User-representative feature selection for keystroke dynamics

Continuous user authentication with keystroke dynamics uses characters sequences as features. Since users can type characters in any order, it is imperative to find character sequences (n-graphs) that are representative of user typing behavior. The contemporary feature selection approaches do not guarantee selecting frequently-typed features which may cause less accurate statistical user-representation. Furthermore, the selected features do not inherently reflect user typing behavior. We propose four statistical based feature selection techniques that mitigate limitations of existing approaches. The first technique selects the most frequently occurring features. The other three consider different user typing behaviors by selecting: n-graphs that are typed quickly; n-graphs that are typed with consistent time; and n-graphs that have large time variance among users. We use Gunetti’s keystroke dataset and k-means clustering algorithm for our experiments. The results show that among the proposed techniques, the most-frequent feature selection technique can effectively find user representative features. We further substantiate our results by comparing the most-frequent feature selection technique with three existing approaches (popular Italian words, common n-graphs, and least frequent ngraphs). We find that it performs better than the existing approaches after selecting a certain number of most-frequent n-graphs.

Feature selection using expected attainable discrimination

We propose expected attainable discrimination (EAD) as a measure to select discrete valued features for reliable discrimination between two classes of data. EAD is an average of the area under the ROC curves obtained when a simple histogram probability density model is trained and tested on many random partitions of a data set. EAD can be incorporated into various stepwise search methods to determine promising subsets of features, particularly when misclassification costs are difficult or impossible to specify. Experimental application to the problem of risk prediction in pregnancy is described.

Discovering the best feature extraction and selection algorithms for spontaneous facial expression recognition

Feature extraction and selection are critical processes in developing facial expression recognition (FER) systems. While many algorithms have been proposed for these processes, direct comparison between texture, geometry and their fusion, as well as between multiple selection algorithms has not been found for spontaneous FER. This paper addresses this issue by proposing a unified framework for a comparative study on the widely used texture (LBP, Gabor and SIFT) and geometric (FAP) features, using Adaboost, mRMR and SVM feature selection algorithms. Our experiments on the Feedtum and NVIE databases demonstrate the benefits of fusing geometric and texture features, where SIFT+FAP shows the best performance, while mRMR outperforms Adaboost and SVM. In terms of computational time, LBP and Gabor perform better than SIFT. The optimal combination of SIFT+FAP+mRMR also exhibits a state-of-the-art performance.

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.

Semantic labelling for document feature patterns using ontological subjects

Finding and labelling semantic features patterns of documents in a large, spatial corpus is a challenging problem. Text documents have characteristics that make semantic labelling difficult; the rapidly increasing volume of online documents makes a bottleneck in finding meaningful textual patterns. Aiming to deal with these issues, we propose an unsupervised documnent labelling approach based on semantic content and feature patterns. A world ontology with extensive topic coverage is exploited to supply controlled, structured subjects for labelling. An algorithm is also introduced to reduce dimensionality based on the study of ontological structure. The proposed approach was promisingly evaluated by compared with typical machine learning methods including SVMs, Rocchio, and kNN.

Enhanced n-gram extraction using relevance feature discovery

Guaranteeing the quality of extracted features that describe relevant knowledge to users or topics is a challenge because of the large number of extracted features. Most popular existing term-based feature selection methods suffer from noisy feature extraction, which is irrelevant to the user needs (noisy). One popular method is to extract phrases or n-grams to describe the relevant knowledge. However, extracted n-grams and phrases usually contain a lot of noise. This paper proposes a method for reducing the noise in n-grams. The method first extracts more specific features (terms) to remove noisy features. The method then uses an extended random set to accurately weight n-grams based on their distribution in the documents and their terms distribution in n-grams. The proposed approach not only reduces the number of extracted n-grams but also improves the performance. The experimental results on Reuters Corpus Volume 1 (RCV1) data collection and TREC topics show that the proposed method significantly outperforms the state-of-art methods underpinned by Okapi BM25, tf*idf and Rocchio.

Automated feature engineering for HTTP tunnel detection

Generating discriminative input features is a key requirement for achieving highly accurate classifiers. The process of generating features from raw data is known as feature engineering and it can take significant manual effort. In this paper we propose automated feature engineering to derive a suite of additional features from a given set of basic features with the aim of both improving classifier accuracy through discriminative features, and to assist data scientists through automation. Our implementation is specific to HTTP computer network traffic. To measure the effectiveness of our proposal, we compare the performance of a supervised machine learning classifier built with automated feature engineering versus one using human-guided features. The classifier addresses a problem in computer network security, namely the detection of HTTP tunnels. We use Bro to process network traffic into base features and then apply automated feature engineering to calculate a larger set of derived features. The derived features are calculated without favour to any base feature and include entropy, length and N-grams for all string features, and counts and averages over time for all numeric features. Feature selection is then used to find the most relevant subset of these features. Testing showed that both classifiers achieved a detection rate above 99.93% at a false positive rate below 0.01%. For our datasets, we conclude that automated feature engineering can provide the advantages of increasing classifier development speed and reducing development technical difficulties through the removal of manual feature engineering. These are achieved while also maintaining classification accuracy.

A user driven data mining process model and learning system

This paper deals with the problem of using the data mining models in a real-world situation where the user can not provide all the inputs with which the predictive model is built. A learning system framework, Query Based Learning System (QBLS), is developed for improving the performance of the predictive models in practice where not all inputs are available for querying to the system. The automatic feature selection algorithm called Query Based Feature Selection (QBFS) is developed for selecting features to obtain a balance between the relative minimum subset of features and the relative maximum classification accuracy. Performance of the QBLS system and the QBFS algorithm is successfully demonstrated with a real-world application

Exploring visual features through Gabor representations for facial expression detection

Gabor representations have been widely used in facial analysis (face recognition, face detection and facial expression detection) due to their biological relevance and computational properties. Two popular Gabor representations used in literature are: 1) Log-Gabor and 2) Gabor energy filters. Even though these representations are somewhat similar, they also have distinct differences as the Log-Gabor filters mimic the simple cells in the visual cortex while the Gabor energy filters emulate the complex cells, which causes subtle differences in the responses. In this paper, we analyze the difference between these two Gabor representations and quantify these differences on the task of facial action unit (AU) detection. In our experiments conducted on the Cohn-Kanade dataset, we report an average area underneath the ROC curve (A`) of 92.60% across 17 AUs for the Gabor energy filters, while the Log-Gabor representation achieved an average A` of 96.11%. This result suggests that small spatial differences that the Log-Gabor filters pick up on are more useful for AU detection than the differences in contours and edges that the Gabor energy filters extract.

Application of data mining techniques in the prediction of coronary artery disease : use of anaesthesia time-series and patient risk factor data

The high morbidity and mortality associated with atherosclerotic coronary vascular disease (CVD) and its complications are being lessened by the increased knowledge of risk factors, effective preventative measures and proven therapeutic interventions. However, significant CVD morbidity remains and sudden cardiac death continues to be a presenting feature for some subsequently diagnosed with CVD. Coronary vascular disease is also the leading cause of anaesthesia related complications. Stress electrocardiography/exercise testing is predictive of 10 year risk of CVD events and the cardiovascular variables used to score this test are monitored peri-operatively. Similar physiological time-series datasets are being subjected to data mining methods for the prediction of medical diagnoses and outcomes. This study aims to find predictors of CVD using anaesthesia time-series data and patient risk factor data. Several pre-processing and predictive data mining methods are applied to this data. Physiological time-series data related to anaesthetic procedures are subjected to pre-processing methods for removal of outliers, calculation of moving averages as well as data summarisation and data abstraction methods. Feature selection methods of both wrapper and filter types are applied to derived physiological time-series variable sets alone and to the same variables combined with risk factor variables. The ability of these methods to identify subsets of highly correlated but non-redundant variables is assessed. The major dataset is derived from the entire anaesthesia population and subsets of this population are considered to be at increased anaesthesia risk based on their need for more intensive monitoring (invasive haemodynamic monitoring and additional ECG leads). Because of the unbalanced class distribution in the data, majority class under-sampling and Kappa statistic together with misclassification rate and area under the ROC curve (AUC) are used for evaluation of models generated using different prediction algorithms. The performance based on models derived from feature reduced datasets reveal the filter method, Cfs subset evaluation, to be most consistently effective although Consistency derived subsets tended to slightly increased accuracy but markedly increased complexity. The use of misclassification rate (MR) for model performance evaluation is influenced by class distribution. This could be eliminated by consideration of the AUC or Kappa statistic as well by evaluation of subsets with under-sampled majority class. The noise and outlier removal pre-processing methods produced models with MR ranging from 10.69 to 12.62 with the lowest value being for data from which both outliers and noise were removed (MR 10.69). For the raw time-series dataset, MR is 12.34. Feature selection results in reduction in MR to 9.8 to 10.16 with time segmented summary data (dataset F) MR being 9.8 and raw time-series summary data (dataset A) being 9.92. However, for all time-series only based datasets, the complexity is high. For most pre-processing methods, Cfs could identify a subset of correlated and non-redundant variables from the time-series alone datasets but models derived from these subsets are of one leaf only. MR values are consistent with class distribution in the subset folds evaluated in the n-cross validation method. For models based on Cfs selected time-series derived and risk factor (RF) variables, the MR ranges from 8.83 to 10.36 with dataset RF_A (raw time-series data and RF) being 8.85 and dataset RF_F (time segmented time-series variables and RF) being 9.09. The models based on counts of outliers and counts of data points outside normal range (Dataset RF_E) and derived variables based on time series transformed using Symbolic Aggregate Approximation (SAX) with associated time-series pattern cluster membership (Dataset RF_ G) perform the least well with MR of 10.25 and 10.36 respectively. For coronary vascular disease prediction, nearest neighbour (NNge) and the support vector machine based method, SMO, have the highest MR of 10.1 and 10.28 while logistic regression (LR) and the decision tree (DT) method, J48, have MR of 8.85 and 9.0 respectively. DT rules are most comprehensible and clinically relevant. The predictive accuracy increase achieved by addition of risk factor variables to time-series variable based models is significant. The addition of time-series derived variables to models based on risk factor variables alone is associated with a trend to improved performance. Data mining of feature reduced, anaesthesia time-series variables together with risk factor variables can produce compact and moderately accurate models able to predict coronary vascular disease. Decision tree analysis of time-series data combined with risk factor variables yields rules which are more accurate than models based on time-series data alone. The limited additional value provided by electrocardiographic variables when compared to use of risk factors alone is similar to recent suggestions that exercise electrocardiography (exECG) under standardised conditions has limited additional diagnostic value over risk factor analysis and symptom pattern. The effect of the pre-processing used in this study had limited effect when time-series variables and risk factor variables are used as model input. In the absence of risk factor input, the use of time-series variables after outlier removal and time series variables based on physiological variable values’ being outside the accepted normal range is associated with some improvement in model performance.

Machine learning techniques to improve software quality

A significant proportion of the cost of software development is due to software testing and maintenance. This is in part the result of the inevitable imperfections due to human error, lack of quality during the design and coding of software, and the increasing need to reduce faults to improve customer satisfaction in a competitive marketplace. Given the cost and importance of removing errors improvements in fault detection and removal can be of significant benefit. The earlier in the development process faults can be found, the less it costs to correct them and the less likely other faults are to develop. This research aims to make the testing process more efficient and effective by identifying those software modules most likely to contain faults, allowing testing efforts to be carefully targeted. This is done with the use of machine learning algorithms which use examples of fault prone and not fault prone modules to develop predictive models of quality. In order to learn the numerical mapping between module and classification, a module is represented in terms of software metrics. A difficulty in this sort of problem is sourcing software engineering data of adequate quality. In this work, data is obtained from two sources, the NASA Metrics Data Program, and the open source Eclipse project. Feature selection before learning is applied, and in this area a number of different feature selection methods are applied to find which work best. Two machine learning algorithms are applied to the data - Naive Bayes and the Support Vector Machine - and predictive results are compared to those of previous efforts and found to be superior on selected data sets and comparable on others. In addition, a new classification method is proposed, Rank Sum, in which a ranking abstraction is laid over bin densities for each class, and a classification is determined based on the sum of ranks over features. A novel extension of this method is also described based on an observed polarising of points by class when rank sum is applied to training data to convert it into 2D rank sum space. SVM is applied to this transformed data to produce models the parameters of which can be set according to trade-off curves to obtain a particular performance trade-off.