stacking strategy, achieved better results than any other features and feature-based strategies, with an accuracy value of 98.32%.

Unveiling anomalies: harnessing machine learning for detection and insights

The rise of Internet of Things (IoT) devices has brought about an increase in security risks, emphasizing the need for effective anomaly detection systems. Previous research introduced a dynamic voting classifier to overcome overfitting or inaccurate accuracies caused by dataset imbalance. This article introduces a new method for IoT anomaly detection that employs a hybrid voting classifier, which combines several machine learning models. To solve the overfitting and class weight issues, an adaptive voting classifier is used that adjusts weights according to the highest preference for accuracy. The developing voting system increases the effectiveness of more accurate classifiers, enhancing the group’s overall capability. A proposed combined classifier combines Logistic Regression, AdaBoost, Gradient Boosting, and Multi-Layer Perceptron models using a soft voting method. To develop and assess this method, the CIC-IoT-2023 dataset is utilized, which contains 33 types of IoT attacks across 7 categories. This process includes thorough data preprocessing and feature selection from a pool of 42 available attributes. The performance of this approach is measured against individual classifiers across binary, 8-class, and 34-class classification tasks. The results highlight the effectiveness of the hybrid model. It achieves 98.95% accuracy, 76.72% recall, and 72.01% F1-score in the 34-class problem, surpassing the performance of all individual models. For the 8-class task, the hybrid classifier attains 99.39% accuracy, 90.89% recall, and an 83.01% F1-score. This demonstrates the high potential of the hybrid approach for IoT anomaly detection.

DeepDate: A deep fusion model based on whale optimization and artificial neural network for Arabian date classification.

As agricultural technology continues to develop, the scale of planting and production of date fruit is increasing, which brings higher yields. However, the increasing yields also put a lot of pressure on the classification step afterward. Image recognition based on deep learning algorithms can help to identify and classify the date fruit species, even in natural light. In this paper, a deep fusion model based on whale optimization and an artificial neural network for Arabian date classification is proposed. The dataset used in this study includes five classes of date fruit images (Barhi, Khalas, Meneifi, Naboot Saif, Sullaj). The process of designing each model can be divided into three phases. The first phase is feature extraction. The second phase is feature selection. The third phase is the training and testing phase. Finally, the best-performing model was selected and compared with the currently established models (Alexnet, Squeezenet, Googlenet, Resnet50). The experimental results show that, after trying different combinations of optimization algorithms and classifiers, the highest test accuracy achieved by DeepDate was 95.9%. It takes less time to achieve a balance between classification accuracy and time consumption. In addition, the performance of DeepDate is better than that of many deep transfer learning models such as Alexnet, Squeezenet, Googlenet, VGG-19, NasNet, and Inception-V3. The proposed DeepDate improves the accuracy and efficiency of classifying date fruits and achieves better results in classification metrics such as accuracy and F1. DeepDate provides a promising classification solution for date fruit classification with higher accuracy. To further advance the industry, it is recommended that stakeholders invest in technology transfer programs to bring advanced image recognition and AI tools to smaller producers, enhancing sustainability and productivity across the sector. Collaborations between agricultural technologists and growers could also foster more tailored solutions that address specific regional challenges in date fruit production.

A new probability adjustment method for combining conflicting evidences: application in classifier combination

A new probability adjustment method for combining conflicting evidences: application in classifier combination

Exploring the potential of Wav2vec 2.0 for speech emotion recognition using classifier combination and attention-based feature fusion

Exploring the potential of Wav2vec 2.0 for speech emotion recognition using classifier combination and attention-based feature fusion

Uncovering phishing attacks using principles of persuasion analysis

With the rising of Internet in early ’90s, many fraudulent activities have migrated from physical to digital: one of them is phishing. Phishing is a deceptive practice focused on exploiting the human factor, which is the most vulnerable aspect of any security process. In this scam, social engineering techniques are extensively utilized, specifically focusing on the principles of persuasion, to deceive individuals into disclosing sensitive information or engaging in malicious actions. This research explores the use of message subjectivity for detecting phishing attacks. It does so by assessing the impact of various data representations and classifiers on automatically identifying principles of persuasion. Furthermore, it investigates how these detected principles of persuasion can be leveraged for identifying phishing attacks. The experiments conducted revealed that there is no universal solution for data representation and classifier selection to effectively detect all principles of persuasion. Instead, a tailored combination of data representation and classifiers is required for detecting each principle. The Machine Learning models created automatically detect principles of persuasion with confidence levels ranging from 0.7306 to 0.8191 for AUC-ROC. Next, principles of persuasion detected are used for phishing detection. This study also emphasizes the need for user-friendly and comprehensible models. To validate the proposal presented, several families of classifiers were tested, but among all of them, tree-based models (and Random Forest in particular) stand out as preferred option. These models achieve similar level of effectiveness as alternative methods while offering improved clarity and user-friendliness, with an AUC-ROC of 0.859842.

Improved source classification and performance analysis using Gaia DR3

The Discrete Source Classifier (DSC) provides probabilistic classification of sources in Gaia Data Release 3 (GDR3) using a Bayesian framework and a global prior. The DSC Combmod classifier in GDR3 achieved for the extragalactic classes (quasars and galaxies) a high completeness of 92%, but a low purity of 22% (all sky, all magnitudes) due to contamination from the far larger star class. However, these single metrics mask significant variation in performance with magnitude and sky position. Furthermore, a better combination of the individual classifiers that comprise Combmod is possible. Here we compute two-dimensional (2D) representations of the completeness and the purity as a function of Galactic latitude and source brightness, and also exclude the Magellanic Clouds where stellar contamination significantly reduces the purity. Reevaluated on a cleaner validation set and without introducing changes to the published GDR3 DSC probabilities themselves, we here achieve for Combmod average 2D completenesses of 92% and 95% and average 2D purities of 55% and 89% for the quasar and galaxy classes, respectively. Since the relative proportions of extragalactic objects to stars in Gaia is expected to vary significantly with brightness and latitude, we then introduce a new prior that is a continuous function of brightness and latitude, and compute new class probabilities from the GDR3 DSC component classifiers, Specmod and Allosmod. Contrary to expectations, this variable prior only improves the performance by a few percentage points, mostly at the faint end. Significant improvement, however, is obtained by a new additive combination of Specmod and Allosmod. This classifier, Combmod-α, achieves average 2D completenesses of 82% and 93% and average 2D purities of 79% and 93% for the quasar and galaxy classes, respectively, when using the global prior. Thus, we achieve a significant improvement in purity for a small loss of completeness. The improvement is most significant for faint quasars (G≥20) where the purity rises from 20% to 62%.