No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
Linear discriminant functions
... Support Vector Machine (SVM) has been successfully applied to classification on many occasions as a machine learning algorithm. It aims at maximizing the margin between the separating hyperplane and the data to minimize an upper bound of the generalization error [42], [43]. The advantages of SVM are effective in high dimensional spaces and still effective in cases where number of dimensions is greater than the number of samples. ...
... In addition, the kernel trick in SVM algorithm is used to build in expert knowledge about the problem via engineering the kernel. The more detailed information about SVM is presented in [42]. In our work, ...
... To avoid the overfitting phenomenon in machine learning, 10-fold cross-validation method for SVM were conducted [45]. The SVM model performance was evaluated by sensitivity (Se), specificity (Sp) and accuracy (Acc) referred to [42]. Sensitivity (Se): (true positive rate) ...
It is well known that electrocardiogram heartbeats are substantial for cardiac disease diagnosis. In this study, the best time scale was investigated to recognize congestive heart failure (CHF) based on heart rate variability (HRV) measures. The classifications were performed on seven different time scales with a support vector machine (SVM) classifier. The nine HRV measures including three time-domain measures, three frequency-domain measures and three nonlinear-domain measures were took as feature vectors for classifier on each time scale. A total of 83 subjects with RR intervals are analyzed of which 54 cases are normal and 29 patients are suffering from congestive heart failure in PhysioNet databases. The classifying results using 10-folds cross validation method achieved the best performance of sensitivity, specificity and accuracy values of 86.7%, 98.3% and 94.4%, respectively, on 2-hour time scale. Moreover, by introducing only three nonstandard standard deviation of NN intervals (SDNN) fitting and vector length index of Poincare plot (VLI) fitting features it could achieve better sensitivity, specificity and accuracy values of 93.3%, 98.3% and 96.7%, respectively. The impressive performance of discrimination power on 2-hour time scale and the trends of HRV measures on the time scales prove that multiple time scales play significant roles in detecting CHF and can be valuable in expressing useful knowledge in medicine.
... These classification methods were selected for their simplicity, robustness, and abilities to fit highdimensional data for binary classification problems, such as the ones proposed in this study (Varmuza and Filzmoser, 2016). LDA is a linear probabilistic classifier that receives an input feature vector and based on the value of the discriminant function, f(), the sample is assigned to class C 1 if f()≥ 0 and to class C 2 otherwise (Bishop, 2006;Duda et al., 2012). K-nearest Neighbour (KNN) is a non-parametric classification technique where each sample point is assigned to a specific class based on its distance to the nearest points within the feature space (Bishop, 2007;Duda et al., 2012). ...
... LDA is a linear probabilistic classifier that receives an input feature vector and based on the value of the discriminant function, f(), the sample is assigned to class C 1 if f()≥ 0 and to class C 2 otherwise (Bishop, 2006;Duda et al., 2012). K-nearest Neighbour (KNN) is a non-parametric classification technique where each sample point is assigned to a specific class based on its distance to the nearest points within the feature space (Bishop, 2007;Duda et al., 2012). Based on preliminary analysis, the k value was chosen in this study to be 4, and the Euclidean distance was considered between each sample point and the centre of each class. ...
Food contamination is a major concern for consumers and food businesses, especially when the contaminant is an allergen. This study focused on detecting and quantifying peanut powder in garlic powder using low-cost Near-Infrared sensors (S2.0-1550-1950 nm, and S2.5-2000-2450 nm) coupled with machine learning methods. Garlic and peanut powders of three different origins were allocated to different data sets, and 37 peanut cotamination concentrations from 0.01-20% were studied. Samples were first assessed to determine if they were contaminated with peanut. Peanut concentration was then determined to be either low (0.01-1%) or high (2-20%). Finally, the peanut concentration was predicted. Classification accuracy of 100% was achieved when assessing models on an individual data set but declined when a second independent data set was used. In general, the models developed from the S2.0 sensor performed marginally better than those developed using the S2.5. Peanut concentration prediction models achieved Correlation Coefficient and Root Mean Square Error values of 70.8%, 0.49%, and 77.5%, 3.53% for low and high peanut concentrations, respectively. The results obtained from this study can be used to develop cost-effective contamination detection technologies for the food sector.
... The LDA classifier is simple to implement and relatively fast to train compared with artificial neural networks [55]. The classifier includes a linear combination of the features and is based on the value of the discriminant function before the sample (i.e., image) is classified into the appropriate class [56]. Ensemble techniques are based on optimizing the classification performance by weighing several individual classifiers and combining them to obtain a classifier that outperforms the single classifier [57]. ...
... In the case Table 6 Regression results using RT and cross-validation for quantifying adulterant level using all or selected features along with regression trees. and this could be decreased further by testing a larger number of samples, which would reduce training errors [56]. Current results indicate that using all features for predicting the adulteration level was the best method as the models had lower errors. ...
Meat products are popular foods and there is a need for cost-effective technologies for rapid quality assessment. In this study, RGB color imaging coupled with machine learning algorithms were investigated to detect plant and animal adulterants with ratios of from 1-50% in minced meat. First, samples were classified as either pure or adulterated, then adulterated samples were classified based on the adulterant's type. Finally, regression models were developed to predict the adulteration quantity. Linear discriminant classifier enhanced by bagging ensembling performed the best with overall classification accuracies for detecting pure or adulterated samples up to 99.1% using all features, and 100% using selected features. Classification accuracies for adulteration origin were 48.9–76.1% using all features and 63.8% for selected features. Regression trees were used for adulterant level quantification and the r (RPD) values were up to 98.0%(5.0) based on all features, and 94.5%(3.2) for selected features. Gray-level and co-occurrence features were more effective than other color channels in building classification and regression models. This study presents a non-invasive, and low-cost system for adulteration detection in minced meats.
... The source code of the proposed algorithm is available at: https://github.com/cil6758/ATTRACTIVE. The training algorithm included four parts -dataset collection, text preprocessing, feature extraction, and building learning models with a linear discriminant function (LDF) [19] (Fig. 2). Two organ/tissue classification models, one using titles and the other using abstracts, were built for comparison. ...
... To improve the static organ/tissue category model, LDFs [19] were used. First, one article was chosen from the training dataset and cosine similarity between the selected article and every organ/tissue model was calculated, and the article was assigned to the category with the highest score. ...
Many research articles are published on regenerative medicine every year. However, only a small proportion of these articles provide experimental methods on organ/tissue differentiation. Therefore, we developed a database – ATTRACTIVE (An auTo-updating daTabase foR experimentAl protoCols in regeneraTIVe mEdicine) – that collects journal articles with differentiation methods in regenerative medicine and updates itself automatically on a regular basis. Since the number of articles in regenerative medicine was insufficient and unbalanced, which limited the performance of the supervised learning algorithms, we proposed an algorithm that combines cosine similarity and linear discriminant functions to classify articles based on their titles and abstracts more efficiently. The results show that our proposed methods out-performed other machine learning algorithms such as k-nearest neighbors, support vector machine, and long short-term memory methods. The classification accuracy reached 94.62%, even with a small and unbalanced dataset. Lastly, we incorporated our classifier into the database for automatic updates. The database is available at http://attractive.cgm.ntu.edu.tw/.
... To this end, we used linear discriminant analysis (LDA) to estimate the accuracy of GVHD predictive genes discovered in microarray and quantitative real-time (qRT)PCR experiments[28]. In addition, we assessed the robustness for all the genes validated by qRT-PCR by performing 500 independent instances of training-test dataset splits crossvalidation to determine empirically through computational resampling the expected generalizable class-prediction accuracy on independent test datasets[29,30]. In LDA with assumed equal class a priori probabilities, the boundary between class P (GVHDþ) and class N (GVHDÀ) is determined by the value of the separatrix, S, which is the point (in univariate analysis) between the class P and N means that is equidistant to both[28]. ...
... However, a single instance of training-test dataset comparison can be considered neither representative nor robust, since it is potentially sensitive to idiosyncratic fluctuations of datapoints around the separatrix. We therefore determined the robust average accuracy over many independently generated test datasets for each gene, on the basis of different selections of training-set data for each gene[30], using conventional cross-validation procedures[29]. These analyses were performed on the 17 single genes (Table 2) and the PIA variables representative of the four gene pairs (Figure 4A) that were predictive of cGVHD occurrence. ...
Graft-versus-host disease (GVHD) results from recognition of host antigens by donor T cells following allogeneic hematopoietic cell transplantation (AHCT). Notably, histoincompatibility between donor and recipient is necessary but not sufficient to elicit GVHD. Therefore, we tested the hypothesis that some donors may be "stronger alloresponders" than others, and consequently more likely to elicit GVHD.
To this end, we measured the gene-expression profiles of CD4(+) and CD8(+) T cells from 50 AHCT donors with microarrays. We report that pre-AHCT gene-expression profiling segregates donors whose recipient suffered from GVHD or not. Using quantitative PCR, established statistical tests, and analysis of multiple independent training-test datasets, we found that for chronic GVHD the "dangerous donor" trait (occurrence of GVHD in the recipient) is under polygenic control and is shaped by the activity of genes that regulate transforming growth factor-beta signaling and cell proliferation.
These findings strongly suggest that the donor gene-expression profile has a dominant influence on the occurrence of GVHD in the recipient. The ability to discriminate strong and weak alloresponders using gene-expression profiling could pave the way to personalized transplantation medicine.
... We apply the Fisher's linear discriminant analysis (LDA) [8] to determine whether the given q q q corresponds to the presence of PLI. For this purpose, we compute r = w w w T q q q + w 0 ...
... Ni(q q q k 0 m m mi)(q q q k 0 m m mi) T : (8) The feature vectors acquired from the ECG signals adopted from MIT-BIH arrhythmia database [10] with supervised PLI corruption are clustered to generate w w w, w0, and S S S W W W of the pattern classifier. In Fig. 3, we illustrated the feature vectors as well as the decision boundary derived using LDA. ...
... The signals are then downsampled by a factor of 4. The xDAWN spatial filter [44] was used to reduce the 32 EEG channels to 3 xDAWN components that maximise the difference between target and non-target trials. A Linear Discriminant Analysis (LDA) classifier [45] uses epochs from stimulus-onset to 600 ms post-stimulus to determine the target row and column, i.e. the row and column that contain the letter the user is focusing on. Both the xDAWN spatial filter and the LDA classifier are trained for each participant at the beginning of the session by using the EEG data recorded from copyspelling runs. ...
Objective. Neurofeedback training through brain-computer interfacing has demonstrated efficacy in treating neurological deficits and diseases, and enhancing cognitive abilities in healthy individuals. It was previously shown that event-related potential (ERP)-based neurofeedback training using a P300 speller can improve attention in healthy adults by incrementally increasing the difficulty of the spelling task. This study aims to assess the impact of task difficulty adaptation on ERP-based attention training in healthy adults. To achieve this, we introduce a novel adaptation employing iterative learning control and compare it against an existing method and a control group with random task difficulty variation. Approach. The study involved 45 healthy participants in a single-blind, 3-arm randomized controlled trial. Each group underwent one neurofeedback training session, using different methods to adapt task difficulty in a P300 spelling task: two groups with personalised difficulty adjustments (our proposed iterative learning control and an existing approach) and one group with random difficulty. Cognitive performance was evaluated before and after the training session using a visual spatial attention task and we gathered participant feedback through questionnaires. Main results. All groups demonstrated a significant performance improvement in the spatial attention task post-training, with an average increase of 12.63%. Notably, the group using the proposed iterative learning controller achieved a 22% increase in P300 amplitude during training and a 17% reduction in post-training alpha power, all while significantly accelerating the training process compared to other groups. Significance. Our results suggest that ERP-based neurofeedback training using a P300 speller effectively enhances attention in healthy adults, with significant improvements observed after a single session. Personalised task difficulty adaptation using iterative learning control not only accelerates the training but also enhances ERPs during the training. Accelerating neurofeedback training, while maintaining its effectiveness, is vital for its acceptability by both end-users and clinicians.
... : LDA, also known as Fisher's linear discriminant, creates a classification method based on a Gaussian distribution for data component x using the discriminant function [20] g ...
... The data were forward and backward filtered using a 3-50 Hz bandpass filter to remove movement and synchronous breathing and heart artifacts, as well as high frequency electronic noise and 60 Hz powerline. In addition, Independent Component Analysis (ICA) was used to remove cardiac artifact from the MEG data and singular value decomposition (SVD) method was used to eliminate high amplitude artifacts associated with head, eye and mouth movement (Duda and Hart 1973;Tufts et al. 1982). All data were visually inspected for epileptic spikes by board-certified neurophysiologists. ...
Magnetoencephalography (MEG) is a noninvasive imaging method for localization of focal epileptiform activity in patients with epilepsy. Diffusion tensor imaging (DTI) is a noninvasive imaging method for measuring the diffusion properties of the underlying white matter tracts through which epileptiform activity is propagated. This study investigates the relationship between the cerebral functional abnormalities quantified by MEG coherence and structural abnormalities quantified by DTI in mesial temporal lobe epilepsy (mTLE). Resting state MEG data was analyzed using MEG coherence source imaging (MEG-CSI) method to determine the coherence in 54 anatomical sites in 17 adult mTLE patients with surgical resection and Engel class I outcome, and 17 age- and gender- matched controls. DTI tractography identified the fiber tracts passing through these same anatomical sites of the same subjects. Then, DTI nodal degree and laterality index were calculated and compared with the corresponding MEG coherence and laterality index. MEG coherence laterality, after Bonferroni adjustment, showed significant differences for right versus left mTLE in insular cortex and both lateral orbitofrontal and superior temporal gyri (p < 0.017). Likewise, DTI nodal degree laterality, after Bonferroni adjustment, showed significant differences for right versus left mTLE in gyrus rectus, insular cortex, precuneus and superior temporal gyrus (p < 0.017). In insular cortex, MEG coherence laterality correlated with DTI nodal degree laterality ([Formula: see text] in the cases of mTLE. None of these anatomical sites showed statistically significant differences in coherence laterality between right and left sides of the controls. Coherence laterality was in agreement with the declared side of epileptogenicity in insular cortex (in 82 % of patients) and both lateral orbitofrontal (88 %) and superior temporal gyri (88 %). Nodal degree laterality was also in agreement with the declared side of epileptogenicity in gyrus rectus (in 88 % of patients), insular cortex (71 %), precuneus (82 %) and superior temporal gyrus (94 %). Combining all significant laterality indices improved the lateralization accuracy to 94 % and 100 % for the coherence and nodal degree laterality indices, respectively. The associated variations in diffusion properties of fiber tracts quantified by DTI and coherence measures quantified by MEG with respect to epileptogenicity possibly reflect the chronic microstructural cerebral changes associated with functional interictal activity. The proposed methodology for using MEG and DTI to investigate diffusion abnormalities related to focal epileptogenicity and propagation may provide a further means of noninvasive lateralization.
... A linear discriminant analysis (LDA) (Duda et al. 2001) classifier in the feature space was computed between the active period for the right hand MI and its corresponding baseline period for each of the 0.4 s segments. To validate the classifiers, the leave-one-trial-out cross validation approach was adopted due to the relatively small number of trials. ...
Objective
Motor Imagination (MI) and Functional Electrical Stimulation (FES) can activate the sensory-motor cortex through efferent and afferent pathways respectively. Motor Imagination can be used as a control strategy to activate FES through a Brain-computer interface as the part of a rehabilitation therapy. It is believed that precise timing between the onset of MI and FES is important for strengthening the cortico-spinal pathways but it is not known whether prolonged MI during FES influences cortical response.
Methods
Electroencephalogram was measured in ten able-bodied participants using MI strategy to control FES through a BCI system. Event related synchronisation/desynchronisation (ERS/ERD) over the sensory-motor cortex was analysed and compared in three paradigms: MI before FES, MI before and during FES and FES alone activated automatically.
Results
MI practiced both before and during FES produced strongest ERD. When MI only preceded FES it resulted in a weaker beta ERD during FES than when FES was activated automatically. Following termination of FES, beta ERD returns to the baseline level within 0.5 s while alpha ERD took longer than 1 s.
Conclusions
When MI and FES are combined for rehabilitation purposes it is recommended that MI is practiced throughout FES activation period.
Significance
The study is relevant for neurorehabilitation of movement.
... Therefore, W is over-determined and no unique solution exists. In this project, the minimum squared error (MSE) technique was adopted(Duda and Hart, 1973) to approximate W. The MSE procedure minimizes the squared error between Y and XW. Using this procedure, the pseudoinverse of X (X"^) is computed by X"*" = (X^X)""^X^.where ...
... We choose the Fisher s linear discriminant analysis (LDA) in this study for its simplicity and computational efficiency. The LDA classifier is briefly described as follows [6]. ...
A novel power-line interference (PLI) detection and suppression algorithm is proposed to pre-process real-time electrocardiogram (ECG) signals. This algorithm first compares the energy at the harmonic frequency against the energy at neighboring frequencies of the ECG power spectrum, and employs an optimal linear discriminant analysis (LDA) algorithm to determine whether PLI interference exists in the ECG signal. If the presence of PLI is detected, it then applies a recursive least square (RLS) adaptive notch filter to suppress the interference. Extensive simulation results indicate that the algorithm consistently exhibits superior performance in terms of less ECG distortion, faster convergence rate and numerical stability.
... A further reduction of the 35 gene signature was performed by ranking using Relief methodology (Kira and Rendel, 1992;Kononeko, 1994) and the top seven genes were selected as the predictor signature genes. The predictor signature thus identified was tested by using linear discriminant analysis (Duda, 2000) in the data obtained from a test set of rats that were exposed to lower concentrations of silica (1 or 2 mg/m 3 , 6 h/ day, 5 days). ...
Blood gene expression profiling was investigated as a minimally invasive surrogate approach to detect silica exposure and resulting pulmonary toxicity. Rats were exposed by inhalation to crystalline silica (15 mg/m³, 6 h/day, 5 days), and pulmonary damage and blood gene expression profiles were determined after latency periods (0-16 weeks). Silica exposure resulted in pulmonary toxicity as evidenced by histological and biochemical changes in the lungs. The number of significantly differentially expressed genes in the blood, identified by microarray analysis, correlated with the severity of silica-induced pulmonary toxicity. Functional analysis of the differentially expressed genes identified activation of inflammatory response as the major biological signal. Induction of pulmonary inflammation, as suggested by the blood gene expression data, was supported by significant increases in the number of macrophages and infiltrating neutrophils as well as the activity of pro-inflammatory chemokines observed in the lungs of the silica-exposed rats. A gene expression signature developed using the blood gene expression data predicted the exposure of rats to lower, minimally toxic and nontoxic concentrations of silica. Taken together, our findings suggest the potential application of peripheral blood gene expression profiling as a minimally invasive surrogate approach to detect pulmonary toxicity induced by silica in the rat. However, further research is required to determine the potential application of our findings specifically to monitor human exposure to silica and the resulting pulmonary effects.
... Linear classification is robust and interpretable, and a linear classifier was chosen (Duda, 2001a). In addition, exclusion of non-linearity created situation of under-training and limited over-fit concerns. ...
... It was also hypothesized thatspatialfrequencies observedwithinhealthyand unhealthy tendons in ultrasound images would vary sufficiently to enable use as a method for discriminating and categorizing between such groups. LDA is a method for classification of data by finding the linear combination of features (discriminant functions) that best separate two or more classes of objects [28]. This tool can be used when the underlying probability distribution of the parameters are not known, and is useful as an initial classifier if the data do not suggest otherwise. ...
The structural characteristics of a healthy tendon are related to the anisotropic speckle patterns observed in ultrasonic images. This speckle orientation is disrupted upon damage to the tendon structure as observed in patients with tendinopathy. Quantification of the structural appearance of tendon shows promise in creating a tool for diagnosing, prognosing, or measuring changes in tendon organization over time. The current work describes a first step taken towards this goal-classification of Achilles tendon images into tendinopathy and control categories. Eight spatial frequency parameters were extracted from regions of interest on tendon images, filtered and classified using linear discriminant analysis. Resulting algorithms had better than 80% accuracy in categorizing tendon image kernels as tendinopathy or control. Tendon images categorized wrongly provided for an interesting clinical association between incorrect classification of tendinopathy kernels as control and the symptom and clinical time history based inclusion criteria. To assess intersession reliability of image acquisition, the first 10 subjects were imaged twice during separate sessions. Test-retest of repeated measures was excellent (tau = 0.996, ICC = (2, 1) = 0.73 with one outlier) indicating a general consistency in imaging techniques.
In the era of Big Data, as vast amounts of data are collected and shared among collaborators or uploaded to the Internet, the attacks on data privacy become more and more serious. Compressive privacy (CP) is a kind of privacypreserving dimension-reduced projection schemes such that the projected data can be well used for the intended utility task but not for malicious applications. Nevertheless, most of existing CP approaches belong to centralized processing, which are not applicable to the cases that data is dispersedly collected/stored at distributed nodes and cannot be centralized to one node for processing due to various reasons. To tackle this problem, we propose a distributed differential utility/cost analysis (dDUCA), in which each node in the network is only allowed to exchange and combine the compressive-and-lossy projection matrix with its one-hop neighbors. Using the projection matrix, the classification of the projected data is performed. Experiments on several datasets confirm the effectiveness of the proposed method in terms of both privacy protection and utility retention.
Near-infrared (NIR) diffuse reflectance has been extensively and successfully applied on quality assurance for fruits, vegetables, and food products. This study is principally aimed to extract the primary wavelengths related to the prediction of glucose and sucrose for potato tubers (of Frito Lay 1879 (FL), a chipping cultivar, and Russet Norkotah (RN), a table use cultivar, and investigating the potential of classification of potatoes based on sugar levels important to the frying industry. Whole tubers, as well as 12.7 mm slices, were scanned using a NIR reflectance spectroscopic system (900–1685 nm). To extract the most influential wavelength in the studied range, interval partial least squares (IPLS), and genetic algorithm (GA) were utilized. Partial least squares regression (PLSR) was applied for building prediction models. Prediction models for RN showed stronger correlation than FL with r(RPD) (correlation coefficient (ratio of reference standard deviation to root mean square error of the model)) values for whole tubers for glucose being as high as 0.81(1.70), and 0.97(3.91) for FL and RN; in the case of sliced samples the values were 0.74(1.49) and 0.94(2.73) for FL and RN. Lower correlation was obtained for sucrose with r(RPD) for whole tubers as high as 0.75(1.52), 0.92(2.57) for FL and RN; and the values for sliced samples were 0.67(1.31) and 0.75(1.41) for FL and RN respectively. Classification of potatoes based on sugar levels was conducted and training models were built using different classifiers (linear discriminant analysis (LDA), K-nearest neighbor (Knn), partial least squares discriminant analysis (PLSDA), and artificial neural network (ANN)), in addition to classifier fusion. To obtain more robust classification models for the training data, 4-fold cross validation was used and results were tested using separate sets of data. Classification rates of the testing set for whole tubers, based on glucose, were as high as 81% and 100% for FL and RN. For sliced samples, the rates were 83% and 81% for FL and RN. Generally, lower classification rates were obtained based on sucrose with values of whole tubers of 71%, and 79% for FL and RN, and for sliced samples the rates were 75%, and 82% which follows a similar trend as PLSR results. This study presents a potential of using selected wavelengths and NIR reflectance spectroscopy to effectively evaluate the sugar content of potatoes and classify potatoes based on thresholds that are crucial for the frying industry.
The aim of this study was to test how the presence of central neuropathic pain (CNP) influences the performance of a motor imagery based Brain Computer Interface (BCI).
In this electroencephalography (EEG) based study, we tested BCI classification accuracy and analysed event related desynchronisation (ERD) in 3 groups of volunteers during imagined movements of their arms and legs. The groups comprised of nine able-bodied people, ten paraplegic patients with CNP (lower abdomen and legs) and nine paraplegic patients without CNP. We tested two types of classifiers: a 3 channel bipolar montage and classifiers based on common spatial patterns (CSPs), with varying number of channels and CSPs.
Paraplegic patients with CNP achieved higher classification accuracy and had stronger ERD than paraplegic patients with no pain for all classifier configurations. Highest 2-class classification accuracy was achieved for CSP classifier covering wider cortical area: 82±7% for patients with CNP, 82±4% for able-bodied and 78±5% for patients with no pain.
Presence of CNP improves BCI classification accuracy due to stronger and more distinct ERD.
Results of the study show that CNP is an important confounding factor influencing the performance of motor imagery based BCI based on ERD.
Copyright © 2015 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved.
A novel multi-hand tracking algorithm for human computer interface (HCI) is proposed, to solve the problem caused by the incompletion of tracking, the overlap of multi-object, the lost of the object and the confusion from background. For tracking the hand completely and effectively, we propose a new region-connection-based tracking method. And a hand motion model is established for tracking failure detection, with the utility of Bayesian classifier. Furthermore, to recover tracking after failure, a monitoring windows method is designed.
Spam mail recognition is a new growing field which brings together the topic of natural language processing and machine learning
as it is in essence a two class classification of natural language texts. An important feature of spam recognition is that
it is a cost-sensitive classification: misclassification of a nonspam mail as spam is generally a more severe error than misclassifying
a spam mail as non-spam. In order to be compared, the methods applied to this field should be all evaluated with the same
corpus and within the same cost-sensitive framework. In this paper, the performances of Support Vector Machines (SVM), Neural
Networks (NN) and Naϊve Bayes (NB) techniques are compared using a publicly available corpus (LINGSPAM) for different cost
scenarios. The training time complexities of the methods are also evaluated. The results show that NN has significantly better
performance than the two other, having acceptable training times. NB gives better results than SVM when the cost is extremely
high while in all other cases SVM outperforms NB.
ResearchGate has not been able to resolve any references for this publication.