Sherine Rady

Information-theoretic environment modeling for efficient topological localization

Place recognition is a vital methodology for modeling environments and localizing autonomous mobile robots topologically. It can also be integrated in a hierarchical framework where it guides a fast and more precise metric position estimation. Especially for those hierarchical frameworks, it is crucial that the place recognition modules be highly accurate. In this paper, an information-theoretic approach that focuses on the efficiency of place recognition for topological environment modeling and localization is presented. The approach relies on a minimal discriminative feature set obtained from an entropy-based qualitative evaluation and a codebook compression. The generated environment feature map achieves a significant combination of high localization accuracy, speed and less memory storage.

An E-mail Filtering Approach Using Classification Techniques

E-mail is one of the most popular ways of communication due to its accessibility, low sending cost and fast message transfer. However, Spam emails appear as a severe problem affecting this application of today’s Internet. Filtering is an important approach to isolate those spam emails. In this paper, an approach for filtering spam email is proposed, which is based on classification techniques. The approach analyses the body of Email messages and assigns weights to terms (features) that can help identifying spam and clean (ham) emails. An adaptation is proposed that tries to reduce the dimensionality of the extracted features, in which only determined (meaningful) terms are regarded by consulting a dictionary. A thorough comparative study has been studied among different classification algorithms that prove the efficiency of the filtering approach proposed. The approach has been evaluated using Enron dataset.

Email filtering based on supervised learning and mutual information feature selection

Electronic mail is one of todays most important ways to communicate and transfer information. Because of fast delivery and easy to access, it is used almost in every aspect of communication in work and life. However, the increase in email users has resulted in a dramatic increase in spam emails during the past few years. In this paper, we propose an email-filtering approach that is based on supervised classifier and mutual information. The proposed model has the advantage of combining machine supervised learning with feature selection. Term frequency (TF) is presented to assign relevance weights to words of each email class. We conduct experiments to compare between six different classifiers. Results show that the proposed approach has high performance in terms of precision, recall and accuracy performance measures.

A study of spatial machine learning for business behavior prediction in location based social networks

Understanding business behaviors requires acquiring huge amounts of data from diverse field studies. Location Based Social Networks can provide such large amounts of data that can be used in urban analysis to understand business behaviors. Towards more insight for business behavior, a novel analytical prespective that exploits data collected from Location Based Social Networks is introduced to predict business turnouts. Prediction is implemented using machine learning techniques. Spatial regression models are investigated through a comparative study to model the dataset features relationships for business behavior prediction. Geographically Weighted Regression model is found to be the most appropriate in predicting business turnouts of objects provided by Location Based Social Networks. Moreover, a Partitioned Geographically Weighted Regression model is proposed to deal with the data heterogeneity nature pursuing more accurate predictions for the business turnouts. An experimental case study, using data about venues registered in Foursquare is conducted to assess the performance of the proposed methods. The experimental results confirm the best performance by the Geographically Weighted Regression compared to Durbin, Durbin Error, Spatial Lag, Spatial Error, and Spatial Lag X regression models presented in this study. Moreover, the proposed Partitioned Geographically Weighted Regression model experimental results showed better prediction accuracy compared to the classical Geographically Weighted Regression model.

Exploiting location based social networks in business predictions

The growing use of Location Based Social Networks especially in recent years provides large amount of data transactions. These data transactions attract many data mining researchers to infer various information from them. In this paper, a geographic business prediction technique is proposed, which infers business usage by exploiting data published about venues in Location Based Social Networks. The proposed technique is beneficial for investors and business decision makers. The proposed geo-business prediction technique considers spatial and categorical factors in the prediction process. Both factors affect the prediction accuracy rather than using traditional spatial prediction techniques, which are usually used where only the location feature is involved in the prediction process. Additionally, an outlier filter is proposed and applied to the data to avoid extreme values involvement in the prediction process in order to achieve better prediction accuracy. To test the proposed technique, an experimental case study is implemented. It uses data extracted from Foursquare about business venues in Texas State in the United States of America. The proposed geo business prediction technique has shown to provide better prediction accuracy than k nearest neighbor spatial prediction. The Application of the outlier filter, results in even higher prediction accuracy for the proposed technique.

Business Behavior Predictions Using Location Based Social Networks in Smart Cities

Understanding business behaviors require acquiring lots of data through numerous amounts of field research. The growing use of mobile devices, especially in recent years, provides large amount of data transactions that can replace the data acquired by field researchers. Data acquired from location based social networks can be exploited in urban analysis for economic reasons. Such research studies the spatial correlation of business turnouts for venues in location based social networks. Towards more insight for business behavior in smart cities, data acquired from location based social networks is used for predicting the business turnouts based on their spatial locations. The presented work moves along this direction by proposing a machine learning approach using spatial interpolation applied to predict business turnouts. In this approach, a similarity embedded spatial interpolation technique is additionally proposed. The proposed technique, with the exploitation of the multiple features provided by location based social networks, can issue better prediction performance. To test the proposed technique, an experimental case study is implemented. The case study uses training data extracted from Foursquare about venues in Texas in the United States of America. The proposed similarity embedded spatial interpolation technique has shown better prediction accuracy for business turnouts than classical spatial interpolation predictions.

Exploring feature dimensionality reduction methods for enhancing automatic sport image annotation

Nowadays, multimedia information requires the demand to investigate and apply efficient techniques for better annotation and retrieval purposes. In the content-based indexing, low-level features are generally extracted from images to serve as image descriptors. Other than the descriptor poses a computational overhead, the learning model may also tend to overfit, resulting in performance degeneration. This work solves such problems in the sport image domain by proposing feature dimensionality reduction techniques for the retrieval and annotation of image datasets. Different techniques are investigated, such as Information Gain, Gain Ratio, Chi-Square, and Latent Semantic Analysis (LSA), and applied for sport images classification using Support Vector Machine (SVM) classifier. A comparison between the performances of applying SVM alone and when incorporating the different reduction methods is presented. Experimental results show that the SVM classification accuracy is 76.4%; while integrating LSA technique manages to raise the accuracy to 96%, with the other techniques recording 74% accuracy at 50% feature space reduction.

A Convolutional Neural Network Model for Emotion Detection from Tweets.

Sentiment analysis and emotion recognition are major indicators of society trends toward certain topics. Analyzing opinions and feelings helps improving the human-computer interaction in several fields ranging from opinion mining to psychological concerns. This paper proposes a deep learning model for emotion detection from short informal sentences. The model consists of three Convolutional Neural Networks (CNNs). Each CNN contains a convolutional layer and a max-pooling layer, followed by a fully-connected layer for classifying the sentences into positive or negative. The model employs the word vector representation as textual features, which works on random initialization for the word vectors, and are set to be trainable and updated through the model training phase. Eventually, task-specific vectors are generated as the model learns to distinguish the meaning of words in the dataset. The model has been tested on the Stanford Twitter Sentiment dataset for classifying sentiment into two classes positive and negative. The presented model achieved to record 80.6% accuracy as a prove that even with randomly initialized word vectors, it can work very well in text classification tasks when trained with CNNs.

Supervised Classification Techniques for Identifying Alzheimer’s Disease

Alzheimer’s Disease is a serious form of dementia. With no current cure, treatments focus on slowing the progression of the disease and controlling its symptoms. Early diagnosis by studying the biomarkers found in structural MRI is the key. This paper proposes a method which combines texture features extracted from gray level co-occurrence matrix and voxel-based morphometry neuroimaging analysis to classify Alzheimer’s disease patients. Different supervised classification techniques are studied, support vector machine, k-nearest neighbor, and decision tree, to obtain best identification accuracy. The paper explores as well the discriminative power for Alzheimer’s disease of certain anatomical regions of interest. The proposed technique is applied on gray matter tissues, and managed successfully to differentiate between Alzheimer’s disease patients and normal controls with accuracy 92%.

Developing an Efficient Clique-Based Algorithm for Community Detection in Large Graphs

Many computer science problems are structured as a network. Mobile, e-mail, social networks (MySpace, Friendster, Facebook, etc.), collaboration networks, and Protein-Protein Interaction (PPI), Gene Regulatory Networks (GRN) and Metabolic Networks (MN) in bioinformatics, are among several applications. Discovering communities in Networks is a recent and critical task in order to understand and model network structures. Several methods exist for community detection, such as modularity, clique, and random walk methods. These methods are somewhat limited because of the time needed to detect communities and their modularity. In this work, a Clique-based Community Detection Algorithm (CCDA) is proposed to overcome time and modularity limitations. The clique method is suitable since it arises in many real-world problems, as in bioinformatics, computational chemistry, and social networks. In definition, clique is a group of individuals who interact with one another and share similar interests. Based on this definition, if one vertex of a clique is assigned to a specific community, all other vertices in this clique often belong to the same community. CCDA develops a simple and fast method to detect maximum clique for specific vertex. In addition, testing is done for the closest neighbor node instead of testing all nodes in the graph. Since neighbor nodes are also sorted in descending order, it contributes to save more execution time. Furthermore, each node will be visited exactly once. To test the performance of CCDA, it is compared with previously proposed community detection algorithms (Louvain, and MACH with DDA-M2), using various datasets: Amazon (262111 nodes/1234877 vertices), DBLP (317080 nodes/1049866 vertices), and LiveJournal (4847571 nodes, 68993773 vertices). Results have proven the efficiency of the proposed method in terms of time performance and modularity.