Machine learning models are increasingly being deployed in real-world applications where their predictions are used to make critical decisions in a variety of domains. The proliferation of such models has led to a burgeoning need to ensure the reliability and safety of these models, given the potential negative consequences of model vulnerabilities. The complexity of machine learning models, along with the extensive data sets they analyze, can result in unpredictable and unintended outcomes. Model vulnerabilities may manifest due to errors in data input, algorithm design, or model deployment, which can have significant implications for both individuals and society. To prevent such negative outcomes, it is imperative to identify model vulnerabilities at an early stage in the development process. This will aid in guaranteeing the integrity, dependability, and safety of the models, thus mitigating potential risks and enabling the full potential of these technologies to be realized. However, enumerating vulnerabilities can be challenging due to the complexity of the real-world environment. Visual analytics, situated at the intersection of human-computer interaction, computer graphics, and artificial intelligence, offers a promising approach for achieving high interpretability of complex black-box models, thus reducing the cost of obtaining insights into potential vulnerabilities of models. This research is devoted to designing novel visual analytics methods to support the identification and analysis of model vulnerabilities. Specifically, generalizable visual analytics frameworks are instantiated to explore vulnerabilities in machine learning models concerning security (adversarial attacks and data perturbation) and fairness (algorithmic bias). In the end, a visual analytics approach is proposed to enable domain experts to explain and diagnose the model improvement of addressing identified vulnerabilities of machine learning models in a human-in-the-loop fashion. The proposed methods hold the potential to enhance the security and fairness of machine learning models deployed in critical real-world applications.

Modern physical systems are experiencing tremendous evolutions with growing size, more and more complex structures, and the incorporation of new devices. This calls for better planning, monitoring, and control. However, achieving these goals is challenging since the system knowledge (e.g., system structures and edge parameters) may be unavailable for a normal system, let alone some dynamic changes like maintenance, reconfigurations, and events, etc. Therefore, extracting system knowledge becomes a central topic. Luckily, advanced metering techniques bring numerous data, leading to the emergence of Machine Learning (ML) methods with efficient learning and fast inference. This work tries to propose a systematic framework of ML-based methods to learn system knowledge under three what-if scenarios: (i) What if the system is normally operated? (ii) What if the system suffers dynamic interventions? (iii) What if the system is new with limited data? For each case, this thesis proposes principled solutions with extensive experiments. Chapter 2 tackles scenario (i) and the golden rule is to learn an ML model that maintains physical consistency, bringing high extrapolation capacity for changing operational conditions. The key finding is that physical consistency can be linked to convexity, a central concept in optimization. Therefore, convexified ML designs are proposed and the global optimality implies faithfulness to the underlying physics. Chapter 3 handles scenario (ii) and the goal is to identify the event time, type, and locations. The problem is formalized as multi-class classification with special attention to accuracy and speed. Subsequently, Chapter 3 builds an ensemble learning framework to aggregate different ML models for better prediction. Next, to tackle high-volume data quickly, a tensor as the multi-dimensional array is used to store and process data, yielding compact and informative vectors for fast inference. Finally, if no labels exist, Chapter 3 uses physical properties to generate labels for learning. Chapter 4 deals with scenario (iii) and a doable process is to transfer knowledge from similar systems, under the framework of Transfer Learning (TL). Chapter 4 proposes cutting-edge system-level TL by considering the network structure, complex spatial-temporal correlations, and different physical information.

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.

Tensors are commonly used for representing multi-dimensional data, such as Web graphs, sensor streams, and social networks. As a consequence of the increase in the use of tensors, tensor decomposition operations began to form the basis for many data analysis and knowledge discovery tasks, from clustering, trend detection, anomaly detection to correlationanalysis [31, 38]. It is well known that Singular Value matrix Decomposition (SVD) [9] is used to extract latent semantics for matrix data. When apply SVD to tensors, which have more than two modes, it is tensor decomposition. The two most popular tensor decomposition algorithms are the Tucker [54] and the CP [19] decompositions. Intuitively, they both generalize SVD to tensors. However, one key problem with tensor decomposition is its computational complexity which may cause system bottleneck. Therefore, two phase block-centric CP tensor decomposition (2PCP) was proposed to partition the tensor into small sub-tensors, execute sub-tensor decomposition in parallel and combine the factors from each sub-tensor into final decomposition factors through iterative rerefinement process. Consequently, I proposed Sub-tensor Impact Graph (SIG) to account for inaccuracy propagation among sub-tensors and measure the impact of decomposition of sub-tensors on the other's decomposition, Based on SIG, I proposed several optimization strategies to optimize 2PCP's phase-2 refinement process. Furthermore, I applied SIG and optimization strategies for data focus, data evolution, and focus shifting in tensor analysis. Personalized Tensor Decomposition (PTD) is proposed to account for the users focus given the observations that in many applications, the user may have a focus of interest i.e., part of the data for which the user needs high accuracy and beyond this area focus, accuracy may not be as critical. PTD takes as input one or more areas of focus and performs the decomposition in such a way that, when reconstructed, the accuracy of the tensor is boosted for these areas of focus. A related challenge of data evolution in tensor analytics is incremental tensor decomposition since re-computation of the whole tensor decomposition with each update will cause high computational costs and incur large memory overheads. Especially for applications where data evolves over time and the tensor-based analysis results need to be continuouslymaintained. To avoid re-decomposition, I propose a two-phase block-incremental CP-based tensor decomposition technique, BICP, that efficiently and effectively maintains tensor decomposition results in the presence of dynamically evolving tensor data. I further extend the research focus on user focus shift. User focus may change over time as data is evolving along the time. Although PTD is efficient, re-computation for each user preference update can be the bottleneck for the system. Therefore I propose dynamic evolving user focus tensor decomposition which can smartly reuse the existing decomposition result to improve the efficiency of evolving user focus block decomposition.

Spatial data is fundamental in many applications like map services, land resource management, etc. Meanwhile, spatial data inherently comes with abundant context information because spatial entities themselves possess different properties, e.g., graph or textual information, etc. Among all these compound spatial data, geospatial graph data is one of the most challenging for the complexity of graph data. Graph data is commonly used to model real scenarios and searching for the matching subgraphs is fundamental in retrieving and analyzing graph data. With the ubiquity of spatial data, vertexes or edges in graphs are enriched with spatial location attributes side by side with other non-spatial attributes. Graph-based applications integrate spatial data into the graph model and provide more spatial-aware services. The co-existence of the graph and spatial data in the same geospatial graph triggers some new applications. To solve new problems in these applications, existing solutions develop an integrated system that incorporates the graph database and spatial database engines. However, existing approaches suffer from the architecture where graph data and spatial data are isolated. In this dissertation, I will explain two indexing frameworks, GeoReach and RisoTree, which can significantly accelerate the queries in geospatial graphs. GeoReach includes a query operator that adds spatial data awareness to a graph database management system. In GeoReach, the neighborhood spatial information is summarized and stored on each vertex in the graph. The summarization includes three different structures according to the location distribution. These spatial summaries are utilized to terminate the graph search early.RisoTree is a hierarchical tree structure where each node is represented by a minimum bounding rectangle (MBR). The MBR of a node is a rectangle that encloses all its children. A key difference between RisoTree and RTree is that RisoTree contains pre-materialized subgraph information to each index node. The subgraph information is utilized during the spatial index search phase to prune search paths that cannot satisfy the query graph pattern. The RisoTree index reduces the search space when the spatial filtering phase is performed with relatively light cost.

Many real-world problems, such as model- and data-driven computer simulation analysis, social and collaborative network analysis, brain data analysis, and so on, benefit from jointly modeling and analyzing the underlying patterns associated with complex, multi-relational data. Tensor decomposition is an ideal mathematical tool for this joint modeling, due to its simultaneous analysis of such multi-relational data, which is made possible by the data's multidimensional, array-based nature. A major challenge in tensor decomposition lies with its computational and space complexity, especially for dense datasets. While the process is comparatively faster for sparse tensors, decomposition is still a major bottleneck for many applications. The tensor decomposition process results in dense (hence, large) intermediate results, even when the input tensor is sparse (or small). Noise is another challenge for most data mining techniques, and many tensor decomposition schemes are sensitive to noisy datasets; this is an inevitable problem for real-world data, which can lead to false conclusions. In this dissertation, I develop innovative tensor decomposition algorithms for mining both sparse and dense multi-relational data in a noise-resistant way. I present novel, scalable, parallelizable tensor decomposition algorithms, specifically tuned to be effective for dense, noisy tensors, and which maintain the quality of the resulting analysis. Furthermore, I present results on multi-relational data applications focusing on model- and data-driven computer simulation analysis, as well as social network and web mining, which demonstrate the effectiveness of these tensor decompositions.

Learning programming involves a variety of complex cognitive activities, from abstract knowledge construction to structural operations, which include program design,modifying, debugging, and documenting tasks. In this work, the objective was to explore and investigate the barriers and obstacles that programming novice learners encountered and how the learners overcome them. Several lab and classroom studies were designed and conducted, the results showed that novice students had different behavior patterns compared to experienced learners, which indicates obstacles encountered. The studies also proved that proper assistance could help novices find helpful materials to read. However, novices still suffered from the lack of background knowledge and the limited cognitive load while learning, which resulted in challenges in understanding programming related materials, especially code examples. Therefore, I further proposed to use the natural language generator (NLG) to generate code explanations for educational purposes. The natural language generator is designed based on Long Short Term Memory (LSTM), a deep-learning translation model. To establish the model, a data set was collected from Amazon Mechanical Turks (AMT) recording explanations from human experts for programming code lines.

To evaluate the model, a pilot study was conducted and proved that the readability of the machine generated (MG) explanation was compatible with human explanations, while its accuracy is still not ideal, especially for complicated code lines. Furthermore, a code-example based learning platform was developed to utilize the explanation generating model in programming teaching. To examine the effect of code example explanations on different learners, two lab-class experiments were conducted separately ii in a programming novices’ class and an advanced students’ class. The experiment result indicated that when learning programming concepts, the MG code explanations significantly improved the learning Predictability for novices compared to control group, and the explanations also extended the novices’ learning time by generating more material to read, which potentially lead to a better learning gain. Besides, a completed correlation model was constructed according to the experiment result to illustrate the connections between different factors and the learning effect.

The recent proliferation of online platforms has not only revolutionized the way people communicate and acquire information but has also led to propagation of malicious information (e.g., online human trafficking, spread of misinformation, etc.). Propagation of such information occurs at unprecedented scale that could ultimately pose imminent societal-significant threats to the public. To better understand the behavior and impact of the malicious actors and counter their activity, social media authorities need to deploy certain capabilities to reduce their threats. Due to the large volume of this data and limited manpower, the burden usually falls to automatic approaches to identify these malicious activities. However, this is a subtle task facing online platforms due to several challenges: (1) malicious users have strong incentives to disguise themselves as normal users (e.g., intentional misspellings, camouflaging, etc.), (2) malicious users are high likely to be key users in making harmful messages go viral and thus need to be detected at their early life span to stop their threats from reaching a vast audience, and (3) available data for training automatic approaches for detecting malicious users, are usually either highly imbalanced (i.e., higher number of normal users than malicious users) or comprise insufficient labeled data.

To address the above mentioned challenges, in this dissertation I investigate the propagation of online malicious information from two broad perspectives: (1) content posted by users and (2) information cascades formed by resharing mechanisms in social media. More specifically, first, non-parametric and semi-supervised learning algorithms are introduced to discern potential patterns of human trafficking activities that are of high interest to law enforcement. Second, a time-decay causality-based framework is introduced for early detection of “Pathogenic Social Media (PSM)” accounts (e.g., terrorist supporters). Third, due to the lack of sufficient annotated data for training PSM detection approaches, a semi-supervised causal framework is proposed that utilizes causal-related attributes from unlabeled instances to compensate for the lack of enough labeled data. Fourth, a feature-driven approach for PSM detection is introduced that leverages different sets of attributes from users’ causal activities, account-level and content-related information as well as those from URLs shared by users.

The pervasive use of the Web has connected billions of people all around the globe and enabled them to obtain information at their fingertips. This results in tremendous amounts of user-generated data which makes users traceable and vulnerable to privacy leakage attacks. In general, there are two types of privacy leakage attacks for user-generated data, i.e., identity disclosure and private-attribute disclosure attacks. These attacks put users at potential risks ranging from persecution by governments to targeted frauds. Therefore, it is necessary for users to be able to safeguard their privacy without leaving their unnecessary traces of online activities. However, privacy protection comes at the cost of utility loss defined as the loss in quality of personalized services users receive. The reason is that this information of traces is crucial for online vendors to provide personalized services and a lack of it would result in deteriorating utility. This leads to a dilemma of privacy and utility.

Protecting users' privacy while preserving utility for user-generated data is a challenging task. The reason is that users generate different types of data such as Web browsing histories, user-item interactions, and textual information. This data is heterogeneous, unstructured, noisy, and inherently different from relational and tabular data and thus requires quantifying users' privacy and utility in each context separately. In this dissertation, I investigate four aspects of protecting user privacy for user-generated data. First, a novel adversarial technique is introduced to assay privacy risks in heterogeneous user-generated data. Second, a novel framework is proposed to boost users' privacy while retaining high utility for Web browsing histories. Third, a privacy-aware recommendation system is developed to protect privacy w.r.t. the rich user-item interaction data by recommending relevant and privacy-preserving items. Fourth, a privacy-preserving framework for text representation learning is presented to safeguard user-generated textual data as it can reveal private information.

Machine learning (ML) and deep neural networks (DNNs) have achieved great success in a variety of application domains, however, despite significant effort to make these networks robust, they remain vulnerable to adversarial attacks in which input that is perceptually indistinguishable from natural data can be erroneously classified with high prediction confidence. Works on defending against adversarial examples can be broadly classified as correcting or detecting, which aim, respectively at negating the effects of the attack and correctly classifying the input, or detecting and rejecting the input as adversarial. In this work, a new approach for detecting adversarial examples is proposed. The approach takes advantage of the robustness of natural images to noise. As noise is added to a natural image, the prediction probability of its true class drops, but the drop is not sudden or precipitous. The same seems to not hold for adversarial examples. In other word, the stress response profile for natural images seems different from that of adversarial examples, which could be detected by their stress response profile. An evaluation of this approach for detecting adversarial examples is performed on the MNIST, CIFAR-10 and ImageNet datasets. Experimental data shows that this approach is effective at detecting some adversarial examples on small scaled simple content images and with little sacrifice on benign accuracy.