Reduced Training Set Research Articles

Leak localization in District Heating Networks integrating physical model-based and data driven-based methods: Impact of dataset construction on model performance

Leaks in district heating networks result in waste of water and heat, and even threaten personal safety. Various machine learning-based methods have been proposed to realize leak localization and gain satisfying performance based on a large amount of leak samples. However, the creation of such samples demands a substantial investment of time and computational resources, presenting challenges for real-time applications. Previous studies have rarely addressed the influence of leak dataset construction and efficient creation methods. To fill this limited area, this paper conducts a comparative analysis on the variables in leak dataset and proposes an effective method for sample dataset generation, complemented by a novel accuracy evaluation that incorporates the topological relationship among pipelines. A multi-source looped heating network is employed to do the case study. The results demonstrate that the proposed method achieves satisfactory leak localization performance with a significantly reduced training set. Furthermore, by segmenting the physical pipeline into multiple virtual pipelines and using the adjacent evaluation metric, the leak localization method achieves a 75 m error with an accuracy of 0.998.

High-throughput seed quality analysis in faba bean: leveraging Near-InfraRed spectroscopy (NIRS) data and statistical methods

Near-infrared spectroscopy (NIRS) provides a high-throughput phenotyping technique to assist breeding for improved faba bean seed quality. We combined chemical analysis of protein, oil content (and composition) with NIRS through chemometrics, employing Partial Least Squares (PLS), Elastic Net (EN), Memory-based Learning (MBL), and Bayes B (BB) as prediction models. Protein was the most reliably predicted trait (R2 = 0.96–0.98) across field trials, followed by oil (R2 = 0.82–0.86) and oleic acid (R2 = 0.31–0.68). Samples for training the models were selected using K-means clustering. The optimal statistical approach for prediction was compound-specific: PLS for protein (Root Mean Squared Error - RMSE = 0.46), BB for oil (RMSE = 0.067), and EN for oleic acid content (RMSE = 2.83). Reduced training set simulations revealed different effects on prediction accuracy depending on the model and compound. Several NIR regions were pinpointed as highly informative for the compounds, using the shrinkage and variable selection capabilities of EN and BB.

Improving snore detection under limited dataset through harmonic/percussive source separation and convolutional neural networks

Snoring, an acoustic biomarker commonly observed in individuals with Obstructive Sleep Apnoea Syndrome (OSAS), holds significant potential for diagnosing and monitoring this recognized clinical disorder. Irrespective of snoring types, most snoring instances exhibit identifiable harmonic patterns manifested through distinctive energy distributions over time. In this work, we propose a novel method to differentiate monaural snoring from non-snoring sounds by analyzing the harmonic content of the input sound using harmonic/percussive sound source separation (HPSS). The resulting feature, based on the harmonic spectrogram from HPSS, is employed as input data for conventional neural network architectures, aiming to enhance snoring detection performance even under a limited data learning framework. To evaluate the performance of our proposal, we studied two different scenarios: 1) using a large dataset of snoring and interfering sounds, and 2) using a reduced training set composed of around 1% of the data material. In the former scenario, the proposed HPSS-based feature provides competitive results compared to other input features from the literature. However, the key advantage of the proposed method lies in the superior performance of the harmonic spectrogram derived from HPSS in a limited data learning context. In this particular scenario, using the proposed harmonic feature significantly enhances the performance of all the studied architectures in comparison to the classical input features documented in the existing literature. This finding clearly demonstrates that incorporating harmonic content enables more reliable learning of the essential time-frequency characteristics that are prevalent in most snoring sounds, even in scenarios where the amount of training data is limited.

A Comparative Survey of Instance Selection Methods applied to Non-Neural and Transformer-Based Text Classification

Progress in natural language processing has been dictated by the rule of more : more data, more computing power, more complexity, best exemplified by deep learning Transformers. However, training (or fine-tuning) large dense models for specific applications usually requires significant amounts of computing resources. One way to ameliorate this problem is through data engineering instead of the algorithmic or hardware perspectives. Our focus here is an under-investigated data engineering technique, with enormous potential in the current scenario – Instance Selection (IS) (a.k.a. Selective Sampling, Prototype Selection). The IS goal is to reduce the training set size by removing noisy or redundant instances while maintaining or improving the effectiveness (accuracy) of the trained models and reducing the training process cost. We survey classical and recent state-of-the-art IS techniques and provide a scientifically sound comparison of IS methods applied to an essential natural language processing task—Automatic Text Classification (ATC). IS methods have been normally applied to small tabular datasets and have not been systematically compared in ATC. We consider several neural and non-neural state-of-the-art ATC solutions and many datasets. We answer several research questions based on tradeoffs induced by a tripod (training set reduction, effectiveness, and efficiency). Our answers reveal an enormous unfulfilled potential for IS solutions. Specially, we show that in 12 out of 19 datasets, specific IS methods—namely, Condensed Nearest Neighbor, Local Set-based Smoother, and Local Set Border Selector—can reduce the size of the training set without effectiveness losses. Furthermore, in the case of fine-tuning the Transformer methods, the IS methods reduce the amount of data needed, without losing effectiveness and with considerable training-time gains.

Global-local least-squares support vector machine (GLocal-LS-SVM).

This study introduces the global-local least-squares support vector machine (GLocal-LS-SVM), a novel machine learning algorithm that combines the strengths of localised and global learning. GLocal-LS-SVM addresses the challenges associated with decentralised data sources, large datasets, and input-space-related issues. The algorithm is a double-layer learning approach that employs multiple local LS-SVM models in the first layer and one global LS-SVM model in the second layer. The key idea behind GLocal-LS-SVM is to extract the most informative data points, known as support vectors, from each local region in the input space. Local LS-SVM models are developed for each region to identify the most contributing data points with the highest support values. The local support vectors are then merged at the final layer to form a reduced training set used to train the global model. We evaluated the performance of GLocal-LS-SVM using both synthetic and real-world datasets. Our results demonstrate that GLocal-LS-SVM achieves comparable or superior classification performance compared to standard LS-SVM and state-of-the-art models. In addition, our experiments show that GLocal-LS-SVM outperforms standard LS-SVM in terms of computational efficiency. For instance, on a training dataset of 9, 000 instances, the average training time for GLocal-LS-SVM was only 2% of the time required to train the LS-SVM model while maintaining classification performance. In summary, the GLocal-LS-SVM algorithm offers a promising solution to address the challenges associated with decentralised data sources and large datasets while maintaining high classification performance. Furthermore, its computational efficiency makes it a valuable tool for practical applications in various domains.

Reconstruction of high-frequency methane atmospheric concentration peaks from measurements using metal oxide low-cost sensors

Abstract. Detecting and quantifying CH4 gas emissions at industrial facilities is an important goal for being able to reduce these emissions. The nature of CH4 emissions through “leaks” is episodic and spatially variable, making their monitoring a complex task; this is partly being addressed by atmospheric surveys with various types of instruments. Continuous records are preferable to snapshot surveys for monitoring a site, and one solution would be to deploy a permanent network of sensors. Deploying such a network with research-level instruments is expensive, so low-cost and low-power sensors could be a good alternative. However, low cost usually entails lower accuracy and the existence of sensor drifts and cross-sensitivity to other gases and environmental parameters. Here we present four tests conducted with two types of Figaro® Taguchi gas sensors (TGSs) in a laboratory experiment. The sensors were exposed to ambient air and peaks of CH4 concentrations. We assembled four chambers, each containing one TGS sensor of each type. The first test consisted in comparing parametric and non-parametric models to reconstruct the CH4 peak signal from observations of the voltage variations of TGS sensors. The obtained relative accuracy is better than 10 % to reconstruct the maximum amplitude of peaks (RMSE ≤2 ppm). Polynomial regression and multilayer perceptron (MLP) models gave the highest performances for one type of sensor (TGS 2611C, RMSE =0.9 ppm) and for the combination of two sensors (TGS 2611C + TGS 2611E, RMSE =0.8 ppm), with a training set size of 70 % of the total observations. In the second test, we compared the performance of the same models with a reduced training set. To reduce the size of the training set, we employed a stratification of the data into clusters of peaks that allowed us to keep the same model performances with only 25 % of the data to train the models. The third test consisted of detecting the effects of age in the sensors after 6 months of continuous measurements. We observed performance degradation through our models of between 0.6 and 0.8 ppm. In the final test, we assessed the capability of a model to be transferred between chambers in the same type of sensor and found that it is only possible to transfer models if the target range of variation of CH4 is similar to the one on which the model was trained.

Core Temperature Estimation Method for Lithium-Ion Battery Based on Long Short-Term Memory Model With Transfer Learning

Temperature is a crucial parameter that determines the safety and reliability of lithium-ion batteries (LIBs) in electric vehicles and energy storage systems. Estimating LIBs temperature for battery management system state monitoring and thermal control, especially the core temperature (CT), is essential. However, the CT cannot be obtained directly and must be estimated via other measurable variables. To this end, this article proposes a method to estimate the CT of LIBs based on the long short-term memory (LSTM) method combined with transfer learning (TL). Through this method, the relationship between CT and measured variables can be obtained. Moreover, the TL procedure, with its fine-tuning strategy, can substantially reduce the computation burden when used to estimate the CT of other LIBs via a reduced training set. In addition, the applicability of the proposed method is verified via a long period cycle test. The experimental results demonstrate that the proposed method can precisely estimate the LIB CT in ambient temperatures of between −10 °C and 55 °C in highly dynamic driving cycles and aging cycle tests, and the maximum root-mean-square error (RMSE) is 0.3302 °C. The LSTM-TL method has a high accuracy compared with other widely used methods.

Instance selection using one‐versus‐all and one‐versus‐one decomposition approaches in multiclass classification datasets

AbstractInstance is important in data analysis and mining; it filters out unrepresentative, redundant, or noisy data from a given training set to obtain effective model learning. Various instance selection algorithms are proposed in the literature, and their potential and applicability in data cleaning and preprocessing steps are demonstrated. For multiclass classification datasets, the existing instance selection algorithms must deal with all the instances across the different classes simultaneously to produce a reduced training set. Generally, every multiclass classification dataset can be regarded as a complex domain problem, which can be effectively solved using the divide‐and‐conquer principle. In this study, the one‐versus‐all (OVA) and one‐versus‐one (OVO) decomposition approaches were used to decompose a multiclass dataset into multiple binary class datasets. These approaches have been widely employed when constructing the classifier but have never been considered in instance selection. The results of instance selection performance obtained with the OVA, OVO, and baseline approaches were assessed and compared for 20 different domain multiclass datasets as the first study and five medical domain datasets as the validation study. Furthermore, three instance selection algorithms were compared, including IB3, DROP3, and GA. The results demonstrate that using the OVO approach to perform instance selection can make the support vector machine (SVM) and k‐nearest neighbour (k‐NN) classifiers perform significantly better than the OVA and baseline approaches in terms of the area under the ROC curve (AUC) rate, regardless of the instance selection algorithm used. Moreover, the OVO approach can provide reasonably good data reduction rates and processing times, which are all better than those of the OVA approach.

Cross-Image Region Mining With Region Prototypical Network for Weakly Supervised Segmentation

Weakly supervised image segmentation trained with image-level labels usually suffers from inaccurate coverage of object areas during the generation of the pseudo groundtruth. This is because the object activation maps are trained with the classification objective and lack the ability to generalize. To improve the generality of the object activation maps, we propose a region prototypical network ( <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">RPNet</b> ) to explore the cross-image object diversity of the training set. Similar object parts across images are identified via region feature comparison. Object confidence is propagated between regions to discover new object areas while background regions are suppressed. Experiments show that the proposed method generates more complete and accurate pseudo object masks while achieving state-of-the-art performance on PASCAL VOC 2012 and MS COCO. In addition, we investigate the robustness of the proposed method on reduced training sets. The code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/liuweide01/RPNet-Weakly-Supervised-Segmentation</uri> .

Augmented PIN Authentication through Behavioral Biometrics.

Personal Identification Numbers (PINs) are widely used today for user authentication on mobile devices. However, this authentication method can be subject to several attacks such as phishing, smudge, and side-channel. In this paper, we increase the security of PIN-based authentication by considering behavioral biometrics, specifically the smartphone movements typical of each user. To this end, we propose a method based on anomaly detection that is capable of recognizing whether the PIN is inserted by the smartphone owner or by an attacker. This decision is taken according to the smartphone movements, which are recorded during the PIN insertion through the built-in motion sensors. For each digit in the PIN, an anomaly score is computed using Machine Learning (ML) techniques. Subsequently, these scores are combined to obtain the final decision metric. Numerical results show that our authentication method can achieve an Equal Error Rate (EER) as low as 5% in the case of 4-digit PINs, and 4% in the case of 6-digit PINs. Considering a reduced training set, composed of solely 50 samples, the EER only slightly worsens, reaching 6%. The practicality of our approach is further confirmed by the low processing time required, on the order of fractions of milliseconds.

Adaptive surrogates of crashworthiness models for multi-purpose engineering analyses accounting for uncertainty

Uncertainty Quantification (UQ) is a booming discipline for complex computational models based on the analysis of robustness, reliability and credibility. UQ analysis for nonlinear crash models with high dimensional outputs presents important challenges. In crashworthiness, nonlinear structural behaviours with multiple hidden modes require expensive models (18 h for a single run). Surrogate models (metamodels) allow substituting the full order model, introducing a response surface for a reduced training set of numerical experiments. Moreover, uncertain input and large number of degrees of freedom result in high dimensional problems, which derives to a bottle neck that blocks the computational efficiency of the metamodels. Kernel Principal Component Analysis (kPCA) is a multidimensionality reduction technique for non-linear problems, with the advantage of capturing the most relevant information from the response and improving the efficiency of the metamodel. Aiming to compute the minimum number of samples with the full order model. The proposed methodology is tested with a practical industrial problem that arises from the automotive industry.

Accelerated computation of lattice thermal conductivity using neural network interatomic potentials

With the development of the density functional theory (DFT) and ever-increasing computational capacity, an accurate prediction of lattice thermal conductivity based on the Boltzmann transport theory becomes computationally feasible, contributing to a fundamental understanding of thermal conductivity as well as a choice of the optimal materials for specific applications. However, steep costs in evaluating third-order force constants limit the theoretical investigation to crystals with high symmetry and few atoms in the unit cell. Currently, machine learning potentials are garnering attention as a computationally efficient high-fidelity model of DFT, and several studies have demonstrated that the lattice thermal conductivity could be computed accurately via the machine learning potentials. However, test materials were mostly crystals with high symmetries, and the applicability of machine learning potentials to a wide range of materials has yet to be demonstrated. Furthermore, establishing a standard training set that provides consistent accuracy and computational efficiencies across a wide range of materials would be useful. To address these issues, herein we compute lattice thermal conductivities at 300 K using neural network interatomic potentials. As test materials, we select 25 materials with diverse symmetries and a wide range of lattice thermal conductivities between 10−1 and 103 Wm−1K−1. Among various choices of training sets, we find that molecular dynamics trajectories at 50–700 K consistently provide results at par with DFT for the test materials. In contrast to pure DFT approaches, the computational cost in the present approach is uniform over the test materials, yielding a speed gain of 2–10 folds. When a smaller reduced training set is used, the relative efficiency increases by up to ∼50 folds without sacrificing accuracy significantly. The current work will broaden the application scope of machine learning potentials by establishing a robust framework for accurately computing lattice thermal conductivity with machine learning potentials.

Quantitative prediction accuracies derived from laser-induced breakdown spectra using optimized multivariate submodels

The accuracy of laser-induced breakdown spectroscopy (LIBS) methods for analyzing geological samples is improved when calibration standards and unknown targets are compositionally similar. A recent study suggests that customized submodels can be used to optimize calibration datasets to achieve more accurate predictions [1]. In practice, this is difficult to implement because the errors inherent in the methods used for sorting unknown targets by composition may affect how successfully this matching can occur. Moreover, creation of submodels intrinsically reduces the size of the dataset on which the model is trained, which has been shown to reduce prediction accuracy. This paper uses LIBS spectra of 2990 unique rock powder standards to compare the accuracy of 1) submodels generated for each element over its geochemical range, 2) submodels created using SiO2 content only, 3) submodels created using the ratio of Si(II)/Si(I) emission lines to group spectra by a proxy for approximate plasma temperature, and 4) models created using all data. Results indicate that prediction accuracies are not always improved by creating submodels because subdividing a dataset to optimize calibrations will always result in a smaller database available for each submodel, and the reduced training set size negatively affects accuracy. Customized LIBS standards for specific applications might overcome this problem in cases where the matrix is similar and the expected concentration range is known. But in a majority of geochemical applications, submodel approaches are only useful in improving prediction accuracies when the initial database is itself extensive enough to support large, robust submodel calibration suites.

Zero-Shot Cross-Modal Retrieval for Remote Sensing Images With Minimal Supervision

The performance of a deep-learning-based model primarily relies on the diversity and size of the training dataset. However, obtaining such a large amount of labeled data for practical remote sensing applications is expensive and labor-intensive. Training protocols have been previously proposed for few-shot learning (FSL) and zero-shot learning (ZSL). However, FSL is not compatible with handling unobserved class data at the inference phase, while ZSL requires many training samples of the seen classes. In this work, we propose a novel training protocol for image retrieval and name it as <i>label-deficit zero-shot learning</i> (LDZSL). We use this novel LDZSL training protocol for the challenging task of cross-sensor data retrieval in remote sensing. This protocol uses very few labeled data samples of the seen classes during training and interprets unobserved class data samples at the inference phase. This strategy is critical as some data modalities are hard to annotate without domain experts. This work proposes a novel bi-level Siamese network to perform the LDZSL cross-sensor retrieval of multispectral and SAR images. We utilize the available geo-referenced SAR and multispectral data to domain align the embedding features of the two modalities. We experimentally demonstrate the proposed model&#x2019;s efficacy using the So2Sat dataset compared to the existing state-of-the-art models of the ZSL framework trained under a reduced training set. We also show the generalizability of the proposed model using a sketch-based image retrieval task. Experimental results on the Earth on Canvas dataset exhibit comparative performance over the literature.

Selective ensemble of classifiers trained on selective samples

Classifier ensembles are characterized by the high quality of classification, thanks to their generalizing ability. Most existing ensemble algorithms use all learning samples to learn the base classifiers that may negatively impact the ensemble’s diversity. Also, the existing ensemble pruning algorithms often return suboptimal solutions that are biased by the selection criteria. In this work, we present a proposal to alleviate these drawbacks. We employ an instance selection method to query a reduced training set that reduces both the space complexity of the formed ensemble members and the time complexity to classify an instance. Additionally, we propose a guided search-based pruning schema that perfectly explores large-size ensembles and brings on a near-optimal subensemble with less computational requirements in reduced memory space and improved prediction time. We show experimentally how the proposed method could be an alternative to large-size ensembles. We demonstrate how to form less-complex, small-size, and high-accurate ensembles through our proposal. Experiments on 25 datasets show that the proposed method can produce effective ensembles better than Random Forest and baseline classifier pruning methods. Moreover, our proposition is comparable with the Extreme Gradient Boosting Algorithm in terms of accuracy.

On the generation of multi-label prototypes

Data reduction techniques play a key role in instance-based classification to lower the amount of data to be processed. Prototype generation aims to obtain a reduced training set in order to obtain accurate results with less effort. This translates into a significant reduction in both algorithms’ spatial and temporal burden. This issue is particularly relevant in multi-label classification, which is a generalization of multiclass classification that allows objects to belong to several classes simultaneously. Although this field is quite active in terms of learning algorithms, there is a lack of data reduction methods. In this paper, we propose several prototype generation methods from multi-label datasets based on Granular Computing. The simulations show that these methods significantly reduce the number of examples to a set of prototypes without significantly affecting classifiers’ performance.

Enhanced prediction of anti-tubercular peptides from sequence information using divergence measure-based intuitionistic fuzzy-rough feature selection

Tuberculosis is one of the leading causes of millions of deaths across the world, mainly due to growth of drug-resistant strains. Anti-tubercular peptides may facilitate an alternate way to combat antibiotic tolerance. This study describes a novel approach for enhancing the prediction of anti-tubercular peptides by feature extraction from sequence of the peptides, selection of optimal features from the extracted features, and selection of suitable learning algorithm. Firstly, we extract different sequence features by using iFeature web server. Then, the optimal features are obtained by using a novel divergence measure-based intuitionistic fuzzy rough sets-assisted feature selection technique. Furthermore, an attempt has been made to develop models using different machine learning techniques for enhancing the prediction of anti-tubercular (or anti-mycobacterial peptides) with other antibacterial peptides (ABP) as well non-antibacterial peptides (non-ABP). Moreover, the best prediction result is obtained by vote-based classifier. Using 80:20 percentage split, the proposed method performs well, with sensitivity of 92.0%, 96.4%, specificity of 83.3%, 88.4%, overall accuracy of 87.80%, 92.90%, Mathews correlation coefficient of 0.757, 0.857, AUC of 0.922, 0.914, and g-means of 87.5%, 92.3% for anti-tubercular and ABP (primary dataset), anti-tubercular and non-ABP (secondary dataset), respectively. Finally, we have evaluated the performances of different machine learning algorithms by using the reduced training sets as produced by our proposed feature selection technique as well as already existing intuitionistic fuzzy rough set based and ensemble feature selection technique. Moreover, the performance of our proposed approach is evaluated on few benchmark and AMP datasets. From the experimental results, it can be observed that our proposed method is outperforming the previous methods.

Decision fusion using virtual dictionary‐based sparse representation for robust SAR automatic target recognition

Sparse representation displays remarkable characteristics when applied to image processing and classification. The critical point in the success of sparse representation-based classification is to learn an authentic dictionary. The present study proposes a virtual dictionary-based sparse representation for automatic target recognition. Based on the properties of the synthetic aperture radar (SAR) images, some low complexity modules including adding speckle noise, histogram equalisation mapping, and bicubic interpolation are applied to construct some virtual compact dictionaries using Fisher discriminative dictionary learning. These dictionaries have different discriminative information on targets, which are used independently in several sparse representation-based classifiers. The reconstruction error vectors of the latter classifiers are then combined to recognise the target using decision fusion. Based on experimental results obtained drawing upon moving and stationary target acquisition and recognition data set, the proposed method presents the highest accuracy in classification reported yet in the literature. Furthermore, the procedure improves the recognition robustness against most commonly extended operating conditions, e.g. speckle noise corruption, depression angle variation and reduced training set. Accordingly, the current study claims a robust parallel method of high real-time ability in the target recognition of SAR images applicable to practical situations.

On a Hybridization of Deep Learning and Rough Set Based Granular Computing

The set of heuristics constituting the methods of deep learning has proved very efficient in complex problems of artificial intelligence such as pattern recognition, speech recognition, etc., solving them with better accuracy than previously applied methods. Our aim in this work has been to integrate the concept of the rough set to the repository of tools applied in deep learning in the form of rough mereological granular computing. In our previous research we have presented the high efficiency of our decision system approximation techniques (creating granular reflections of systems), which, with a large reduction in the size of the training systems, maintained the internal knowledge of the original data. The current research has led us to the question whether granular reflections of decision systems can be effectively learned by neural networks and whether the deep learning will be able to extract the knowledge from the approximated decision systems. Our results show that granulated datasets perform well when mined by deep learning tools. We have performed exemplary experiments using data from the UCI repository—Pytorch and Tensorflow libraries were used for building neural network and classification process. It turns out that deep learning method works effectively based on reduced training sets. Approximation of decision systems before neural networks learning can be important step to give the opportunity to learn in reasonable time.

Optimisation of cancer classification by machine learning generates an enriched list of candidate drug targets and biomarkers.

The Cancer Genome Atlas has provided expression values of 18 015 genes for different cancer types. Studies on the classification of cancers by machine learning algorithms have used different data and methods, which makes it difficult to compare their performance. It is unclear, which algorithm performs best and if maximum levels of accuracy have been obtained. In this study, we aimed to optimise the diagnosis of cancer by comparing the performance of five algorithms using the same data, and by identifying the smallest possible number of differentiator genes. Classification accuracies of five algorithms of cancer type and primary site were determined using a gene expression dataset of 5629 samples and a dataset of 9144 samples, respectively. When trained with sample sets ranging from 16 718 to 40 genes, Random Forest (RF), Gradient Boosting Machine (GBM), and Neural Network (NN) consistently achieved 100% or near 100% accuracy in the classification of both cancer type and primary site. Reduction of training sets to the 40 highest-ranked genes resulted in 78-fold and 45-fold faster processing times for RF and GBM, respectively. The olfactory receptor family, keratin associated proteins, and defensin beta family were among the highest ranked genes. The ensemble and NN algorithms were the most accurate at distinguishing between cancer types and primary sites, whereas KNN was the fastest. Training sets can be reduced to the 40 highest-ranked differentiator genes without any significant loss of accuracy, amongst which there are potential drug targets and biomarkers.