Theses and dissertations completed in the Data Lab, newest first. Every entry links to the full text on SURFACE, Syracuse University’s institutional repository. For the lab’s papers, see our publications page.
PhD Theses
Tian, Hao
Exploring Higher-Order Networks PhD Thesis
Syracuse University, 2025, (Ph.D. Dissertation, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.).
@phdthesis{tian2025exploringhigherordernetworks,
title = {Exploring Higher-Order Networks},
author = {Hao Tian},
url = {https://surface.syr.edu/cgi/viewcontent.cgi?article=3147&context=etd, PDF
https://surface.syr.edu/etd/2062/, SURFACE record},
year = {2025},
date = {2025-05-11},
address = {Syracuse, NY, USA},
school = {Syracuse University},
abstract = {Networks are natural representations of interactions in the real world (social networks, bio-networks, road networks, and the like) and are utilized across various disciplines. By default, network interactions are pairwise; in recent years, the demand for model ing higher-order interactions has kept increasing. For example, in collaboration networks, we aim to distinguish between one publication coauthored by three or three publications coauthored by two in a triangle. In this work, we perform higher-order network analysis in the following two directions. First, we explore the influence of higher-order structures on dyadic (pairwise) graphs; second, we model higher-order interactions by ordered hy pergraphs and analyze their cross-order properties. Followed by applications such as link prediction and representation learning, we verify the effectiveness of higher-order network modelings. In dyadic graphs, specific subgraph patterns with high frequencies are called network mo tifs. Inspired by the study of motifs, we first introduce a neighborhood motif-based link prediction method. Through extensive observation of real-world graphs, we found that the network structure of the neighborhood has a strong influence on getting two nodes connected. Specifically, a more well-connected and compact structure leads to a higher probability of wiring rather than a sparser structure. Based on this discovery, we propose a novel scoring method based on neighborhood motifs. By learning a γ-decaying model, we measure the pairwise similarities between nodes more accurately, even when only using the information of the common neighbors, which is often used in current techniques. We further model higher-order interactions directly using ordered hypergraphs. We ex plore and quantify similarities among various orders of the network. Our goal is to build relationships between different network orders and to solve higher-order problems using lower-order information. Similarities between different orders are not directly compara ble. Hence, we introduce a set of general cross-order similarities and a measure: subedge rate. Our experiments on multiple real-world datasets demonstrate that most higher-order networks have considerable consistency as we move from higher orders to lower orders. Using this discovery, we develop a new cross-order framework for a higher-order link pre diction method. These methods can predict higher-order links from lower-order edges, which cannot be attained by current higher-order methods that rely on data from a single order. Using the same topology, we explore the relationship between higher-order structure and spectral properties. These spectral properties are closely related to the structural properties of dyadic graphs. We generalize such connections and characterize higher-order networks by their spectral information. Specifically, we prove that the second moment provides an upper bound on the degree distributions while the third moment provides an upper bound on the number of on triangles. With the theoretical support, we use all such spec tral moments across different orders as the higher-order graph representation. Extensive experiments show the utility of this new representation in various settings. For instance, graph classification on higher-order graphs shows that this representation significantly out performs other techniques.},
note = {Ph.D. Dissertation, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}
Networks are natural representations of interactions in the real world (social networks, bio-networks, road networks, and the like) and are utilized across various disciplines. By default, network interactions are pairwise; in recent years, the demand for model ing higher-order interactions has kept increasing. For example, in collaboration networks, we aim to distinguish between one publication coauthored by three or three publications coauthored by two in a triangle. In this work, we perform higher-order network analysis in the following two directions. First, we explore the influence of higher-order structures on dyadic (pairwise) graphs; second, we model higher-order interactions by ordered hy pergraphs and analyze their cross-order properties. Followed by applications such as link prediction and representation learning, we verify the effectiveness of higher-order network modelings. In dyadic graphs, specific subgraph patterns with high frequencies are called network mo tifs. Inspired by the study of motifs, we first introduce a neighborhood motif-based link prediction method. Through extensive observation of real-world graphs, we found that the network structure of the neighborhood has a strong influence on getting two nodes connected. Specifically, a more well-connected and compact structure leads to a higher probability of wiring rather than a sparser structure. Based on this discovery, we propose a novel scoring method based on neighborhood motifs. By learning a γ-decaying model, we measure the pairwise similarities between nodes more accurately, even when only using the information of the common neighbors, which is often used in current techniques. We further model higher-order interactions directly using ordered hypergraphs. We ex plore and quantify similarities among various orders of the network. Our goal is to build relationships between different network orders and to solve higher-order problems using lower-order information. Similarities between different orders are not directly compara ble. Hence, we introduce a set of general cross-order similarities and a measure: subedge rate. Our experiments on multiple real-world datasets demonstrate that most higher-order networks have considerable consistency as we move from higher orders to lower orders. Using this discovery, we develop a new cross-order framework for a higher-order link pre diction method. These methods can predict higher-order links from lower-order edges, which cannot be attained by current higher-order methods that rely on data from a single order. Using the same topology, we explore the relationship between higher-order structure and spectral properties. These spectral properties are closely related to the structural properties of dyadic graphs. We generalize such connections and characterize higher-order networks by their spectral information. Specifically, we prove that the second moment provides an upper bound on the degree distributions while the third moment provides an upper bound on the number of on triangles. With the theoretical support, we use all such spec tral moments across different orders as the higher-order graph representation. Extensive experiments show the utility of this new representation in various settings. For instance, graph classification on higher-order graphs shows that this representation significantly out performs other techniques.
Abdolazimi, Reyhaneh
Noise-Enhanced Network Science PhD Thesis
Syracuse University, 2024, (Ph.D. Dissertation, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.).
@phdthesis{abdolazimi2024noiseenhancednetworkscience,
title = {Noise-Enhanced Network Science},
author = {Reyhaneh Abdolazimi},
url = {https://surface.syr.edu/cgi/viewcontent.cgi?article=3132&context=etd, PDF
https://surface.syr.edu/etd/2047/, SURFACE record},
year = {2024},
date = {2024-12-20},
address = {Syracuse, NY, USA},
school = {Syracuse University},
abstract = {Graphs are a versatile and powerful data structure used to model complex relationships in diverse domains such as social networks, biological systems, and transportation networks. In graphs, entities are represented by nodes, and interactions or relationships between them are represented by edges. For example, in the World Wide Web, web pages are considered vertices, and if there is a link from one page to another page on the web, there will be a directed edge between those pages in its graph data structure. These structures allow us to analyze and solve important problems like community detection (identifying closely related clusters of nodes) and link prediction (forecasting missing or future connections between nodes). Graph algorithms, therefore, are fundamental tools in applications ranging from e-commerce recommendation engines to biological research and road network optimization. Despite significant advances, many challenges remain in improving the effectiveness and accuracy of graph algorithms. Traditional methods for community detection, link prediction, and classification have limitations, particularly in handling noise—typically viewed as a disruptive element that decreases performance. However, insights from fields like signal processing suggest that noise, when applied appropriately, can enhance system performance. This thesis introduces a novel framework that leverages noise to improve graph algorithms. By using noise as a tool rather than an obstacle, the thesis presents new techniques that modify graph data structures to enhance the performance of community detection, link prediction, and classification. Through theoretical justification and empirical experiments on real-world and synthetic datasets, it is demonstrated that adding noise to graph data can enhance the robustness and accuracy of these algorithms. The proposed noise-enhanced methods not only improve the quality of results in tasks like community detection and link prediction but also mitigate over-smoothing in GNNs, enabling them to generalize better to large and complex graphs. This research offers a paradigm shift in how noise is perceived in graph algorithms, transforming it from an obstacle to a powerful tool for optimization. The findings have broad implications for improving graph-based models and applications, paving the way for future advancements in fields ranging from social network analysis to biological research and beyond.},
note = {Ph.D. Dissertation, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}
Graphs are a versatile and powerful data structure used to model complex relationships in diverse domains such as social networks, biological systems, and transportation networks. In graphs, entities are represented by nodes, and interactions or relationships between them are represented by edges. For example, in the World Wide Web, web pages are considered vertices, and if there is a link from one page to another page on the web, there will be a directed edge between those pages in its graph data structure. These structures allow us to analyze and solve important problems like community detection (identifying closely related clusters of nodes) and link prediction (forecasting missing or future connections between nodes). Graph algorithms, therefore, are fundamental tools in applications ranging from e-commerce recommendation engines to biological research and road network optimization. Despite significant advances, many challenges remain in improving the effectiveness and accuracy of graph algorithms. Traditional methods for community detection, link prediction, and classification have limitations, particularly in handling noise—typically viewed as a disruptive element that decreases performance. However, insights from fields like signal processing suggest that noise, when applied appropriately, can enhance system performance. This thesis introduces a novel framework that leverages noise to improve graph algorithms. By using noise as a tool rather than an obstacle, the thesis presents new techniques that modify graph data structures to enhance the performance of community detection, link prediction, and classification. Through theoretical justification and empirical experiments on real-world and synthetic datasets, it is demonstrated that adding noise to graph data can enhance the robustness and accuracy of these algorithms. The proposed noise-enhanced methods not only improve the quality of results in tasks like community detection and link prediction but also mitigate over-smoothing in GNNs, enabling them to generalize better to large and complex graphs. This research offers a paradigm shift in how noise is perceived in graph algorithms, transforming it from an obstacle to a powerful tool for optimization. The findings have broad implications for improving graph-based models and applications, paving the way for future advancements in fields ranging from social network analysis to biological research and beyond.
Li, Jiayu
Enhancing Graph Neural Networks by Editing Graphs PhD Thesis
Syracuse University, 2024, (Ph.D. Dissertation, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.).
@phdthesis{li2024enhancinggraphneuralnetworks,
title = {Enhancing Graph Neural Networks by Editing Graphs},
author = {Jiayu Li},
url = {https://surface.syr.edu/cgi/viewcontent.cgi?article=3130&context=etd, PDF
https://surface.syr.edu/etd/2045/, SURFACE record},
year = {2024},
date = {2024-12-20},
address = {Syracuse, NY, USA},
school = {Syracuse University},
abstract = {Graphs are pervasive in both the natural world and various domains of science and engineering. Numerous advanced classifiers, such as Graph Neural Networks (GNNs), have been developed to perform node classification on these graphs. However, as graphs become denser with an increasing number of edges, GNNs often suffer from suboptimal generalization performance due to the presence of task-irrelevant connections. These redundant connections can introduce noise, consume excessive computational resources, and degrade performance. Identifying and preserving critical connections in large-scale graphs, while pruning unnecessary ones, is crucial for enhancing the efficiency and accuracy of GNNs in node classification, particularly for GCNs. As a result, the first challenge is determining whether it is possible to identify skeletal substructures within graphs—key subsets of connections—that can preserve the predictive performance of node classification. This involves either maintaining accuracy when using these critical subgraphs instead of the entire graph or training classifiers on these subgraphs with minimal performance degradation. To address the challenge, we propose the Sparsified Graph Convolutional Network (SGCN), a neural network-based graph sparsifier. SGCN effectively reduces graph density by pruning certain edges while still maintaining comparable performance in node classification tasks. To further improve the efficiency of GNN with even more sparse graph, we introduce the Graph Ultra-sparsifier, a semi-supervised graph sparsification method that incorporates dynamically-updated regularization terms derived from graph convolution. This approach preserves the properties of graph filters, enabling the generation of sparser graphs that maintain the performance of GCN models when used as input. To further explore the significance of key subsets of connections, we propose a sparse adversarial attack framework, AdverSparse. This framework demonstrates how the removal of just a few critical connections can severely disrupt the spatial dependencies learned by spatial-temporal models, leading to issues such as increased prediction errors. While graph sparsification effectively addresses the space and time consumption issues of GCNs while maintaining their performance, existing graph editing methods are specifically designed for the low-pass filter structure of GCNs, which inherently favor homophilic graphs. However, real-world graphs often exhibit varying levels of homophily. To further enhance the capability of GCNs on graphs with diverse homophily ratios, it is essential to go beyond merely removing edges. Instead, we propose dynamically assigning weights to both existing edges and hidden edges (i.e., edges from complement graphs). This introduces a second challenge: dynamically editing graphs. The goal is to modify graphs by incorporating edges from both the original graph and its complement in a way that allows them to serve as adaptive graph filters. Such filters should dynamically adjust their behavior based on the homophily ratio of the graph, improving node classification performance. To address the second challenge, we introduce Complement Graph Convolution (CGC), a method designed to overcome the limitations of GCNs and enhance the performance of GNNs that use GCNs as basic modules for node classification tasks across graphs with varying homophily levels. CGC incorporates a trainable, adaptive frequency-response filter that utilizes both an original graph and its complement graph. Spectral analysis reveals that this filter can be expressed as a Bernstein polynomial approximation, allowing it to dynamically adjust the importance of different frequency components for spectral convolution, effectively accommodating diverse homophily and heterophily ratios. Furthermore, we demonstrate that CGC mitigates the over-smoothing problem commonly associated with GCNs. The overall objective of this thesis is to explore methods for enhancing the efficiency and capability of GNNs, particularly the GCN family, by editing graph structures to affect the properties of graph filters.},
note = {Ph.D. Dissertation, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}
Graphs are pervasive in both the natural world and various domains of science and engineering. Numerous advanced classifiers, such as Graph Neural Networks (GNNs), have been developed to perform node classification on these graphs. However, as graphs become denser with an increasing number of edges, GNNs often suffer from suboptimal generalization performance due to the presence of task-irrelevant connections. These redundant connections can introduce noise, consume excessive computational resources, and degrade performance. Identifying and preserving critical connections in large-scale graphs, while pruning unnecessary ones, is crucial for enhancing the efficiency and accuracy of GNNs in node classification, particularly for GCNs. As a result, the first challenge is determining whether it is possible to identify skeletal substructures within graphs—key subsets of connections—that can preserve the predictive performance of node classification. This involves either maintaining accuracy when using these critical subgraphs instead of the entire graph or training classifiers on these subgraphs with minimal performance degradation. To address the challenge, we propose the Sparsified Graph Convolutional Network (SGCN), a neural network-based graph sparsifier. SGCN effectively reduces graph density by pruning certain edges while still maintaining comparable performance in node classification tasks. To further improve the efficiency of GNN with even more sparse graph, we introduce the Graph Ultra-sparsifier, a semi-supervised graph sparsification method that incorporates dynamically-updated regularization terms derived from graph convolution. This approach preserves the properties of graph filters, enabling the generation of sparser graphs that maintain the performance of GCN models when used as input. To further explore the significance of key subsets of connections, we propose a sparse adversarial attack framework, AdverSparse. This framework demonstrates how the removal of just a few critical connections can severely disrupt the spatial dependencies learned by spatial-temporal models, leading to issues such as increased prediction errors. While graph sparsification effectively addresses the space and time consumption issues of GCNs while maintaining their performance, existing graph editing methods are specifically designed for the low-pass filter structure of GCNs, which inherently favor homophilic graphs. However, real-world graphs often exhibit varying levels of homophily. To further enhance the capability of GCNs on graphs with diverse homophily ratios, it is essential to go beyond merely removing edges. Instead, we propose dynamically assigning weights to both existing edges and hidden edges (i.e., edges from complement graphs). This introduces a second challenge: dynamically editing graphs. The goal is to modify graphs by incorporating edges from both the original graph and its complement in a way that allows them to serve as adaptive graph filters. Such filters should dynamically adjust their behavior based on the homophily ratio of the graph, improving node classification performance. To address the second challenge, we introduce Complement Graph Convolution (CGC), a method designed to overcome the limitations of GCNs and enhance the performance of GNNs that use GCNs as basic modules for node classification tasks across graphs with varying homophily levels. CGC incorporates a trainable, adaptive frequency-response filter that utilizes both an original graph and its complement graph. Spectral analysis reveals that this filter can be expressed as a Bernstein polynomial approximation, allowing it to dynamically adjust the importance of different frequency components for spectral convolution, effectively accommodating diverse homophily and heterophily ratios. Furthermore, we demonstrate that CGC mitigates the over-smoothing problem commonly associated with GCNs. The overall objective of this thesis is to explore methods for enhancing the efficiency and capability of GNNs, particularly the GCN family, by editing graph structures to affect the properties of graph filters.
Jin, Shengmin
Interpretable Network Representations PhD Thesis
Syracuse University, 2022, (Ph.D. Dissertation, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.).
@phdthesis{jin2022interpretablenetworkrepresentations,
title = {Interpretable Network Representations},
author = {Shengmin Jin},
url = {https://surface.syr.edu/cgi/viewcontent.cgi?article=2607&context=etd, PDF
https://surface.syr.edu/etd/1606/, SURFACE record},
year = {2022},
date = {2022-12-16},
address = {Syracuse, NY, USA},
school = {Syracuse University},
abstract = {Networks (or interchangeably graphs) have been ubiquitous across the globe and within science and engineering: social networks, collaboration networks, protein-protein interaction networks, infrastructure networks, among many others. Machine learning on graphs, especially network representation learning, has shown remarkable performance in network-based applications, such as node/graph classification, graph clustering, and link prediction. Like performance, it is equally crucial for individuals to understand the behavior of machine learning models and be able to explain how these models arrive at a certain decision. Such needs have motivated many studies on interpretability in machine learning. For example, for social network analysis, we may need to know the reasons why certain users (or groups) are classified or clustered together by the machine learning models, or why a friend recommendation system considers some users similar so that they are recommended to connect with each other. Therefore, an interpretable network representation is necessary and it should carry the graph information to a level understandable by humans. Here, we first introduce our method on interpretable network representations: the network shape. It provides a framework to represent a network with a 3-dimensional shape, and one can customize network shapes for their need, by choosing various graph sampling methods, 3D network embedding methods and shape-fitting methods. In this thesis, we introduce the two types of network shape: a Kronecker hull which represents a network as a 3D convex polyhedron using stochastic Kronecker graphs as the network embedding method, and a Spectral Path which represents a network as a 3D path connecting the spectral moments of the network and its subgraphs. We demonstrate that network shapes can capture various properties of not only the network, but also its subgraphs. For instance, they can provide the distribution of subgraphs within a network, e.g., what proportion of subgraphs are structurally similar to the whole network? Network shapes are interpretable on different levels, so one can quickly understand the structural properties of a network and its subgraphs by its network shape. Using experiments on real-world networks, we demonstrate that network shapes can be used in various applications, including (1) network visualization, the most intuitive way for users to understand a graph; (2) network categorization (e.g., is this a social or a biological network?); (3) computing similarity between two graphs. Moreover, we utilize network shapes to extend biometrics studies to network data, by solving two problems: network identification (Given an anonymized graph, can we identify the network from which it is collected? i.e., answering questions such as ``where is this anonymized graph sampled from, Twitter or Facebook?") and network authentication (If one claims the graph is sampled from a certain network, can we verify this claim?). The overall objective of the thesis is to provide a compact, interpretable, visualizable, comparable and efficient representation of networks.},
note = {Ph.D. Dissertation, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}
Networks (or interchangeably graphs) have been ubiquitous across the globe and within science and engineering: social networks, collaboration networks, protein-protein interaction networks, infrastructure networks, among many others. Machine learning on graphs, especially network representation learning, has shown remarkable performance in network-based applications, such as node/graph classification, graph clustering, and link prediction. Like performance, it is equally crucial for individuals to understand the behavior of machine learning models and be able to explain how these models arrive at a certain decision. Such needs have motivated many studies on interpretability in machine learning. For example, for social network analysis, we may need to know the reasons why certain users (or groups) are classified or clustered together by the machine learning models, or why a friend recommendation system considers some users similar so that they are recommended to connect with each other. Therefore, an interpretable network representation is necessary and it should carry the graph information to a level understandable by humans. Here, we first introduce our method on interpretable network representations: the network shape. It provides a framework to represent a network with a 3-dimensional shape, and one can customize network shapes for their need, by choosing various graph sampling methods, 3D network embedding methods and shape-fitting methods. In this thesis, we introduce the two types of network shape: a Kronecker hull which represents a network as a 3D convex polyhedron using stochastic Kronecker graphs as the network embedding method, and a Spectral Path which represents a network as a 3D path connecting the spectral moments of the network and its subgraphs. We demonstrate that network shapes can capture various properties of not only the network, but also its subgraphs. For instance, they can provide the distribution of subgraphs within a network, e.g., what proportion of subgraphs are structurally similar to the whole network? Network shapes are interpretable on different levels, so one can quickly understand the structural properties of a network and its subgraphs by its network shape. Using experiments on real-world networks, we demonstrate that network shapes can be used in various applications, including (1) network visualization, the most intuitive way for users to understand a graph; (2) network categorization (e.g., is this a social or a biological network?); (3) computing similarity between two graphs. Moreover, we utilize network shapes to extend biometrics studies to network data, by solving two problems: network identification (Given an anonymized graph, can we identify the network from which it is collected? i.e., answering questions such as ``where is this anonymized graph sampled from, Twitter or Facebook?") and network authentication (If one claims the graph is sampled from a certain network, can we verify this claim?). The overall objective of the thesis is to provide a compact, interpretable, visualizable, comparable and efficient representation of networks.
Zhou, Xinyi
Algorithmic Solutions to Combat Online Fake News PhD Thesis
Syracuse University, 2022, (Ph.D. Dissertation, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.).
@phdthesis{zhou2022algorithmicsolutionsfakenews,
title = {Algorithmic Solutions to Combat Online Fake News},
author = {Xinyi Zhou},
url = {https://surface.syr.edu/cgi/viewcontent.cgi?article=2600&context=etd, PDF
https://surface.syr.edu/etd/1599/, SURFACE record},
year = {2022},
date = {2022-12-09},
address = {Syracuse, NY, USA},
school = {Syracuse University},
abstract = {The unprecedented growth of new information producing, distributing, and consuming every moment on the Web has fostered the rise of ``fake news.'' Because of its detrimental effect on democracy, global economies, and public health, effectively combating online fake news has become an essential and urgent task. This dissertation starts with making typological, theoretical, and empirical efforts to promote the public's comprehension of fake news and lay the foundation for algorithmically combating fake news. As there has been no universal definition of fake news, this dissertation discusses the definition of fake news from three dimensions: veracity, intention, and news, comparing it with related terms, such as misinformation and disinformation. The dissertation first probes and collects extensive theories in social sciences, presenting or interpreting the psychology, behavior, and motivations of human beings as fake news producers, distributors, and consumers. It creates real-world multimodal, multilingual, and cross-site datasets, with which the dissertation empirically characterizes the language of fake news and its propagation on social networks differential from the truth. Beyond understanding fake news, this dissertation presents novel machine (deep) learning algorithms for accurate, explainable, early, and robust prediction of fake news. It first introduces social theories and empirical patterns of fake news into feature extraction. It designs the neural networks that explicitly and adaptively capture the linguistic style of various news articles (i.e., the usage of words and the linguistically meaningful way they are structured into documents). It first leverages multimodal news content and cross-modal consistency to predict fake news. The proposed algorithms comprehensively investigate news language across the lexical, syntactic, semantic, and discourse levels, the visual information within news content, and news diffusion on social networks across the node, ego, triad, community, and network levels. Their effectiveness in predicting fake news is demonstrated with real-world datasets publicly available. Furthermore, this dissertation strives for proactive fake news mitigation, considering that predicting fake news can be effective but reactive in countering online fake news. It formulates a new task of assessing the intent of fake news spreaders to keep social media users from unintentionally circulating any future fake news without realizing its fakeness. It proposes a social-theory-informed AI-powered solution. Specifically, social theories interpret why a human unintentionally spreads fake news (i.e., preexisting beliefs and social influence). Advanced AI (artificial intelligence) techniques are employed to compute one's beliefs and received social influence. It first annotates the intent of fake news spreaders as ground truth, with which we demonstrate the proposed solution's effectiveness.},
note = {Ph.D. Dissertation, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}
The unprecedented growth of new information producing, distributing, and consuming every moment on the Web has fostered the rise of ``fake news.'' Because of its detrimental effect on democracy, global economies, and public health, effectively combating online fake news has become an essential and urgent task. This dissertation starts with making typological, theoretical, and empirical efforts to promote the public's comprehension of fake news and lay the foundation for algorithmically combating fake news. As there has been no universal definition of fake news, this dissertation discusses the definition of fake news from three dimensions: veracity, intention, and news, comparing it with related terms, such as misinformation and disinformation. The dissertation first probes and collects extensive theories in social sciences, presenting or interpreting the psychology, behavior, and motivations of human beings as fake news producers, distributors, and consumers. It creates real-world multimodal, multilingual, and cross-site datasets, with which the dissertation empirically characterizes the language of fake news and its propagation on social networks differential from the truth. Beyond understanding fake news, this dissertation presents novel machine (deep) learning algorithms for accurate, explainable, early, and robust prediction of fake news. It first introduces social theories and empirical patterns of fake news into feature extraction. It designs the neural networks that explicitly and adaptively capture the linguistic style of various news articles (i.e., the usage of words and the linguistically meaningful way they are structured into documents). It first leverages multimodal news content and cross-modal consistency to predict fake news. The proposed algorithms comprehensively investigate news language across the lexical, syntactic, semantic, and discourse levels, the visual information within news content, and news diffusion on social networks across the node, ego, triad, community, and network levels. Their effectiveness in predicting fake news is demonstrated with real-world datasets publicly available. Furthermore, this dissertation strives for proactive fake news mitigation, considering that predicting fake news can be effective but reactive in countering online fake news. It formulates a new task of assessing the intent of fake news spreaders to keep social media users from unintentionally circulating any future fake news without realizing its fakeness. It proposes a social-theory-informed AI-powered solution. Specifically, social theories interpret why a human unintentionally spreads fake news (i.e., preexisting beliefs and social influence). Advanced AI (artificial intelligence) techniques are employed to compute one's beliefs and received social influence. It first annotates the intent of fake news spreaders as ground truth, with which we demonstrate the proposed solution's effectiveness.
Masters Theses
Cai, Weibin
Harnessing LLMs to Detect Hate Speech Masters Thesis
Syracuse University, Syracuse, NY, USA, 2025, (M.S. Thesis, Department of Information Science & Technology, Syracuse University. Advisor: Reza Zafarani.).
@mastersthesis{cai2025harnessingllmshatespeech,
title = {Harnessing LLMs to Detect Hate Speech},
author = {Weibin Cai},
url = {https://surface.syr.edu/cgi/viewcontent.cgi?article=1934&context=thesis, PDF
https://surface.syr.edu/thesis/933/, SURFACE record},
year = {2025},
date = {2025-05-11},
address = {Syracuse, NY, USA},
school = {Syracuse University},
abstract = {Hate is a sentiment, while hate speech refers to the expression of hate in a form that targets and attacks specific groups, such as race, religion, or gender. With the rise of the internet and social media, hate speech has spread rapidly, gaining wide exposure and posing threats to individual well-being, the profits of major tech companies, and social stability. As a result, both industry and academia have turned their attention to the study of hate speech. One of the most active areas is hate speech detection, which involves training models to predict whether a given piece of content is hateful. However, the choice of models and methods can vary depending on the form in which the hate speech is conveyed. In this thesis, we address two key issues: 1.In the task of hateful meme classification, many existing approaches focus on stacking model parameters to achieve better performance, but lack a deep understanding of how hateful memes are constructed. Furthermore, they have not effectively leveraged large language models (LLMs) for this task. 2.Although the definitions of hate and hate speech are well-established, individuals’ perceptions of hate can vary due to differences in cultural background. As a result, judgments about whether a piece of content is hateful may differ from person to person. However, current hate speech detection models typically rely on labels obtained through majority voting, without accounting for the cultural specificity of individual annotators. To address the first issue, we observe that creators of hateful memes often exaggerate their emotions and reinforce stereotypes, leading to a mismatch between the text and image within the meme. We refer to this phenomenon as a false claim. Based on this insight, we propose the FACT model(FAlse Claim haTeful meme classification model), a model built upon a large language model(LLM), which identifies false claims in memes to assist the classification process. To address the second issue, we first evaluate LLMs and find that they are unable to effectively utilize cultural background information to support reasoning, while historical labeling proves to be useful. Based on this, we hypothesize that an individual’s perception of hate is influenced by specific combinations of cultural background factors. To incorporate this insight, we apply matrix factorization techniques from recommender systems to learn interaction features for each cultural background combination. These features are then used to support culture-aware hate speech detection.},
note = {M.S. Thesis, Department of Information Science & Technology, Syracuse University. Advisor: Reza Zafarani.},
keywords = {},
pubstate = {published},
tppubtype = {mastersthesis}
}
Hate is a sentiment, while hate speech refers to the expression of hate in a form that targets and attacks specific groups, such as race, religion, or gender. With the rise of the internet and social media, hate speech has spread rapidly, gaining wide exposure and posing threats to individual well-being, the profits of major tech companies, and social stability. As a result, both industry and academia have turned their attention to the study of hate speech. One of the most active areas is hate speech detection, which involves training models to predict whether a given piece of content is hateful. However, the choice of models and methods can vary depending on the form in which the hate speech is conveyed. In this thesis, we address two key issues: 1.In the task of hateful meme classification, many existing approaches focus on stacking model parameters to achieve better performance, but lack a deep understanding of how hateful memes are constructed. Furthermore, they have not effectively leveraged large language models (LLMs) for this task. 2.Although the definitions of hate and hate speech are well-established, individuals’ perceptions of hate can vary due to differences in cultural background. As a result, judgments about whether a piece of content is hateful may differ from person to person. However, current hate speech detection models typically rely on labels obtained through majority voting, without accounting for the cultural specificity of individual annotators. To address the first issue, we observe that creators of hateful memes often exaggerate their emotions and reinforce stereotypes, leading to a mismatch between the text and image within the meme. We refer to this phenomenon as a false claim. Based on this insight, we propose the FACT model(FAlse Claim haTeful meme classification model), a model built upon a large language model(LLM), which identifies false claims in memes to assist the classification process. To address the second issue, we first evaluate LLMs and find that they are unable to effectively utilize cultural background information to support reasoning, while historical labeling proves to be useful. Based on this, we hypothesize that an individual’s perception of hate is influenced by specific combinations of cultural background factors. To incorporate this insight, we apply matrix factorization techniques from recommender systems to learn interaction features for each cultural background combination. These features are then used to support culture-aware hate speech detection.
Gopalakrishnan, Atul Anand
Graph Augmentation using Spectral Moments Masters Thesis
Syracuse University, Syracuse, NY, USA, 2023, (M.S. Thesis, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.).
@mastersthesis{gopalakrishnan2023graphaugmentationspectralmoments,
title = {Graph Augmentation using Spectral Moments},
author = {Atul Anand Gopalakrishnan},
url = {https://surface.syr.edu/cgi/viewcontent.cgi?article=1702&context=thesis, PDF
https://surface.syr.edu/thesis/701/, SURFACE record},
year = {2023},
date = {2023-05-14},
address = {Syracuse, NY, USA},
school = {Syracuse University},
abstract = {Graph representational learning focuses on learning real value vectors that for nodes,edges or the graph, such that these vectors capture adequate information about these entities. Graph data augmentation, focuses on changing the structure or features in a graph to help improve classification performance and become more generalizable. This can be broadly categorized into feature based augmentation and structure based augmentation. Feature augmentation focuses on changing the feature matrix, without changing the structure of the graph to help improve the performance of the graph neural network. Graph structure augmentation refers to the manipulation of the adjacency matrix of a given graph to achieve better classification performance. Our approach focuses on the problem of graph augmentation but from a spectral standpoint. More specifically, we attempt to augment a graph using spectral moments. Recent results have indicated that the second, third and fourth spectral moments of a graph, have strong connections to the graph's properties, such as degree distribution, clustering coefficient, and connectivity[1]. Our contribution is two fold: First, we explain a formal method to find a spectral moment that helps maximize node classification performance. Second, we also provide an algorithm to augment the graph using it's spectral moments, and therefore augment the graph to the spectral point that helps maximize classification performance while making the graph sparse. For the purpose of node classification, we use the GraphSAGE model with no node sampling and the mean aggregator. We notice that the node classification performance after augmentation goes up in a majority of our datasets, and furthermore, the graph also gets sparser across all our datasets.},
note = {M.S. Thesis, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.},
keywords = {},
pubstate = {published},
tppubtype = {mastersthesis}
}
Graph representational learning focuses on learning real value vectors that for nodes,edges or the graph, such that these vectors capture adequate information about these entities. Graph data augmentation, focuses on changing the structure or features in a graph to help improve classification performance and become more generalizable. This can be broadly categorized into feature based augmentation and structure based augmentation. Feature augmentation focuses on changing the feature matrix, without changing the structure of the graph to help improve the performance of the graph neural network. Graph structure augmentation refers to the manipulation of the adjacency matrix of a given graph to achieve better classification performance. Our approach focuses on the problem of graph augmentation but from a spectral standpoint. More specifically, we attempt to augment a graph using spectral moments. Recent results have indicated that the second, third and fourth spectral moments of a graph, have strong connections to the graph's properties, such as degree distribution, clustering coefficient, and connectivity[1]. Our contribution is two fold: First, we explain a formal method to find a spectral moment that helps maximize node classification performance. Second, we also provide an algorithm to augment the graph using it's spectral moments, and therefore augment the graph to the spectral point that helps maximize classification performance while making the graph sparse. For the purpose of node classification, we use the GraphSAGE model with no node sampling and the mean aggregator. We notice that the node classification performance after augmentation goes up in a majority of our datasets, and furthermore, the graph also gets sparser across all our datasets.
Mulay, Apurva Shriniwas
On Stability and Similarity of Network Embeddings Masters Thesis
Syracuse University, Syracuse, NY, USA, 2021, (M.S. Thesis, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.).
@mastersthesis{mulay2021stabilitysimilaritynetworkembeddings,
title = {On Stability and Similarity of Network Embeddings},
author = {Apurva Shriniwas Mulay},
url = {https://surface.syr.edu/cgi/viewcontent.cgi?article=1528&context=thesis, PDF
https://surface.syr.edu/thesis/527/, SURFACE record},
year = {2021},
date = {2021-05-23},
address = {Syracuse, NY, USA},
school = {Syracuse University},
abstract = {Machine Learning on graphs has become an active research area due to the prevailing graph-structured data in the real world. Many real-world applications can be modeled with graphs. Modern application domains include web-scale social networks [26], recommender systems, knowledge graphs, and biological or protein networks. However, there are various challenges. First, the graphs generated from such applications are often large. Moreover, in some scenarios, the complete graph is not available, e.g., for privacy reasons. Thus, it becomes impractical to perform network analysis or compute various graph measures. Hence, graph sampling becomes an important task.Sampling is often the first step to handle any type of massive data. The same applies to graphs as well, which leads to many graph sampling techniques. Sampling Techniques include Node-based (e.g., Random Node Sampling), Edge-based (e.g., Random Edge Sampling) and Traversal-based (e.g., Random Walk Sampling). Graphs are often analyzed by first embedding (i.e., representing) them in some matrix/vector form with some number of dimensions. Various graph embedding methods have been developed to convert raw graph data into high dimensional vectors while preserving intrinsic graph properties [3]. The embedding methods focus on the node-level, edge-level [28], a hybrid, or at the graph level. This thesis focuses on graph-level embeddings which allows calculating similarity between two graphs. With the knowledge of embedding and sampling methods, the natural questions to ask are: 1) What is a good sampling size to ensure embeddings are similar enough to that of the original graph? 2) Do results depend on the sampling method? 3) Do they depend on the embedding method? 4) As we have embeddings, can we find some similarity between the original graph and sample? 5) How do we decide if the sample is good or not? How do we decide if the embedding is good or not? Essentially, if we have an embedding method and a sampling strategy, can we find the smallest sampling size that will give an ε-similar embedding to that of the original graph? We will try to answer the above questions in the thesis and give a new perspective on graph sampling. The experiments are conducted on graphs with thousands of edges and nodes. The datasets include graphs from social networks, autonomous systems, peer-to-peer networks, and collabo- ration networks. Two sampling methods are targeted namely - Random Node Sampling, and Random Edge Sampling. Euclidean distance is used as a similarity metric. Experiments are car- ried out on Graph2vec, and Spectral Features(SF) graph embedding methods. Univariate analysis is performed to decide a minimum sample which gives, e.g., 40% minimum sample for 80% similarity. We also design a Regression model which predicts similarity for a given sampling size and graph properties. Finally, we analyze the stability of the embedding methods, where we find that that e.g., Graph2Vec is a stable embedding method.},
note = {M.S. Thesis, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.},
keywords = {},
pubstate = {published},
tppubtype = {mastersthesis}
}
Machine Learning on graphs has become an active research area due to the prevailing graph-structured data in the real world. Many real-world applications can be modeled with graphs. Modern application domains include web-scale social networks [26], recommender systems, knowledge graphs, and biological or protein networks. However, there are various challenges. First, the graphs generated from such applications are often large. Moreover, in some scenarios, the complete graph is not available, e.g., for privacy reasons. Thus, it becomes impractical to perform network analysis or compute various graph measures. Hence, graph sampling becomes an important task.Sampling is often the first step to handle any type of massive data. The same applies to graphs as well, which leads to many graph sampling techniques. Sampling Techniques include Node-based (e.g., Random Node Sampling), Edge-based (e.g., Random Edge Sampling) and Traversal-based (e.g., Random Walk Sampling). Graphs are often analyzed by first embedding (i.e., representing) them in some matrix/vector form with some number of dimensions. Various graph embedding methods have been developed to convert raw graph data into high dimensional vectors while preserving intrinsic graph properties [3]. The embedding methods focus on the node-level, edge-level [28], a hybrid, or at the graph level. This thesis focuses on graph-level embeddings which allows calculating similarity between two graphs. With the knowledge of embedding and sampling methods, the natural questions to ask are: 1) What is a good sampling size to ensure embeddings are similar enough to that of the original graph? 2) Do results depend on the sampling method? 3) Do they depend on the embedding method? 4) As we have embeddings, can we find some similarity between the original graph and sample? 5) How do we decide if the sample is good or not? How do we decide if the embedding is good or not? Essentially, if we have an embedding method and a sampling strategy, can we find the smallest sampling size that will give an ε-similar embedding to that of the original graph? We will try to answer the above questions in the thesis and give a new perspective on graph sampling. The experiments are conducted on graphs with thousands of edges and nodes. The datasets include graphs from social networks, autonomous systems, peer-to-peer networks, and collabo- ration networks. Two sampling methods are targeted namely - Random Node Sampling, and Random Edge Sampling. Euclidean distance is used as a similarity metric. Experiments are car- ried out on Graph2vec, and Spectral Features(SF) graph embedding methods. Univariate analysis is performed to decide a minimum sample which gives, e.g., 40% minimum sample for 80% similarity. We also design a Regression model which predicts similarity for a given sampling size and graph properties. Finally, we analyze the stability of the embedding methods, where we find that that e.g., Graph2Vec is a stable embedding method.
Deshpande, Alpana
Sentiment Classification Bias in User Generated Content Masters Thesis
Syracuse University, Syracuse, NY, USA, 2021, (M.S. Thesis, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.).
@mastersthesis{deshpande2021sentimentclassificationbias,
title = {Sentiment Classification Bias in User Generated Content},
author = {Alpana Deshpande},
url = {https://surface.syr.edu/cgi/viewcontent.cgi?article=1479&context=thesis, PDF
https://surface.syr.edu/thesis/478/, SURFACE record},
year = {2021},
date = {2021-05-22},
address = {Syracuse, NY, USA},
school = {Syracuse University},
abstract = {Interactive websites generate terabytes of data on a daily basis. This data canbe used in multiple analytical applications to teach computers more about human behavior. Text classification is such an application. Multiple freely available user-generated text data can be used to teach computers to identify the sentiments behind a user’s on-screen interactions without the need of any human intervention. Sentiment analysis is an interesting problem, solving which would theoretically get a computer closer to passing the Turing test. Through this thesis, we test the ability of a classifier to accurately identify user sentiments. However, we do not focus on standard classification settings and the aim is to train the classifier in such a way that it would also be effective in identifying sentiment behind user generated text generated from a completely new social media platform. To be able to do this, we must first identify behavioral bias based on user interactions in two different social media sites as well as websites that accept user reviews. This bias must then be mitigated in order to obtain an unbiased classifier that can then be used to identify user sentiments on any social media platform. For the research in this thesis, such user-generated text is obtained from the social media sites Reddit and Twitter. We also obtain product review data related to both books and wine. Various natural language processing techniques are then employed to process the data and extract similar and dissimilar trends. Vectorized user text would be used to train sentiment classifiers. Finally, classification bias would be identified and mitigated in order to obtain classifiers that can identify human sentiments in real-time with an improved accuracy with limited dependency on source information.},
note = {M.S. Thesis, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.},
keywords = {},
pubstate = {published},
tppubtype = {mastersthesis}
}
Interactive websites generate terabytes of data on a daily basis. This data canbe used in multiple analytical applications to teach computers more about human behavior. Text classification is such an application. Multiple freely available user-generated text data can be used to teach computers to identify the sentiments behind a user’s on-screen interactions without the need of any human intervention. Sentiment analysis is an interesting problem, solving which would theoretically get a computer closer to passing the Turing test. Through this thesis, we test the ability of a classifier to accurately identify user sentiments. However, we do not focus on standard classification settings and the aim is to train the classifier in such a way that it would also be effective in identifying sentiment behind user generated text generated from a completely new social media platform. To be able to do this, we must first identify behavioral bias based on user interactions in two different social media sites as well as websites that accept user reviews. This bias must then be mitigated in order to obtain an unbiased classifier that can then be used to identify user sentiments on any social media platform. For the research in this thesis, such user-generated text is obtained from the social media sites Reddit and Twitter. We also obtain product review data related to both books and wine. Various natural language processing techniques are then employed to process the data and extract similar and dissimilar trends. Vectorized user text would be used to train sentiment classifiers. Finally, classification bias would be identified and mitigated in order to obtain classifiers that can identify human sentiments in real-time with an improved accuracy with limited dependency on source information.
Sambatur, Siri Chandana
Computational Analysis of Developmental Disorders in Children Masters Thesis
Syracuse University, Syracuse, NY, USA, 2018, (M.S. Thesis, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.).
@mastersthesis{sambatur2018developmentaldisorderschildren,
title = {Computational Analysis of Developmental Disorders in Children},
author = {Siri Chandana Sambatur},
url = {https://surface.syr.edu/cgi/viewcontent.cgi?article=1251&context=thesis, PDF
https://surface.syr.edu/thesis/250/, SURFACE record},
year = {2018},
date = {2018-06-01},
address = {Syracuse, NY, USA},
school = {Syracuse University},
abstract = {Early developmental disorders are common in children between the ages of 3 through 17. These developmental disorders begin at early ages and affect the day-to-day activities of children. These disorders also impact the growth and lifestyle of children. Most of the time these developmental disorders co-exist in children. The main focus of our research lies in Autism Spectrum Disorder, Attention-Deficit/Hyperactivity Disorder, Deletion syndrome (22q) and their co-occurrences. Most child psychologists and pediatricians diagnose these disorders in children through parent-based surveys. Our research uses three different parent-based reports: (1) Autism Diagnostic Interview (ADI), (2) Behavioral Assessment Schedule for Children (BASC), and (3) Vineland Adaptive Behavior Scales. These reports are questionnaires filled by parents under the inspection of certified professionals. These examinations require substantial amount of time and yield results after at least 13 months of wait time; hence, there is a pressing need to expedite the disorder detection process. Here, we address this challenge by utilizing machine learning techniques. We utilize Machine learning to parent-reviews to help understand the relevance and importance of parental assessments in diagnosing these disorders. Furthermore, we study the co-occurrence of these disorders and identify their indicators in parental-surveys using a variety of machine learning techniques. Our main objective is to determine whether one can accurately predict the occurrence of these disorders.},
note = {M.S. Thesis, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.},
keywords = {},
pubstate = {published},
tppubtype = {mastersthesis}
}
Early developmental disorders are common in children between the ages of 3 through 17. These developmental disorders begin at early ages and affect the day-to-day activities of children. These disorders also impact the growth and lifestyle of children. Most of the time these developmental disorders co-exist in children. The main focus of our research lies in Autism Spectrum Disorder, Attention-Deficit/Hyperactivity Disorder, Deletion syndrome (22q) and their co-occurrences. Most child psychologists and pediatricians diagnose these disorders in children through parent-based surveys. Our research uses three different parent-based reports: (1) Autism Diagnostic Interview (ADI), (2) Behavioral Assessment Schedule for Children (BASC), and (3) Vineland Adaptive Behavior Scales. These reports are questionnaires filled by parents under the inspection of certified professionals. These examinations require substantial amount of time and yield results after at least 13 months of wait time; hence, there is a pressing need to expedite the disorder detection process. Here, we address this challenge by utilizing machine learning techniques. We utilize Machine learning to parent-reviews to help understand the relevance and importance of parental assessments in diagnosing these disorders. Furthermore, we study the co-occurrence of these disorders and identify their indicators in parental-surveys using a variety of machine learning techniques. Our main objective is to determine whether one can accurately predict the occurrence of these disorders.
Sinha, Souradeep
Assessing Gender Inequality from Large-Scale Online Student Reviews Masters Thesis
Syracuse University, Syracuse, NY, USA, 2016, (M.S. Thesis, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.).
@mastersthesis{sinha2016assessinggenderinequality,
title = {Assessing Gender Inequality from Large-Scale Online Student Reviews},
author = {Souradeep Sinha},
url = {https://surface.syr.edu/cgi/viewcontent.cgi?article=1485&context=etd, PDF
https://surface.syr.edu/etd/485/, SURFACE record},
year = {2016},
date = {2016-05-01},
address = {Syracuse, NY, USA},
school = {Syracuse University},
abstract = {Career growth in academia is often dependent on student reviews of university profes- sors. A growing concern is how evaluation of teaching has been affected by gender biases throughout the reviewing process. However, pinpointing the exact causes and consequen- tial effects of this form of gender inequality has been a hard task. Current work focusses on university-wide student reviewing system, that depends on objective responses on a Likert scale to measure various aspects of an instructor’s qual- ity. Through our work, we access online student review data which are not limited by geographies, universities, or disciplines. Thereafter, we come up with a systematic approach to assess the various ways in which gender inequality is apparent from the student reviews. We also suggest a possible way in which bias related to the gender of a professor could be detected from both objective numerical measures and subjective opinions in reviews. Finally, we assess a logistic re- gression learning algorithm to find the most important factors that can help in identifying gender inequality.},
note = {M.S. Thesis, Department of Electrical Engineering and Computer Science, Syracuse University. Advisor: Reza Zafarani.},
keywords = {},
pubstate = {published},
tppubtype = {mastersthesis}
}
Career growth in academia is often dependent on student reviews of university profes- sors. A growing concern is how evaluation of teaching has been affected by gender biases throughout the reviewing process. However, pinpointing the exact causes and consequen- tial effects of this form of gender inequality has been a hard task. Current work focusses on university-wide student reviewing system, that depends on objective responses on a Likert scale to measure various aspects of an instructor’s qual- ity. Through our work, we access online student review data which are not limited by geographies, universities, or disciplines. Thereafter, we come up with a systematic approach to assess the various ways in which gender inequality is apparent from the student reviews. We also suggest a possible way in which bias related to the gender of a professor could be detected from both objective numerical measures and subjective opinions in reviews. Finally, we assess a logistic re- gression learning algorithm to find the most important factors that can help in identifying gender inequality.