Bayesian Learning with Dependency Structures via Latent Factors, Mixtures, and Copulas

Bayesian methods offer a flexible and convenient probabilistic learning framework to extract interpretable knowledge from complex and structured data. Such methods can characterize dependencies among multiple levels of hidden variables and share statistical strength across heterogeneous sources. In the first part of this dissertation, we develop two dependent variational inference methods for full posterior approximation in non-conjugate Bayesian models through hierarchical mixture- and copula-based variational proposals, respectively. The proposed methods move beyond the widely used factorized approximation to the posterior and provide generic applicability to a broad class of probabilistic models with minimal model-specific derivations. In the second part of this dissertation, we design probabilistic graphical models to accommodate multimodal data, describe dynamical behaviors and account for task heterogeneity. In particular, the sparse latent factor model is able to reveal common low-dimensional structures from high-dimensional data. We demonstrate the effectiveness of the proposed statistical learning methods on both synthetic and real-world data.

N-opcode analysis for android malware classification and categorization

Malware detection is a growing problem particularly on the Android mobile platform due to its increasing popularity and accessibility to numerous third party app markets. This has also been made worse by the increasingly sophisticated detection avoidance techniques employed by emerging malware families. This calls for more effective techniques for detection and classification of Android malware. Hence, in this paper we present an n-opcode analysis based approach that utilizes machine learning to classify and categorize Android malware. This approach enables automated feature discovery that eliminates the need for applying expert or domain knowledge to define the needed features. Our experiments on 2520 samples that were performed using up to 10-gram opcode features showed that an f-measure of 98% is achievable using this approach.

Sentiment classification with case-base approach

L’augmentation de la croissance des réseaux, des blogs et des utilisateurs des sites d’examen sociaux font d’Internet une énorme source de données, en particulier sur la façon dont les gens pensent, sentent et agissent envers différentes questions. Ces jours-ci, les opinions des gens jouent un rôle important dans la politique, l’industrie, l’éducation, etc. Alors, les gouvernements, les grandes et petites industries, les instituts universitaires, les entreprises et les individus cherchent à étudier des techniques automatiques fin d’extraire les informations dont ils ont besoin dans les larges volumes de données. L’analyse des sentiments est une véritable réponse à ce besoin. Elle est une application de traitement du langage naturel et linguistique informatique qui se compose de techniques de pointe telles que l’apprentissage machine et les modèles de langue pour capturer les évaluations positives, négatives ou neutre, avec ou sans leur force, dans des texte brut. Dans ce mémoire, nous étudions une approche basée sur les cas pour l’analyse des sentiments au niveau des documents. Notre approche basée sur les cas génère un classificateur binaire qui utilise un ensemble de documents classifies, et cinq lexiques de sentiments différents pour extraire la polarité sur les scores correspondants aux commentaires. Puisque l’analyse des sentiments est en soi une tâche dépendante du domaine qui rend le travail difficile et coûteux, nous appliquons une approche «cross domain» en basant notre classificateur sur les six différents domaines au lieu de le limiter à un seul domaine. Pour améliorer la précision de la classification, nous ajoutons la détection de la négation comme une partie de notre algorithme. En outre, pour améliorer la performance de notre approche, quelques modifications innovantes sont appliquées. Il est intéressant de mentionner que notre approche ouvre la voie à nouveaux développements en ajoutant plus de lexiques de sentiment et ensembles de données à l’avenir.

Learning to Interpret and Generate Instructional Recipes

Thesis (Ph.D.)--University of Washington, 2016-08

Techniques for batch reinforcement learning in robotics

This thesis addresses the Batch Reinforcement Learning methods in Robotics. This sub-class of Reinforcement Learning has shown promising results and has been the focus of recent research. Three contributions are proposed that aim to extend the state-of-art methods allowing for a faster and more stable learning process, such as required for learning in Robotics. The Q-learning update-rule is widely applied, since it allows to learn without the presence of a model of the environment. However, this update-rule is transition-based and does not take advantage of the underlying episodic structure of collected batch of interactions. The Q-Batch update-rule is proposed in this thesis, to process experiencies along the trajectories collected in the interaction phase. This allows a faster propagation of obtained rewards and penalties, resulting in faster and more robust learning. Non-parametric function approximations are explored, such as Gaussian Processes. This type of approximators allows to encode prior knowledge about the latent function, in the form of kernels, providing a higher level of exibility and accuracy. The application of Gaussian Processes in Batch Reinforcement Learning presented a higher performance in learning tasks than other function approximations used in the literature. Lastly, in order to extract more information from the experiences collected by the agent, model-learning techniques are incorporated to learn the system dynamics. In this way, it is possible to augment the set of collected experiences with experiences generated through planning using the learned models. Experiments were carried out mainly in simulation, with some tests carried out in a physical robotic platform. The obtained results show that the proposed approaches are able to outperform the classical Fitted Q Iteration.

Spectral Frame Analysis and Learning through Graph Structure

This dissertation investigates the connection between spectral analysis and frame theory. When considering the spectral properties of a frame, we present a few novel results relating to the spectral decomposition. We first show that scalable frames have the property that the inner product of the scaling coefficients and the eigenvectors must equal the inverse eigenvalues. From this, we prove a similar result when an approximate scaling is obtained. We then focus on the optimization problems inherent to the scalable frames by first showing that there is an equivalence between scaling a frame and optimization problems with a non-restrictive objective function. Various objective functions are considered, and an analysis of the solution type is presented. For linear objectives, we can encourage sparse scalings, and with barrier objective functions, we force dense solutions. We further consider frames in high dimensions, and derive various solution techniques. From here, we restrict ourselves to various frame classes, to add more specificity to the results. Using frames generated from distributions allows for the placement of probabilistic bounds on scalability. For discrete distributions (Bernoulli and Rademacher), we bound the probability of encountering an ONB, and for continuous symmetric distributions (Uniform and Gaussian), we show that symmetry is retained in the transformed domain. We also prove several hyperplane-separation results. With the theory developed, we discuss graph applications of the scalability framework. We make a connection with graph conditioning, and show the in-feasibility of the problem in the general case. After a modification, we show that any complete graph can be conditioned. We then present a modification of standard PCA (robust PCA) developed by Cand\`es, and give some background into Electron Energy-Loss Spectroscopy (EELS). We design a novel scheme for the processing of EELS through robust PCA and least-squares regression, and test this scheme on biological samples. Finally, we take the idea of robust PCA and apply the technique of kernel PCA to perform robust manifold learning. We derive the problem and present an algorithm for its solution. There is also discussion of the differences with RPCA that make theoretical guarantees difficult.

Dynamic Multi-Objective Optimization With jMetal and Spark: a Case Study

Technologies for Big Data and Data Science are receiving increasing research interest nowadays. This paper introduces the prototyping architecture of a tool aimed to solve Big Data Optimization problems. Our tool combines the jMetal framework for multi-objective optimization with Apache Spark, a technology that is gaining momentum. In particular, we make use of the streaming facilities of Spark to feed an optimization problem with data from different sources. We demonstrate the use of our tool by solving a dynamic bi-objective instance of the Traveling Salesman Problem (TSP) based on near real-time traffic data from New York City, which is updated several times per minute. Our experiment shows that both jMetal and Spark can be integrated providing a software platform to deal with dynamic multi-optimization problems.

Ensemble Stream Model for Data-Cleaning in Sensor Networks

Ensemble Stream Modeling and Data-cleaning are sensor information processing systems have different training and testing methods by which their goals are cross-validated. This research examines a mechanism, which seeks to extract novel patterns by generating ensembles from data. The main goal of label-less stream processing is to process the sensed events to eliminate the noises that are uncorrelated, and choose the most likely model without over fitting thus obtaining higher model confidence. Higher quality streams can be realized by combining many short streams into an ensemble which has the desired quality. The framework for the investigation is an existing data mining tool. First, to accommodate feature extraction such as a bush or natural forest-fire event we make an assumption of the burnt area (BA*), sensed ground truth as our target variable obtained from logs. Even though this is an obvious model choice the results are disappointing. The reasons for this are two: One, the histogram of fire activity is highly skewed. Two, the measured sensor parameters are highly correlated. Since using non descriptive features does not yield good results, we resort to temporal features. By doing so we carefully eliminate the averaging effects; the resulting histogram is more satisfactory and conceptual knowledge is learned from sensor streams. Second is the process of feature induction by cross-validating attributes with single or multi-target variables to minimize training error. We use F-measure score, which combines precision and accuracy to determine the false alarm rate of fire events. The multi-target data-cleaning trees use information purity of the target leaf-nodes to learn higher order features. A sensitive variance measure such as f-test is performed during each node’s split to select the best attribute. Ensemble stream model approach proved to improve when using complicated features with a simpler tree classifier. The ensemble framework for data-cleaning and the enhancements to quantify quality of fitness (30% spatial, 10% temporal, and 90% mobility reduction) of sensor led to the formation of streams for sensor-enabled applications. Which further motivates the novelty of stream quality labeling and its importance in solving vast amounts of real-time mobile streams generated today.

Life cycle modelling and design knowledge development in 3D virtual environments : final report

Experience plays an important role in building management. “How often will this asset need repair?” or “How much time is this repair going to take?” are types of questions that project and facility managers face daily in planning activities. Failure or success in developing good schedules, budgets and other project management tasks depend on the project manager's ability to obtain reliable information to be able to answer these types of questions. Young practitioners tend to rely on information that is based on regional averages and provided by publishing companies. This is in contrast to experienced project managers who tend to rely heavily on personal experience. Another aspect of building management is that many practitioners are seeking to improve available scheduling algorithms, estimating spreadsheets and other project management tools. Such “micro-scale” levels of research are important in providing the required tools for the project manager's tasks. However, even with such tools, low quality input information will produce inaccurate schedules and budgets as output. Thus, it is also important to have a broad approach to research at a more “macro-scale.” Recent trends show that the Architectural, Engineering, Construction (AEC) industry is experiencing explosive growth in its capabilities to generate and collect data. There is a great deal of valuable knowledge that can be obtained from the appropriate use of this data and therefore the need has arisen to analyse this increasing amount of available data. Data Mining can be applied as a powerful tool to extract relevant and useful information from this sea of data. Knowledge Discovery in Databases (KDD) and Data Mining (DM) are tools that allow identification of valid, useful, and previously unknown patterns so large amounts of project data may be analysed. These technologies combine techniques from machine learning, artificial intelligence, pattern recognition, statistics, databases, and visualization to automatically extract concepts, interrelationships, and patterns of interest from large databases. The project involves the development of a prototype tool to support facility managers, building owners and designers. This final report presents the AIMMTM prototype system and documents how and what data mining techniques can be applied, the results of their application and the benefits gained from the system. The AIMMTM system is capable of searching for useful patterns of knowledge and correlations within the existing building maintenance data to support decision making about future maintenance operations. The application of the AIMMTM prototype system on building models and their maintenance data (supplied by industry partners) utilises various data mining algorithms and the maintenance data is analysed using interactive visual tools. The application of the AIMMTM prototype system to help in improving maintenance management and building life cycle includes: (i) data preparation and cleaning, (ii) integrating meaningful domain attributes, (iii) performing extensive data mining experiments in which visual analysis (using stacked histograms), classification and clustering techniques, associative rule mining algorithm such as “Apriori” and (iv) filtering and refining data mining results, including the potential implications of these results for improving maintenance management. Maintenance data of a variety of asset types were selected for demonstration with the aim of discovering meaningful patterns to assist facility managers in strategic planning and provide a knowledge base to help shape future requirements and design briefing. Utilising the prototype system developed here, positive and interesting results regarding patterns and structures of data have been obtained.

Life Cycle Modelling and Design Knowledge Development in 3D Virtual Environments

Experience plays an important role in building management. “How often will this asset need repair?” or “How much time is this repair going to take?” are types of questions that project and facility managers face daily in planning activities. Failure or success in developing good schedules, budgets and other project management tasks depend on the project manager's ability to obtain reliable information to be able to answer these types of questions. Young practitioners tend to rely on information that is based on regional averages and provided by publishing companies. This is in contrast to experienced project managers who tend to rely heavily on personal experience. Another aspect of building management is that many practitioners are seeking to improve available scheduling algorithms, estimating spreadsheets and other project management tools. Such “micro-scale” levels of research are important in providing the required tools for the project manager's tasks. However, even with such tools, low quality input information will produce inaccurate schedules and budgets as output. Thus, it is also important to have a broad approach to research at a more “macro-scale.” Recent trends show that the Architectural, Engineering, Construction (AEC) industry is experiencing explosive growth in its capabilities to generate and collect data. There is a great deal of valuable knowledge that can be obtained from the appropriate use of this data and therefore the need has arisen to analyse this increasing amount of available data. Data Mining can be applied as a powerful tool to extract relevant and useful information from this sea of data. Knowledge Discovery in Databases (KDD) and Data Mining (DM) are tools that allow identification of valid, useful, and previously unknown patterns so large amounts of project data may be analysed. These technologies combine techniques from machine learning, artificial intelligence, pattern recognition, statistics, databases, and visualization to automatically extract concepts, interrelationships, and patterns of interest from large databases. The project involves the development of a prototype tool to support facility managers, building owners and designers. This Industry focused report presents the AIMMTM prototype system and documents how and what data mining techniques can be applied, the results of their application and the benefits gained from the system. The AIMMTM system is capable of searching for useful patterns of knowledge and correlations within the existing building maintenance data to support decision making about future maintenance operations. The application of the AIMMTM prototype system on building models and their maintenance data (supplied by industry partners) utilises various data mining algorithms and the maintenance data is analysed using interactive visual tools. The application of the AIMMTM prototype system to help in improving maintenance management and building life cycle includes: (i) data preparation and cleaning, (ii) integrating meaningful domain attributes, (iii) performing extensive data mining experiments in which visual analysis (using stacked histograms), classification and clustering techniques, associative rule mining algorithm such as “Apriori” and (iv) filtering and refining data mining results, including the potential implications of these results for improving maintenance management. Maintenance data of a variety of asset types were selected for demonstration with the aim of discovering meaningful patterns to assist facility managers in strategic planning and provide a knowledge base to help shape future requirements and design briefing. Utilising the prototype system developed here, positive and interesting results regarding patterns and structures of data have been obtained.

Rough set-based reasoning and pattern mining for information filtering

An information filtering (IF) system monitors an incoming document stream to find the documents that match the information needs specified by the user profiles. To learn to use the user profiles effectively is one of the most challenging tasks when developing an IF system. With the document selection criteria better defined based on the users’ needs, filtering large streams of information can be more efficient and effective. To learn the user profiles, term-based approaches have been widely used in the IF community because of their simplicity and directness. Term-based approaches are relatively well established. However, these approaches have problems when dealing with polysemy and synonymy, which often lead to an information overload problem. Recently, pattern-based approaches (or Pattern Taxonomy Models (PTM) [160]) have been proposed for IF by the data mining community. These approaches are better at capturing sematic information and have shown encouraging results for improving the effectiveness of the IF system. On the other hand, pattern discovery from large data streams is not computationally efficient. Also, these approaches had to deal with low frequency pattern issues. The measures used by the data mining technique (for example, “support” and “confidences”) to learn the profile have turned out to be not suitable for filtering. They can lead to a mismatch problem. This thesis uses the rough set-based reasoning (term-based) and pattern mining approach as a unified framework for information filtering to overcome the aforementioned problems. This system consists of two stages - topic filtering and pattern mining stages. The topic filtering stage is intended to minimize information overloading by filtering out the most likely irrelevant information based on the user profiles. A novel user-profiles learning method and a theoretical model of the threshold setting have been developed by using rough set decision theory. The second stage (pattern mining) aims at solving the problem of the information mismatch. This stage is precision-oriented. A new document-ranking function has been derived by exploiting the patterns in the pattern taxonomy. The most likely relevant documents were assigned higher scores by the ranking function. Because there is a relatively small amount of documents left after the first stage, the computational cost is markedly reduced; at the same time, pattern discoveries yield more accurate results. The overall performance of the system was improved significantly. The new two-stage information filtering model has been evaluated by extensive experiments. Tests were based on the well-known IR bench-marking processes, using the latest version of the Reuters dataset, namely, the Reuters Corpus Volume 1 (RCV1). The performance of the new two-stage model was compared with both the term-based and data mining-based IF models. The results demonstrate that the proposed information filtering system outperforms significantly the other IF systems, such as the traditional Rocchio IF model, the state-of-the-art term-based models, including the BM25, Support Vector Machines (SVM), and Pattern Taxonomy Model (PTM).

Fast content-based file type identification

Digital forensic examiners often need to identify the type of a file or file fragment based only on the content of the file. Content-based file type identification schemes typically use a byte frequency distribution with statistical machine learning to classify file types. Most algorithms analyze the entire file content to obtain the byte frequency distribution, a technique that is inefficient and time consuming. This paper proposes two techniques for reducing the classification time. The first technique selects a subset of features based on the frequency of occurrence. The second speeds classification by sampling several blocks from the file. Experimental results demonstrate that up to a fifteen-fold reduction in file size analysis time can be achieved with limited impact on accuracy.

The sample complexity of pattern classification with neural networks: the size of the weights is more important than the size of the network

Sample complexity results from computational learning theory, when applied to neural network learning for pattern classification problems, suggest that for good generalization performance the number of training examples should grow at least linearly with the number of adjustable parameters in the network. Results in this paper show that if a large neural network is used for a pattern classification problem and the learning algorithm finds a network with small weights that has small squared error on the training patterns, then the generalization performance depends on the size of the weights rather than the number of weights. For example, consider a two-layer feedforward network of sigmoid units, in which the sum of the magnitudes of the weights associated with each unit is bounded by A and the input dimension is n. We show that the misclassification probability is no more than a certain error estimate (that is related to squared error on the training set) plus A3 √((log n)/m) (ignoring log A and log m factors), where m is the number of training patterns. This may explain the generalization performance of neural networks, particularly when the number of training examples is considerably smaller than the number of weights. It also supports heuristics (such as weight decay and early stopping) that attempt to keep the weights small during training. The proof techniques appear to be useful for the analysis of other pattern classifiers: when the input domain is a totally bounded metric space, we use the same approach to give upper bounds on misclassification probability for classifiers with decision boundaries that are far from the training examples.

Model selection and error estimation

We study model selection strategies based on penalized empirical loss minimization. We point out a tight relationship between error estimation and data-based complexity penalization: any good error estimate may be converted into a data-based penalty function and the performance of the estimate is governed by the quality of the error estimate. We consider several penalty functions, involving error estimates on independent test data, empirical VC dimension, empirical VC entropy, and margin-based quantities. We also consider the maximal difference between the error on the first half of the training data and the second half, and the expected maximal discrepancy, a closely related capacity estimate that can be calculated by Monte Carlo integration. Maximal discrepancy penalty functions are appealing for pattern classification problems, since their computation is equivalent to empirical risk minimization over the training data with some labels flipped.