Comment

The support vector machine (SVM) has played an important role in bringing certain themes to the fore in computationally oriented statistics. However, it is important to place the SVM in context as but one member of a class of closely related algorithms for nonlinear classification. As we discuss, several of the “open problems” identified by the authors have in fact been the subject of a significant literature, a literature that may have been missed because it has been aimed not only at the SVM but at a broader family of algorithms. Keeping the broader class of algorithms in mind also helps to make clear that the SVM involves certain specific algorithmic choices, some of which have favorable consequences and others of which have unfavorable consequences—both in theory and in practice. The broader context helps to clarify the ties of the SVM to the surrounding statistical literature.

Prediction of cis/trans isomerization in proteins using PSI-BLAST profiles and secondary structure information

Background The majority of peptide bonds in proteins are found to occur in the trans conformation. However, for proline residues, a considerable fraction of Prolyl peptide bonds adopt the cis form. Proline cis/trans isomerization is known to play a critical role in protein folding, splicing, cell signaling and transmembrane active transport. Accurate prediction of proline cis/trans isomerization in proteins would have many important applications towards the understanding of protein structure and function. Results In this paper, we propose a new approach to predict the proline cis/trans isomerization in proteins using support vector machine (SVM). The preliminary results indicated that using Radial Basis Function (RBF) kernels could lead to better prediction performance than that of polynomial and linear kernel functions. We used single sequence information of different local window sizes, amino acid compositions of different local sequences, multiple sequence alignment obtained from PSI-BLAST and the secondary structure information predicted by PSIPRED. We explored these different sequence encoding schemes in order to investigate their effects on the prediction performance. The training and testing of this approach was performed on a newly enlarged dataset of 2424 non-homologous proteins determined by X-Ray diffraction method using 5-fold cross-validation. Selecting the window size 11 provided the best performance for determining the proline cis/trans isomerization based on the single amino acid sequence. It was found that using multiple sequence alignments in the form of PSI-BLAST profiles could significantly improve the prediction performance, the prediction accuracy increased from 62.8% with single sequence to 69.8% and Matthews Correlation Coefficient (MCC) improved from 0.26 with single local sequence to 0.40. Furthermore, if coupled with the predicted secondary structure information by PSIPRED, our method yielded a prediction accuracy of 71.5% and MCC of 0.43, 9% and 0.17 higher than the accuracy achieved based on the singe sequence information, respectively. Conclusion A new method has been developed to predict the proline cis/trans isomerization in proteins based on support vector machine, which used the single amino acid sequence with different local window sizes, the amino acid compositions of local sequence flanking centered proline residues, the position-specific scoring matrices (PSSMs) extracted by PSI-BLAST and the predicted secondary structures generated by PSIPRED. The successful application of SVM approach in this study reinforced that SVM is a powerful tool in predicting proline cis/trans isomerization in proteins and biological sequence analysis.

Application of higher order spectra to identify epileptic EEG

Epilepsy is characterized by the spontaneous and seemingly unforeseeable occurrence of seizures, during which the perception or behavior of patients is disturbed. An automatic system that detects seizure onsets would allow patients or the people near them to take appropriate precautions, and could provide more insight into this phenomenon. Various methods have been proposed to predict the onset of seizures based on EEG recordings. The use of nonlinear features motivated by the higher order spectra (HOS) has been reported to be a promising approach to differentiate between normal, background (pre-ictal) and epileptic EEG signals. In this work, we made a comparative study of the performance of Gaussian mixture model (GMM) and Support Vector Machine (SVM) classifiers using the features derived from HOS and from the power spectrum. Results show that the selected HOS based features achieve 93.11% classification accuracy compared to 88.78% with features derived from the power spectrum for a GMM classifier. The SVM classifier achieves an improvement from 86.89% with features based on the power spectrum to 92.56% with features based on the bispectrum.

Sparse temporal representations for facial expression recognition

In automatic facial expression recognition, an increasing number of techniques had been proposed for in the literature that exploits the temporal nature of facial expressions. As all facial expressions are known to evolve over time, it is crucially important for a classifier to be capable of modelling their dynamics. We establish that the method of sparse representation (SR) classifiers proves to be a suitable candidate for this purpose, and subsequently propose a framework for expression dynamics to be efficiently incorporated into its current formulation. We additionally show that for the SR method to be applied effectively, then a certain threshold on image dimensionality must be enforced (unlike in facial recognition problems). Thirdly, we determined that recognition rates may be significantly influenced by the size of the projection matrix \Phi. To demonstrate these, a battery of experiments had been conducted on the CK+ dataset for the recognition of the seven prototypic expressions - anger, contempt, disgust, fear, happiness, sadness and surprise - and comparisons have been made between the proposed temporal-SR against the static-SR framework and state-of-the-art support vector machine.

Relevance feature discovery for text analysis

It is a big challenge to guarantee the quality of discovered relevance features in text documents for describing user preferences because of the large number of terms, patterns, and noise. Most existing popular text mining and classification methods have adopted term-based approaches. However, they have all suffered from the problems of polysemy and synonymy. Over the years, people have often held the hypothesis that pattern-based methods should perform better than term- based ones in describing user preferences, but many experiments do not support this hypothesis. This research presents a promising method, Relevance Feature Discovery (RFD), for solving this challenging issue. It discovers both positive and negative patterns in text documents as high-level features in order to accurately weight low-level features (terms) based on their specificity and their distributions in the high-level features. The thesis also introduces an adaptive model (called ARFD) to enhance the exibility of using RFD in adaptive environment. ARFD automatically updates the system's knowledge based on a sliding window over new incoming feedback documents. It can efficiently decide which incoming documents can bring in new knowledge into the system. Substantial experiments using the proposed models on Reuters Corpus Volume 1 and TREC topics show that the proposed models significantly outperform both the state-of-the-art term-based methods underpinned by Okapi BM25, Rocchio or Support Vector Machine and other pattern-based methods.

An integrated approach for precise road reconstruction from aerial imagery and LiDAR data

Accurate and detailed road models play an important role in a number of geospatial applications, such as infrastructure planning, traffic monitoring, and driver assistance systems. In this thesis, an integrated approach for the automatic extraction of precise road features from high resolution aerial images and LiDAR point clouds is presented. A framework of road information modeling has been proposed, for rural and urban scenarios respectively, and an integrated system has been developed to deal with road feature extraction using image and LiDAR analysis. For road extraction in rural regions, a hierarchical image analysis is first performed to maximize the exploitation of road characteristics in different resolutions. The rough locations and directions of roads are provided by the road centerlines detected in low resolution images, both of which can be further employed to facilitate the road information generation in high resolution images. The histogram thresholding method is then chosen to classify road details in high resolution images, where color space transformation is used for data preparation. After the road surface detection, anisotropic Gaussian and Gabor filters are employed to enhance road pavement markings while constraining other ground objects, such as vegetation and houses. Afterwards, pavement markings are obtained from the filtered image using the Otsu's clustering method. The final road model is generated by superimposing the lane markings on the road surfaces, where the digital terrain model (DTM) produced by LiDAR data can also be combined to obtain the 3D road model. As the extraction of roads in urban areas is greatly affected by buildings, shadows, vehicles, and parking lots, we combine high resolution aerial images and dense LiDAR data to fully exploit the precise spectral and horizontal spatial resolution of aerial images and the accurate vertical information provided by airborne LiDAR. Objectoriented image analysis methods are employed to process the feature classiffcation and road detection in aerial images. In this process, we first utilize an adaptive mean shift (MS) segmentation algorithm to segment the original images into meaningful object-oriented clusters. Then the support vector machine (SVM) algorithm is further applied on the MS segmented image to extract road objects. Road surface detected in LiDAR intensity images is taken as a mask to remove the effects of shadows and trees. In addition, normalized DSM (nDSM) obtained from LiDAR is employed to filter out other above-ground objects, such as buildings and vehicles. The proposed road extraction approaches are tested using rural and urban datasets respectively. The rural road extraction method is performed using pan-sharpened aerial images of the Bruce Highway, Gympie, Queensland. The road extraction algorithm for urban regions is tested using the datasets of Bundaberg, which combine aerial imagery and LiDAR data. Quantitative evaluation of the extracted road information for both datasets has been carried out. The experiments and the evaluation results using Gympie datasets show that more than 96% of the road surfaces and over 90% of the lane markings are accurately reconstructed, and the false alarm rates for road surfaces and lane markings are below 3% and 2% respectively. For the urban test sites of Bundaberg, more than 93% of the road surface is correctly reconstructed, and the mis-detection rate is below 10%.

Towards spontaneous facial expression recognition in real-world video

Facial expression is an important channel of human social communication. Facial expression recognition (FER) aims to perceive and understand emotional states of humans based on information in the face. Building robust and high performance FER systems that can work in real-world video is still a challenging task, due to the various unpredictable facial variations and complicated exterior environmental conditions, as well as the difficulty of choosing a suitable type of feature descriptor for extracting discriminative facial information. Facial variations caused by factors such as pose, age, gender, race and occlusion, can exert profound influence on the robustness, while a suitable feature descriptor largely determines the performance. Most present attention on FER has been paid to addressing variations in pose and illumination. No approach has been reported on handling face localization errors and relatively few on overcoming facial occlusions, although the significant impact of these two variations on the performance has been proved and highlighted in many previous studies. Many texture and geometric features have been previously proposed for FER. However, few comparison studies have been conducted to explore the performance differences between different features and examine the performance improvement arisen from fusion of texture and geometry, especially on data with spontaneous emotions. The majority of existing approaches are evaluated on databases with posed or induced facial expressions collected in laboratory environments, whereas little attention has been paid on recognizing naturalistic facial expressions on real-world data. This thesis investigates techniques for building robust and high performance FER systems based on a number of established feature sets. It comprises of contributions towards three main objectives: (1) Robustness to face localization errors and facial occlusions. An approach is proposed to handle face localization errors and facial occlusions using Gabor based templates. Template extraction algorithms are designed to collect a pool of local template features and template matching is then performed to covert these templates into distances, which are robust to localization errors and occlusions. (2) Improvement of performance through feature comparison, selection and fusion. A comparative framework is presented to compare the performance between different features and different feature selection algorithms, and examine the performance improvement arising from fusion of texture and geometry. The framework is evaluated for both discrete and dimensional expression recognition on spontaneous data. (3) Evaluation of performance in the context of real-world applications. A system is selected and applied into discriminating posed versus spontaneous expressions and recognizing naturalistic facial expressions. A database is collected from real-world recordings and is used to explore feature differences between standard database images and real-world images, as well as between real-world images and real-world video frames. The performance evaluations are based on the JAFFE, CK, Feedtum, NVIE, Semaine and self-collected QUT databases. The results demonstrate high robustness of the proposed approach to the simulated localization errors and occlusions. Texture and geometry have different contributions to the performance of discrete and dimensional expression recognition, as well as posed versus spontaneous emotion discrimination. These investigations provide useful insights into enhancing robustness and achieving high performance of FER systems, and putting them into real-world applications.

V1-inspired features induce a weighted margin in SVMs

Image representations derived from simplified models of the primary visual cortex (V1), such as HOG and SIFT, elicit good performance in a myriad of visual classification tasks including object recognition/detection, pedestrian detection and facial expression classification. A central question in the vision, learning and neuroscience communities regards why these architectures perform so well. In this paper, we offer a unique perspective to this question by subsuming the role of V1-inspired features directly within a linear support vector machine (SVM). We demonstrate that a specific class of such features in conjunction with a linear SVM can be reinterpreted as inducing a weighted margin on the Kronecker basis expansion of an image. This new viewpoint on the role of V1-inspired features allows us to answer fundamental questions on the uniqueness and redundancies of these features, and offer substantial improvements in terms of computational and storage efficiency.

An ensemble method for predicting subnuclear localizations from primary protein structures

Background Predicting protein subnuclear localization is a challenging problem. Some previous works based on non-sequence information including Gene Ontology annotations and kernel fusion have respective limitations. The aim of this work is twofold: one is to propose a novel individual feature extraction method; another is to develop an ensemble method to improve prediction performance using comprehensive information represented in the form of high dimensional feature vector obtained by 11 feature extraction methods. Methodology/Principal Findings A novel two-stage multiclass support vector machine is proposed to predict protein subnuclear localizations. It only considers those feature extraction methods based on amino acid classifications and physicochemical properties. In order to speed up our system, an automatic search method for the kernel parameter is used. The prediction performance of our method is evaluated on four datasets: Lei dataset, multi-localization dataset, SNL9 dataset and a new independent dataset. The overall accuracy of prediction for 6 localizations on Lei dataset is 75.2% and that for 9 localizations on SNL9 dataset is 72.1% in the leave-one-out cross validation, 71.7% for the multi-localization dataset and 69.8% for the new independent dataset, respectively. Comparisons with those existing methods show that our method performs better for both single-localization and multi-localization proteins and achieves more balanced sensitivities and specificities on large-size and small-size subcellular localizations. The overall accuracy improvements are 4.0% and 4.7% for single-localization proteins and 6.5% for multi-localization proteins. The reliability and stability of our classification model are further confirmed by permutation analysis. Conclusions It can be concluded that our method is effective and valuable for predicting protein subnuclear localizations. A web server has been designed to implement the proposed method. It is freely available at http://bioinformatics.awowshop.com/snlpr​ed_page.php.

Bearing fault prognosis based on health state probability estimation

In condition-based maintenance (CBM), effective diagnostic and prognostic tools are essential for maintenance engineers to identify imminent fault and predict the remaining useful life before the components finally fail. This enables remedial actions to be taken in advance and reschedule of production if necessary. All machine components are subjected to degradation processes in real environments and they have certain failure characteristics which can be related to the operating conditions. This paper describes a technique for accurate assessment of the remnant life of bearings based on health state probability estimation and historical knowledge embedded in the closed loop diagnostics and prognostics system. The technique uses the Support Vector Machine (SVM) classifier as a tool for estimating health state probability of machine degradation process to provide long term prediction. To validate the feasibility of the proposed model, real life fault historical data from bearings of High Pressure-Liquefied Natural Gas (HP-LNG) pumps were analysed and used to obtain the optimal prediction of remaining useful life (RUL). The results obtained were very encouraging and showed that the proposed prognosis system based on health state probability estimation has the potential to be used as an estimation tool for remnant life prediction in industrial machinery.

Predicting fault-prone software modules with rank sum classification

The detection and correction of defects remains among the most time consuming and expensive aspects of software development. Extensive automated testing and code inspections may mitigate their effect, but some code fragments are necessarily more likely to be faulty than others, and automated identification of fault prone modules helps to focus testing and inspections, thus limiting wasted effort and potentially improving detection rates. However, software metrics data is often extremely noisy, with enormous imbalances in the size of the positive and negative classes. In this work, we present a new approach to predictive modelling of fault proneness in software modules, introducing a new feature representation to overcome some of these issues. This rank sum representation offers improved or at worst comparable performance to earlier approaches for standard data sets, and readily allows the user to choose an appropriate trade-off between precision and recall to optimise inspection effort to suit different testing environments. The method is evaluated using the NASA Metrics Data Program (MDP) data sets, and performance is compared with existing studies based on the Support Vector Machine (SVM) and Naïve Bayes (NB) Classifiers, and with our own comprehensive evaluation of these methods.

Learning detectors quickly with stationary statistics

Computer vision is increasingly becoming interested in the rapid estimation of object detectors. The canonical strategy of using Hard Negative Mining to train a Support Vector Machine is slow, since the large negative set must be traversed at least once per detector. Recent work has demonstrated that, with an assumption of signal stationarity, Linear Discriminant Analysis is able to learn comparable detectors without ever revisiting the negative set. Even with this insight, the time to learn a detector can still be on the order of minutes. Correlation filters, on the other hand, can produce a detector in under a second. However, this involves the unnatural assumption that the statistics are periodic, and requires the negative set to be re-sampled per detector size. These two methods differ chie y in the structure which they impose on the co- variance matrix of all examples. This paper is a comparative study which develops techniques (i) to assume periodic statistics without needing to revisit the negative set and (ii) to accelerate the estimation of detectors with aperiodic statistics. It is experimentally verified that periodicity is detrimental.

Automatic classification of free-text radiology reports to identify limb fractures using machine learning and the SNOMED CT ontology

Objective To develop and evaluate machine learning techniques that identify limb fractures and other abnormalities (e.g. dislocations) from radiology reports. Materials and Methods 99 free-text reports of limb radiology examinations were acquired from an Australian public hospital. Two clinicians were employed to identify fractures and abnormalities from the reports; a third senior clinician resolved disagreements. These assessors found that, of the 99 reports, 48 referred to fractures or abnormalities of limb structures. Automated methods were then used to extract features from these reports that could be useful for their automatic classification. The Naive Bayes classification algorithm and two implementations of the support vector machine algorithm were formally evaluated using cross-fold validation over the 99 reports. Result Results show that the Naive Bayes classifier accurately identifies fractures and other abnormalities from the radiology reports. These results were achieved when extracting stemmed token bigram and negation features, as well as using these features in combination with SNOMED CT concepts related to abnormalities and disorders. The latter feature has not been used in previous works that attempted classifying free-text radiology reports. Discussion Automated classification methods have proven effective at identifying fractures and other abnormalities from radiology reports (F-Measure up to 92.31%). Key to the success of these techniques are features such as stemmed token bigrams, negations, and SNOMED CT concepts associated with morphologic abnormalities and disorders. Conclusion This investigation shows early promising results and future work will further validate and strengthen the proposed approaches.

Classification of cancer-related death certificates using machine learning

Background Cancer monitoring and prevention relies on the critical aspect of timely notification of cancer cases. However, the abstraction and classification of cancer from the free-text of pathology reports and other relevant documents, such as death certificates, exist as complex and time-consuming activities. Aims In this paper, approaches for the automatic detection of notifiable cancer cases as the cause of death from free-text death certificates supplied to Cancer Registries are investigated. Method A number of machine learning classifiers were studied. Features were extracted using natural language techniques and the Medtex toolkit. The numerous features encompassed stemmed words, bi-grams, and concepts from the SNOMED CT medical terminology. The baseline consisted of a keyword spotter using keywords extracted from the long description of ICD-10 cancer related codes. Results Death certificates with notifiable cancer listed as the cause of death can be effectively identified with the methods studied in this paper. A Support Vector Machine (SVM) classifier achieved best performance with an overall F-measure of 0.9866 when evaluated on a set of 5,000 free-text death certificates using the token stem feature set. The SNOMED CT concept plus token stem feature set reached the lowest variance (0.0032) and false negative rate (0.0297) while achieving an F-measure of 0.9864. The SVM classifier accounts for the first 18 of the top 40 evaluated runs, and entails the most robust classifier with a variance of 0.001141, half the variance of the other classifiers. Conclusion The selection of features significantly produced the most influences on the performance of the classifiers, although the type of classifier employed also affects performance. In contrast, the feature weighting schema created a negligible effect on performance. Specifically, it is found that stemmed tokens with or without SNOMED CT concepts create the most effective feature when combined with an SVM classifier.

Random Gabor based templates for facial expression recognition in images with facial occlusion

Robust facial expression recognition (FER) under occluded face conditions is challenging. It requires robust algorithms of feature extraction and investigations into the effects of different types of occlusion on the recognition performance to gain insight. Previous FER studies in this area have been limited. They have spanned recovery strategies for loss of local texture information and testing limited to only a few types of occlusion and predominantly a matched train-test strategy. This paper proposes a robust approach that employs a Monte Carlo algorithm to extract a set of Gabor based part-face templates from gallery images and converts these templates into template match distance features. The resulting feature vectors are robust to occlusion because occluded parts are covered by some but not all of the random templates. The method is evaluated using facial images with occluded regions around the eyes and the mouth, randomly placed occlusion patches of different sizes, and near-realistic occlusion of eyes with clear and solid glasses. Both matched and mis-matched train and test strategies are adopted to analyze the effects of such occlusion. Overall recognition performance and the performance for each facial expression are investigated. Experimental results on the Cohn-Kanade and JAFFE databases demonstrate the high robustness and fast processing speed of our approach, and provide useful insight into the effects of occlusion on FER. The results on the parameter sensitivity demonstrate a certain level of robustness of the approach to changes in the orientation and scale of Gabor filters, the size of templates, and occlusions ratios. Performance comparisons with previous approaches show that the proposed method is more robust to occlusion with lower reductions in accuracy from occlusion of eyes or mouth.