Feature Weighting

Case-based reasoning (CBR) is one of the most popular prediction techniques in medical domains because it is easy to apply, has no possibility of overfitting, and provides a good explanation for the output. However, it has a critical limitation -its prediction performance is generally lower than other AI techniques like artificial neural networks (ANN). In order to obtain accurate results from CBR, effective retrieval and matching of useful prior cases for the problem is essential, but it is still a controversial issue to design a good matching and retrieval mechanism for CBR systems. In this study, we propose a novel approach to enhance the prediction performance of CBR. Our suggestion is the simultaneous optimization of feature weights, instance selection, and the number of neighbors that combine using genetic algorithms (GA). Our model improves the prediction performance in three ways -(1) measuring similarity between cases more accurately by considering relative importance of each feature, (2) eliminating useless or erroneous reference cases, and (3) combining several similar cases represent significant patterns. To validate the usefulness of our model, this study applied it to a realworld case for evaluating cytological features derived directly from a digital scan of breast fine needle aspirate (FNA) slides. Experimental results showed that the prediction accuracy of conventional CBR may be improved significantly by using our model. We also found that our proposed model outperformed all the other optimized models for CBR using GA.

The aim of this paper is to study and compare several machine learning methods for implementing a Thai terrorism event extraction system. The main function of the system is to extract information related to terrorism events found in Thai news articles. The terrorism events can then be classified and presented to intelligence officers who can further analyze and predict terrorism events. This paper compares three named entity feature selection techniques provided by terrorism gazetteer, terrorism ontology and terrorism grammar rules, for entity recognition. The machine learning algorithms use for event extraction include Naïve Bayes (NB), K Nearest

viii ix x The IADIS European Conference on Data Mining 2009 received 63 submissions from more than 19 countries. Each submission has been anonymously reviewed by an average of five independent reviewers, to ensure that accepted submissions were of a high standard. Consequently only 14 full papers were published which means an acceptance rate of about 22 %.

Gaussian classifiers are strongly dependent on their underlying distance method, namely the Mahalanobis distance. Even though widely used, in the presence of noise this distance measure loses dramatically in performance, due to equal summation of the squared distances over all features. The features with large distance can mask all the other features so that the classification considers only these features, neglecting the information provided by the other features. To overcome this drawback we propose to weight the different features in the Mahalanobis distance according to their distances after the variance normalization. The idea behind this is to give less weight to noisy features and high weight to noise free features which are more reliable. Thereafter, we replace the traditional distance measure in a Gaussian classifier with the proposed. In a series of experiments we show the improved noise robustness of Gaussian classifiers by the proposed modifications in contrast to the traditional approach.

In this paper, we argue to learn dissimilarity for interactive search in content based image retrieval. In literature, dissimilarity is often learned via the feature space by feature selection, feature weighting or by adjusting the parameters of a function of the features. Other than existing techniques, we use feedback to adjust the dissimilarity space independent of feature space. This has the great advantage that it manipulates dissimilarity directly. To create a dissimilarity space, we use the method proposed by Pekalska and Duin, selecting a set of images called prototypes and computing distances to those prototypes for all images in the collection. After the user gives feedback, we apply active learning with a one-class support vector machine to decide the movement of images such that relevant images stay close together while irrelevant ones are pushed away (the work of Guo et al.). The dissimilarity space is then adjusted accordingly. Results on a Corel dataset of 10000 images and a TrecVid collection of 43907 keyframes show that our proposed approach is not only intuitive, it also significantly improves the retrieval performance.

Selection of feature subset is a preprocessing step in computational learning, and it serves several purposes like reducing the dimensionality of a dataset, decreasing the computational time required for classification and enhancing the classification accuracy of a classifier by removing redundant and misleading or erroneous features. This paper presents a new feature selection and weighting method aided with the decomposition based evolutionary multi-objective algorithm called MOEA/D. The feature vectors are selected and weighted
or scaled simultaneously to project the data points to such a hyper space, where the distance between data points of non-identical classes is increased, thus, making them easier to classify. The inter-class and intraclass distances are simultaneously optimized by using MOEA/D to obtain the optimal features and the scaling factor associated with them. Finally, k-NN (k-Nearest Neighbor) is used to classify the data points having the reduced and weighted feature set. The proposed algorithm is tested with several practical datasets from the well-known data repositories like UCI and LIBSVM. The results are compared with those obtained with the state-of-the-art algorithms to demonstrate the superiority of the proposed algorithm.

In this paper, we propose a generic text summarization method that generates summaries of Turkish texts by ranking sentences according to their scores calculated using their surface level features and extracting the highest ranked ones from the original documents. In order to extract sentences which form a summary with an extensive coverage of main content of the text and less redundancy, we use the features such as term frequency, key phrase, centrality, title similarity and position of the sentence in the original text. Sentence rank is computed using a score function that uses its feature values and the weights of the features. The best feature weights are learned using machine learning techniques with the help of human constructed summaries. Performance evaluation is conducted by comparing summarization outputs with manual summaries generated by 25 independent human evaluators. This paper presents one of the first Turkish summarization systems, and its results are promising.

Automatic document classification due to its various applications in data mining and information technology is one of the important topics in computer science. Classification plays a vital role in many information management and retrieval tasks. Document classification, also known as document categorization, is the process of assigning a document to one or more predefined category labels. Classification is often posed as a supervised learning problem in which a set of labeled data is used to train a classifier which can be applied to label future examples [1]. Document classification includes different parts such as text processing, feature extraction, feature vector construction and final classification. Thus improvement in each part should lead to better results in document classification. In this paper, we apply machine learning methods for automatic Persian news classification. In this regard, we first try to exert some language preprocess in Hamshahri dataset [2], and then we extract a feature vector for each news text by using feature weighting and feature selection algorithms. After that we train our classifier by support vector machine (SVM) and K-nearest neighbor (KNN) algorithms. In Experiments, although both algorithms show acceptable results for Persian text classification, the performance of KNN is better in comparison to SVM.

Interfering noise severely degrades the performance of a speech recognition system. The Parallel Model Compensation (PMC) technique is one of the most efficient techniques for dealing with such noise. Another approach is to use features local in the frequency domain, such as Mel-Frequency Discrete Wavelet Coefficients (MFDWCs). In this paper, we investigate the use of PMC and MFDWC features to take advantage of both noise compensation and local features (MFDWCs) to decrease the effect of noise on recognition performance. We also introduce a practical weighting technique based on the noise level of each coefficient. We evaluate the performance of several wavelet-schemes using the NOISEX-92 database for various noise types and noise levels. Finally, we compare the performance of these versus Mel-Frequency Cepstral Coefficients (MFCCs), both using PMC. Experimental results show significant performance improvements for MFDWCs versus MFCCs, particularly after compensating the HMMs using the PMC technique. The best feature vector among the six MFDWCs we tried gave 13.72 and 5.29 points performance improvement, on the average, over MFCCs for À6 and 0 dB SNR, respectively. This corresponds to 39.9% and 62.8% error reductions, respectively. Weighting the partial score of each coefficient based on the noise level further improves the performance. The average error rates for the best MFDWCs dropped from 19.57% to 16.71% and from 3.14% to 2.14% for À6 dB and 0 dB noise levels, respectively, using the weighting scheme. These improvements correspond to 14.6% and 31.8% error reductions for À6 dB and 0 dB noise levels, respectively.

Data collected from a paper mill using a WIC-100 process analyzer was divided into six classes, each representing a dif-Ž . ferent kind of paper grade or quality. Each of the six classes were modeled separately by principal component analysis PCA .

Heuristic search effectiveness depends directly upon the quality of heuristic evaluations of states in a search space. Given the large amount of research effort devoted to computer chess throughout the past half-century, insufficient attention has been paid to the issue of determining if a proposed change to an evaluation function is beneficial.

This paper describes ACE, a framework for automatically finding effective classification methodologies for arbitrary supervised classification problems. ACE performs experiments with both individual classifiers and classifier ensembles in order to find the approaches best suited to particular problems. A special emphasis is placed on classifier ensembles, as they can be powerful tools, yet are currently rarely used in MIR research. In addition to evaluating various classification methodologies in terms of success rates, ACE also allows users to specify constraints on training and classification times.

In content-based image retrieval, understanding the user's needs is a challenging task that requires integrating him in the process of retrieval. Relevance feedback (RF) has proven to be an effective tool for taking the user's judgement into account. In this paper, we present a new RF framework based on a feature selection algorithm that nicely combines the advantages of a probabilistic formulation with those of using both the positive example (PE) and the negative example (NE). Through interaction with the user, our algorithm learns the importance he assigns to image features, and then applies the results obtained to define similarity measures that correspond better to his judgement. The use of the NE allows images undesired by the user to be discarded, thereby improving retrieval accuracy. As for the probabilistic formulation of the problem, it presents a multitude of advantages and opens the door to more modeling possibilities that achieve a good feature selection. It mak...

The distance metric is the corner stone of nearest neighbor (NN)-based methods, and therefore, of nearest prototype (NP) algorithms. That is because they classify depending on the similarity of the data. When the data is characterized by a set of features which may contribute to the classification task in different levels, feature weighting or selection is required, sometimes in a local sense. However, local weighting is typically restricted to NN approaches. In this paper, we introduce local feature weighting (LFW) in NP classification. LFW provides each prototype its own weight vector, opposite to typical global weighting methods found in the NP literature, where all the prototypes share the same one. Providing each prototype its own weight vector has a novel effect in the borders of the Voronoi regions generated: They become nonlinear. We have integrated LFW with a previously developed evolutionary nearest prototype classifier (ENPC). The experiments performed both in artificial and real data sets demonstrate that the resulting algorithm that we call LFW in nearest prototype classification (LFW-NPC) avoids overfitting on training data in domains where the features may have different contribution to the classification task in different areas of the feature space. This generalization capability is also reflected in automatically obtaining an accurate and reduced set of prototypes.

In this paper, fusion of Principal Component Analysis (PCA) and generalization of Linear Discriminant Analysis (LDA) in the context of multiview face recognition is proposed. The generalization of LDA is extended to establish correlation between face classes in the transformed representation, which is called canonical covariate. The proposed work uses Gabor filter bank for extracting facial features characterized by spatial frequency, spatial locality and orientation to compensate the variations in face that occur due to change in illumination, pose and facial expression. Convolution of Gabor filter bank with face images produces Gabor face representations with high dimensional feature vectors. PCA and canonical covariate are then applied on the Gabor face representations to reduce the high dimensional feature spaces into low dimensional Gabor eigenfaces and Gabor canonical faces. Reduced eigenface vector and canonical face vector are fused together using weighted mean fusion rule. Finally, support vector machines have been trained with augmented fused set of features to perform recognition task. The proposed system has been evaluated with UMIST face database and performs with higher recognition accuracy for multi-view face images.

The Feature Weighting Classifier (FWC) is an efficient multi-class classification algorithm for text data that uses Information Gain to directly estimate per-class feature weights in the classifier. This classifier requires only a single pass over the dataset to compute the feature frequencies per class, is easy to implement, and has memory usage that is linear in the number of features. Results of experiments performed on 128 binary and multi-class text and web datasets show that FWC's performance is at least comparable to, and often better than that of Naive Bayes, TWCNB, Winnow, Balanced Winnow and linear SVM. On a large-scale web dataset with 12,294 classes and 135,973 training instances, FWC trained in 13 seconds and yielded comparable classification performance to a state of the art multi-class SVM implementation, which took over 15 minutes to train.

Vehicle detection in traffic scenes is an important issue in driver assistance systems and self-guided vehicles that includes two stages of Hypothesis Generation (HG) and Hypothesis Verification (HV). The both stages are important and challenging. In the first stage, potential vehicles are hypothesized and in the second stage, all hypotheses are verified and classified into vehicle and non-vehicle classes. In this paper, we present a method for detecting front and rear on-road vehicles without lane information and prior knowledge about the position of the road. In the HG stage, a three-step method including shadow, texture and symmetry clues is applied. In the HV stage, we extract Pyramid Histograms of Oriented Gradients (PHOG) features from a traffic image as basic features to detect vehicles. Principle Component Analysis (PCA) is applied to these PHOG feature vectors as a dimension reduction tool to obtain the PHOG-PCA vectors. Then, we use Genetic Algorithm (GA) and linear Support Vector Machine (SVM) to improve the performance and generalization of the PHOG-PCA features. Experimental results of the proposed HV stage showed good classification accuracy of more than 97% correct classification on realistic on-road vehicle dataset images and also it has better classification accuracy in comparison with other approaches.

In this work we propose a feature weighting method for classification tasks by extracting relevant information from a trained neural network. This method weights an attribute based on strengths (weights) of related links in the neural network, in which an important feature is typically connected to strong links and has more impact on the outputs. This method is applied to feature weighting for the nearest neighbor classifier and is tested on 15 real-world classification tasks. The results show that it can improve the nearest neighbor classifier on 14 of the 15 tested tasks, and also outperforms the neural network on 9 tasks.

Classification fusion combines multiple classifications of data into a single classification solution of greater accuracy. Feature extraction aims to reduce the computational cost of feature measurement, increase classifier efficiency, and allow greater classification accuracy based on the process of deriving new features from the original features. This paper represents an approach for classifying students in order to predict their final grades based on features extracted from logged data in an educational web-based system. A combination of multiple classifiers leads to a significant improvement in classification performance. By weighing feature vectors representing feature importance using a Genetic Algorithm (GA) we can optimize the prediction accuracy and obtain a marked improvement over raw classification. We further show that when the number of features is few, feature weighting and transformation into a new space works efficiently compared to the feature subset selection. This approach is easily adaptable to different types of courses, different population sizes, and allows for different features to be analyzed.

We tend to think of what we 'really' know as what we can talk about, and disparage knowledge that we can' t verbalize." ((Dowling 1989), 252) The exemplar-based learning model is proposed here as an alternative approach to modeling many aspects of music cognition. The implementation of this model is based on a k-NN (nearest neighbor) classifier, and on a genetic algorithm which is used for feature weighting. Although humans are capable of consciously abstracting concepts and deriving rules, there are other cognitive tasks such as music knowledge acquisition that are largely non-verbal and defy generalizations, consequently making the application of traditional rule-based AI models problematic. In exemplar-based learning, such as the k-NN rule, objects are categorized by their similarity to one or more sets of stored examples, which can be represented as weighted feature vectors. Similarity is often defined as the distance between the vectors. In the current implementation, the genetic algorithm is used to find a near-optimal set of weights. This paradigm, also known as the lazy learning model, is attractive because training is not necessary, learning is extremely fast, algorithms are simple and intuitive, rules are not sought, and learning is incremental. The major drawback has been the high memory requirement, since all examples must be stored, but the recent decrease in memory cost makes this model quite feasible. Exemplar-based recognition models have been successfully applied in many pattern recognition and classification tasks. Furthermore, cognitive psychologists have found this model evident in human and animal learning. In music, style recognition, harmonization, expressive performance, instrument recognition, and structural analysis are some of the obvious targets for the deployment of this model. Use of this model is illustrated with an optical music recognition system and a musical instrument identifier that uses only the steady-state (post-attack) portion of instrumental sounds.

This paper analyses the relation between the use of similarity in Memory-Based Learning and the notion of backed-off smoothing in statistical language model-ing. We show that the two approaches are closely related, and we argue that feature weighting methods in the Memory-Based ...

In this paper, we argue to learn dissimilarity for interactive search in content based image retrieval. In literature, dissimilarity is often learned via the feature space by feature selection, feature weighting or by adjusting the parameters of a function of the features. Other than existing techniques, we use feedback to adjust the dissimilarity space independent of feature space. This has the great advantage that it manipulates dissimilarity directly. To create a dissimilarity space, we use the method proposed by Pekalska and Duin, selecting a set of images called prototypes and computing distances to those prototypes for all images in the collection. After the user gives feedback, we apply active learning with a one-class support vector machine to decide the movement of images such that relevant images stay close together while irrelevant ones are pushed away (the work of Guo et al.). The dissimilarity space is then adjusted accordingly. Results on a Corel dataset of 10000 images and a TrecVid collection of 43907 keyframes show that our proposed approach is not only intuitive, it also significantly improves the retrieval performance.

except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights.

The Feature Weighting Classifier (FWC) is an efficient multi-class classification algorithm for text data that uses Information Gain to directly estimate per-class feature weights in the classifier. This classifier requires only a single pass over the dataset to compute the feature frequencies per class, is easy to implement, and has memory usage that is linear in the number of features.

Automatic document classification due to its various applications in data mining and information technology is one of the important topics in computer science. Classification plays a vital role in many information management and retrieval tasks. Document classification, also known as document categorization, is the process of assigning a document to one or more predefined category labels. Classification is often posed as a supervised learning problem in which a set of labeled data is used to train a classifier which can be applied to label future examples [1]. Document classification includes different parts such as text processing, feature extraction, feature vector construction and final classification. Thus improvement in each part should lead to better results in document classification. In this paper, we apply machine learning methods for automatic Persian news classification. In this regard, we first try to exert some language preprocess in Hamshahri dataset [2], and then we extract a feature vector for each news text by using feature weighting and feature selection algorithms. After that we train our classifier by support vector machine (SVM) and K-nearest neighbor (KNN) algorithms. In Experiments, although both algorithms show acceptable results for Persian text classification, the performance of KNN is better in comparison to SVM.

We present a gesture recognition approach for depth video data based on a novel Feature Weighting approach within the Dynamic Time Warping framework. Depth features from human joints are compared through video sequences using Dynamic Time Warping, and weights are assigned to features based on inter-intra class gesture variability. Feature Weighting in Dynamic Time Warping is then applied for recognizing begin-end of gestures in data sequences. The obtained results recognizing several gestures in depth data show high performance compared with classical Dynamic Time Warping approach.

Detection of targets concealed in foliage is a challenging problem and is critical for ground surveillance. To detect foliage-concealed targets, we need to address two major challenges, namely, 1) how to remotely acquire information that contains important features of foliage-concealed targets, and 2) how to distinguish targets from background and clutter. Synthetic aperture radar operated in low VHF-band has shown very good penetration capability in the forest environment, and hence the first problem can be satisfactorily addressed. The second problem is the focus of this paper. Existing detection schemes can achieve good detection performance but at the cost of high false alarm rate. To address the limitation of the existing schemes, in this paper, we develop a target detection algorithm based on a supervised learning technique that maximizes the margin between two classes, i.e., the target class and the non-target class. Specifically, our target detection algorithm consists of 1) image differencing, 2) maximum-margin classifier, and 3) diversity combining. The image differencing is to enhance and highlight the targets so that the targets are more distinguishable from the background. The maximum-margin classifier is based on a recently developed feature weighting technique called I-RELIEF; the objective of the maximum-margin classifier is to achieve robustness against uncertainties and clutter. The diversity combining utilizes multiple images to further improve the performance of detection, and hence it is a type of multi-pass change detection. We evaluate the performance of our proposed detection algorithm, using the SAR image data collected by Swedish CARABAS-II systems which operates at low VHF-band around 20-90 MHz. The experimental results demonstrate superior performance of our algorithm, compared to the benchmark algorithm associated with the CARABAS-II SAR image data. For example, for the same level of target detection probability, our algorithm only produces 11 false alarms while the benchmark algorithm produces 86 false alarms.

The exploration of three-dimensional (3D) anthropometry scanning data along with other existing subject medical profiles using data mining techniques becomes an important research issue for medical decision support. This research attempts to construct a classification approach based on the hybrid use of case-based reasoning (CBR) and genetic algorithms (GAs) for hypertension detection using anthropometric body surface scanning data. The obtained result reveals the relationship between a subject's 3D scanning data and hypertension disease. The GA is adopted to determine the appropriate feature weights for CBR. The proposed approaches were experimented and compared with a regular CBR and other widely used approaches including neural nets and decision trees. The experiment showed that applying GA to determine the suitable weights in CBR is a feasible approach to improving the effectiveness of case matching of hypertension disease. It also demonstrated that different weighted CBR approach presents better classification accuracy over the results obtained from other approaches.

The flexibility of activated factor X (fXa) binding site was assessed employing ligand-based pharmacophor modeling combined with genetic algorithm (GA)-based QSAR modeling. Four training subsets of wide structural diversity were selected from a total of 199 direct fXa inhibitors and were employed to generate different fXa pharmacophoric hypotheses using CATALYST ® software over two subsequent stages. In the first stage, high quality binding models (hypotheses) were identified. However, in the second stage, these models were refined by applying variable feature weight analysis to assess the relative significance of their features in the ligand-target affinity. The binding models were validated according to their coverage (capacity as a three-dimensional (3D) database search queries) and predictive potential as threedimensional quantitative structure-activity relationship (3D-QSAR) models. Subsequently, GA and multiple linear regression (MLR) analysis were employed to construct different QSAR models from high quality pharmacophores and explore the statistical significance of combination models in explaining bioactivity variations across 199 fXa inhibitors. Three orthogonal pharmacophoric models emerged in the optimal QSAR equation suggesting they represent three binding modes accessible to ligands in the binding pocket within fXa.

The Mel-Frequency Cepstral Coefficients (MFCC) and their derivatives are commonly used as acoustic features for speaker recognition. The issue arises of whether some of those features are redundant or dependent on other features. Probably, not all of them are equally relevant for speaker recognition. Reduced feature sets allow more robust estimates of the model parameters. Also, less computational resources are required, which is crucial for real-time speaker recognition applications using low-resource devices. In this paper, we use feature weighting as an intermediate step towards feature selection. Genetic algorithms are used to find the optimal set of weights for a 38-dimensional feature set, consisting of 12 MFCC, their first and second derivatives, energy and its first derivative. To evaluate each set of weights, speaker recognition errors are counted over a validation dataset. Speaker models are based on empirical distributions of acoustic labels, obtained through vector quantization. On average, weighting acoustic features yields between 15% and 25% error reduction in speaker recognition tests. Finally, features are sorted according to their weights, and the K features with greatest average ranks are retained and evaluated. We conclude that combining feature weighting and feature selection allows to reduce costs without degrading performance. 1

This is the first paper on textual case-based reasoning to employ collective classification, a methodology for simultaneously classifying related cases that has consistently attained higher accuracies than standard classification approaches when cases are related. Thus far, case-based classifiers have not been examined for their use in collective classification. We introduce Case-Based Collective Classification (CBCC) and report that it outperforms a traditional case-based classifier on three tasks. We also address issues of case representation and feature weight learning for CBCC. In particular, we describe a cross-validation approach for tuning feature weights and show that it increases CBCC accuracy on these tasks.

Many medical diagnosis applications are characterized by datasets that contain under-represented classes due to the fact that the disease is much rarer than the normal case. In such a situation classifiers such as decision trees and Naive Bayesian that generalize over the data are not the proper choice as classification methods. Case-based classifiers that can work on the samples seen so far are more appropriate for such a task. We propose to calculate the contingency table and class specific evaluation measures despite the overall accuracy for evaluation purposes of classifiers for these specific data characteristics. We evaluate the different options of our case-based classifier and compare the performance to decision trees and Naive Bayesian. Finally, we give an outlook for further work.

In this paper, we propose a feature-weighted mountain clustering method. The proposed method can work well when there are noisy feature variables and could be useful for obtaining initial estimat of cluster centers for other clustering algorithms. Results from color image segmentation illustrate the proposed method actually produces better segmentation than previous methods.

Fusionplex is a system for integrating multiple heterogeneous and autonomous information sources that uses data fusion to resolve factual inconsistencies among the individual sources. To accomplish this, the system relies on source features, which are meta-data on the merits of each information source; for example, the recentness of the data, its accuracy, its availability, or its cost. The fusion process is controlled with several parameters: (1) with a vector of feature weights, each user defines an individual notion of data utility; (2) with thresholds of acceptance, users ensure minimal performance of their data, excluding from the fusion process data that are too old, too costly, or lacking in authority, or numeric data that are too high, too low, or obvious outliers; and, ultimately, (3) in naming a particular fusion function to be used for each attribute (for example, average, maximum, or simply any) users implement their own interpretation of fusion. Several simple extensions to SQL are all that is needed to allow users to state these resolution parameters, thus ensuring that the system is easy to use. Altogether, Fusionplex provides its users with powerful and flexible, yet simple, control over the fusion process. In addition, Fusionplex supports other critical integration requirements, such as information source heterogeneity, dynamic evolution of the information environment, quick ad-hoc integration, and intermittent source availability. The methods described in this paper were implemented in a prototype system that provides complete Web-based integration services for remote clients. (A. Motro). www.elsevier.com/locate/inffus Information Fusion 7 (2006) 176-196 1 These aspects are explained in detail in Section 2.3. 2 A somewhat similar approach can be seen in the area of Internet search engines, where Google, one of the prominent search engines, weighs heavily the authority or importance of individual links in the ranking of its search results .

In this paper, we present a feature combination approach to object tracking based upon graph embedding techniques. The method presented here abstracts the low complexity features used for purposes of tracking to a relational structure and employs graph-spectral methods to combine them. This gives rise to a feature combination scheme which minimises the mutual cross-correlation between features and is devoid of free parameters. It also allows an analytical solution making use of matrix factorisation techniques. The new target location is recovered making use of a weighted combination of target-centre shifts corresponding to each of the features under study, where the feature weights arise from a cost function governed by the embedding process. This treatment permits the update of the feature weights in an on-line fashion in a straightforward manner. We illustrate the performance of our method in real-world image sequences and compare our results to a number of alternatives.