Data mining, also known as big data analysis, has been identified as a critical and challenging process for a variety of applications in real-world problems. Numerous datasets are collected and generated every day to store the information. The rise in the number of data volumes and data modality has resulted in the increased demand for data mining methods and strategies of finding anomalies, patterns, and correlations within large data sets to predict outcomes. Effective machine learning methods are widely adapted to build the data mining pipeline for various purposes like business understanding, data understanding, data preparation, modeling, evaluation, and deployment. The major challenges for effectively and efficiently mining big data include (1) data heterogeneity and (2) missing data. Heterogeneity is the natural characteristic of big data, as the data is typically collected from different sources with diverse formats. The missing value is the most common issue faced by the heterogeneous data analysis, which resulted from variety of factors including the data collecting processing, user initiatives, erroneous data entries, and so on. In response to these challenges, in this thesis, three main research directions with application scenarios have been investigated: (1) Mining and Formulating Heterogeneous Data, (2) missing value imputation strategy in various application scenarios in both offline and online manner, and (3) missing value imputation for multi-modality data. Multiple strategies with theoretical analysis are presented, and the evaluation of the effectiveness of the proposed algorithms compared with state-of-the-art methods is discussed.

Causality analysis is the process of identifying cause-effect relationships among variables. This process is challenging because causal relationships cannot be tested solely based on statistical indicators as additional information is always needed to reduce the ambiguity caused by factors beyond those covered by the statistical test. Traditionally, controlled experiments are carried out to identify causal relationships, but recently there is a growing interest in causality analysis with observational data due to the increasing availability of data and tools. This type of analysis will often involve automatic algorithms that extract causal relations from large amounts of data and rely on expert judgment to scrutinize and verify the relations. Over-reliance on these automatic algorithms is dangerous because models trained on observational data are susceptible to bias that can be difficult to spot even with expert oversight. Visualization has proven to be effective at bridging the gap between human experts and statistical models by enabling an interactive exploration and manipulation of the data and models. This thesis develops a visual analytics framework to support the interaction between human experts and automatic models in causality analysis. Three case studies were conducted to demonstrate the application of the visual analytics framework in which feature engineering, insight generation, correlation analysis, and causality inspections were showcased.

Attributes - that delineating the properties of data, and connections - that describing the dependencies of data, are two essential components to characterize most real-world phenomena. The synergy between these two principal elements renders a unique data representation - the attributed networks. In many cases, people are inundated with vast amounts of data that can be structured into attributed networks, and their use has been attractive to researchers and practitioners in different disciplines. For example, in social media, users interact with each other and also post personalized content; in scientific collaboration, researchers cooperate and are distinct from peers by their unique research interests; in complex diseases studies, rich gene expression complements to the gene-regulatory networks. Clearly, attributed networks are ubiquitous and form a critical component of modern information infrastructure. To gain deep insights from such networks, it requires a fundamental understanding of their unique characteristics and be aware of the related computational challenges.

My dissertation research aims to develop a suite of novel learning algorithms to understand, characterize, and gain actionable insights from attributed networks, to benefit high-impact real-world applications. In the first part of this dissertation, I mainly focus on developing learning algorithms for attributed networks in a static environment at two different levels: (i) attribute level - by designing feature selection algorithms to find high-quality features that are tightly correlated with the network topology; and (ii) node level - by presenting network embedding algorithms to learn discriminative node embeddings by preserving node proximity w.r.t. network topology structure and node attribute similarity. As changes are essential components of attributed networks and the results of learning algorithms will become stale over time, in the second part of this dissertation, I propose a family of online algorithms for attributed networks in a dynamic environment to continuously update the learning results on the fly. In fact, developing application-aware learning algorithms is more desired with a clear understanding of the application domains and their unique intents. As such, in the third part of this dissertation, I am also committed to advancing real-world applications on attributed networks by incorporating the objectives of external tasks into the learning process.

In recent years, the rise in social media usage both vertically in terms of the number of users by platform and horizontally in terms of the number of platforms per user has led to data explosion.

User-generated social media content provides an excellent opportunity to mine data of interest and to build resourceful applications. The rise in the number of healthcare-related social media platforms and the volume of healthcare knowledge available online in the last decade has resulted in increased social media usage for personal healthcare. In the United States, nearly ninety percent of adults, in the age group 50-75, have used social media to seek and share health information. Motivated by the growth of social media usage, this thesis focuses on healthcare-related applications, study various challenges posed by social media data, and address them through novel and effective machine learning algorithms.



The major challenges for effectively and efficiently mining social media data to build functional applications include: (1) Data reliability and acceptance: most social media data (especially in the context of healthcare-related social media) is not regulated and little has been studied on the benefits of healthcare-specific social media; (2) Data heterogeneity: social media data is generated by users with both demographic and geographic diversity; (3) Model transparency and trustworthiness: most existing machine learning models for addressing heterogeneity are considered as black box models, not many providing explanations for why they do what they do to trust them.

In response to these challenges, three main research directions have been investigated in this thesis: (1) Analyzing social media influence on healthcare: to study the real world impact of social media as a source to offer or seek support for patients with chronic health conditions; (2) Learning from task heterogeneity: to propose various models and algorithms that are adaptable to new social media platforms and robust to dynamic social media data, specifically on modeling user behaviors, identifying similar actors across platforms, and adapting black box models to a specific learning scenario; (3) Explaining heterogeneous models: to interpret predictive models in the presence of task heterogeneity. In this thesis, novel algorithms with theoretical analysis from various aspects (e.g., time complexity, convergence properties) have been proposed. The effectiveness and efficiency of the proposed algorithms is demonstrated by comparison with state-of-the-art methods and relevant case studies.

Multi-view learning, a subfield of machine learning that aims to improve model performance by training on multiple views of the data, has been studied extensively in the past decades. It is typically applied in contexts where the input features naturally form multiple groups or views. An example of a naturally multi-view context is a data set of websites, where each website is described not only by the text on the page, but also by the text of hyperlinks pointing to the page. More recently, various studies have demonstrated the initial success of applying multi-view learning on single-view data with multiple artificially constructed views. However, there lacks a systematic study regarding the effectiveness of such artificially constructed views. To bridge this gap, this thesis begins by providing a high-level overview of multi-view learning with the co-training algorithm. Co-training is a classic semi-supervised learning algorithm that takes advantage of both labelled and unlabelled examples in the data set for training. Then, the thesis presents a web-based tool developed in Python allowing users to experiment with and compare the performance of multiple view construction approaches on various data sets. The supported view construction approaches in the web-based tool include subsampling, Optimal Feature Set Partitioning, and the genetic algorithm. Finally, the thesis presents an empirical comparison of the performance of these approaches, not only against one another, but also against traditional single-view models. The findings show that a simple subsampling approach combined with co-training often outperforms both the other view construction approaches, as well as traditional single-view methods.

Despite the fact that machine learning supports the development of computer vision applications by shortening the development cycle, finding a general learning algorithm that solves a wide range of applications is still bounded by the ”no free lunch theorem”. The search for the right algorithm to solve a specific problem is driven by the problem itself, the data availability and many other requirements.

Automated visual inspection (AVI) systems represent a major part of these challenging computer vision applications. They are gaining growing interest in the manufacturing industry to detect defective products and keep these from reaching customers. The process of defect detection and classification in semiconductor units is challenging due to different acceptable variations that the manufacturing process introduces. Other variations are also typically introduced when using optical inspection systems due to changes in lighting conditions and misalignment of the imaged units, which makes the defect detection process more challenging.

In this thesis, a BagStack classification framework is proposed, which makes use of stacking and bagging concepts to handle both variance and bias errors. The classifier is designed to handle the data imbalance and overfitting problems by adaptively transforming the

multi-class classification problem into multiple binary classification problems, applying a bagging approach to train a set of base learners for each specific problem, adaptively specifying the number of base learners assigned to each problem, adaptively specifying the number of samples to use from each class, applying a novel data-imbalance aware cross-validation technique to generate the meta-data while taking into account the data imbalance problem at the meta-data level and, finally, using a multi-response random forest regression classifier as a meta-classifier. The BagStack classifier makes use of multiple features to solve the defect classification problem. In order to detect defects, a locally adaptive statistical background modeling is proposed. The proposed BagStack classifier outperforms state-of-the-art image classification techniques on our dataset in terms of overall classification accuracy and average per-class classification accuracy. The proposed detection method achieves high performance on the considered dataset in terms of recall and precision.

This thesis presents a family of adaptive curvature methods for gradient-based stochastic optimization. In particular, a general algorithmic framework is introduced along with a practical implementation that yields an efficient, adaptive curvature gradient descent algorithm. To this end, a theoretical and practical link between curvature matrix estimation and shrinkage methods for covariance matrices is established. The use of shrinkage improves estimation accuracy of the curvature matrix when data samples are scarce. This thesis also introduce several insights that result in data- and computation-efficient update equations. Empirical results suggest that the proposed method compares favorably with existing second-order techniques based on the Fisher or Gauss-Newton and with adaptive stochastic gradient descent methods on both supervised and reinforcement learning tasks.

Spike sorting is a critical step for single-unit-based analysis of neural activities extracellularly and simultaneously recorded using multi-channel electrodes. When dealing with recordings from very large numbers of neurons, existing methods, which are mostly semiautomatic in nature, become inadequate.

This dissertation aims at automating the spike sorting process. A high performance, automatic and computationally efficient spike detection and clustering system, namely, the M-Sorter2 is presented. The M-Sorter2 employs the modified multiscale correlation of wavelet coefficients (MCWC) for neural spike detection. At the center of the proposed M-Sorter2 are two automatic spike clustering methods. They share a common hierarchical agglomerative modeling (HAM) model search procedure to strategically form a sequence of mixture models, and a new model selection criterion called difference of model evidence (DoME) to automatically determine the number of clusters. The M-Sorter2 employs two methods differing by how they perform clustering to infer model parameters: one uses robust variational Bayes (RVB) and the other uses robust Expectation-Maximization (REM) for Student’s 𝑡-mixture modeling. The M-Sorter2 is thus a significantly improved approach to sorting as an automatic procedure.

M-Sorter2 was evaluated and benchmarked with popular algorithms using simulated, artificial and real data with truth that are openly available to researchers. Simulated datasets with known statistical distributions were first used to illustrate how the clustering algorithms, namely REMHAM and RVBHAM, provide robust clustering results under commonly experienced performance degrading conditions, such as random initialization of parameters, high dimensionality of data, low signal-to-noise ratio (SNR), ambiguous clusters, and asymmetry in cluster sizes. For the artificial dataset from single-channel recordings, the proposed sorter outperformed Wave_Clus, Plexon’s Offline Sorter and Klusta in most of the comparison cases. For the real dataset from multi-channel electrodes, tetrodes and polytrodes, the proposed sorter outperformed all comparison algorithms in terms of false positive and false negative rates. The software package presented in this dissertation is available for open access.

Unsupervised learning of time series data, also known as temporal clustering, is a challenging problem in machine learning. This thesis presents a novel algorithm, Deep Temporal Clustering (DTC), to naturally integrate dimensionality reduction and temporal clustering into a single end-to-end learning framework, fully unsupervised. The algorithm utilizes an autoencoder for temporal dimensionality reduction and a novel temporal clustering layer for cluster assignment. Then it jointly optimizes the clustering objective and the dimensionality reduction objective. Based on requirement and application, the temporal clustering layer can be customized with any temporal similarity metric. Several similarity metrics and state-of-the-art algorithms are considered and compared. To gain insight into temporal features that the network has learned for its clustering, a visualization method is applied that generates a region of interest heatmap for the time series. The viability of the algorithm is demonstrated using time series data from diverse domains, ranging from earthquakes to spacecraft sensor data. In each case, the proposed algorithm outperforms traditional methods. The superior performance is attributed to the fully integrated temporal dimensionality reduction and clustering criterion.

Network mining has been attracting a lot of research attention because of the prevalence of networks. As the world is becoming increasingly connected and correlated, networks arising from inter-dependent application domains are often collected from different sources, forming the so-called multi-sourced networks. Examples of such multi-sourced networks include critical infrastructure networks, multi-platform social networks, cross-domain collaboration networks, and many more. Compared with single-sourced network, multi-sourced networks bear more complex structures and therefore could potentially contain more valuable information.

This thesis proposes a multi-layered HITS (Hyperlink-Induced Topic Search) algorithm to perform the ranking task on multi-sourced networks. Specifically, each node in the network receives an authority score and a hub score for evaluating the value of the node itself and the value of its outgoing links respectively. Based on a recent multi-layered network model, which allows more flexible dependency structure across different sources (i.e., layers), the proposed algorithm leverages both within-layer smoothness and cross-layer consistency. This essentially allows nodes from different layers to be ranked accordingly. The multi-layered HITS is formulated as a regularized optimization problem with non-negative constraint and solved by an iterative update process. Extensive experimental evaluations demonstrate the effectiveness and explainability of the proposed algorithm.