Graph Mining / Complex Networks
Possioning network embeddings
Study existing network embeddings against a supervised poisoning attack strategy
Given a network, how easy is it to change a few edges to fool a classifier based on node embeddings? We explore a supervised poisoning attack strategy to observe this particular decrease in the quality of embeddings across various embedding methods, including modern GCNs. By altering specific edges within the network, we aim to understand the vulnerability of different embedding techniques. Our study focuses on how these changes can significantly impact the performance of classifiers using these embeddings.
Community Deception
Hiding a target community from community detection algorithms
We endeavor to mislead community detection algorithms by subtly altering the connections within a specified network, with the primary goal of obscuring the identity of a specific group of nodes. We describe this issue as Community Deception, alternatively known as Hide and Seek Community (HSC). This strategy is crucial in the realm of privacy and security. For example, it can shield users of social networks from intrusive surveillance practices, thereby protecting their personal privacy. Additionally, the HSC approach serves as a vital instrument in securing sensitive data against discovery by entities without proper authorization.
Citation network analysis
Detecting academic inbreeding in institutions and proving its ill effects on the overall quality of research in the institution
Academic inbreeding is the practice of academia of an institution hiring its own graduates as professors which is considered as an unhealthy and insular practice which might lead to overall degrade in the quality of research in the institution. Through our novel dataset, we will be aiming to provide proof to various hypotheses stating how excessive inbreeding can be bad for an institution. We will be dividing our data into 4 inbreed types namely 'Pure Inbreed', 'Mobile Inbreed', 'Silver Corded', and 'Non Inbreed'. We will be using the Semantic Scholar open corpus dataset to further improve our data and provide deeper analysis about the negative effects of inbreeding.
Node Influence in Network Growth Models
We study how influential nodes or "leaders" are able to gather and maintain their prominence, by attracting new connections in a dynamic network under different network growth models -- Barabasi-Albert, fitness-based, and spatial models
Several network growth models have been proposed in literature that capture growth of evolving networks. An interesting characteristic that can be studied for growing networks is how influential nodes, or “leaders” are able to gather and maintain their prominence over time (by attracting new connections). Looking into the behavior of influential nodes is important since they are potential spreaders of information that play an essential role in real-world applications like viral marketing, epidemic control, and protection from spam attacks.
To study them, we introduce a new metric called “visibility of a node” that captures the probability for a node to form new connections in an evolving network. Different classes of network growth models were studied to see how influential nodes behave under their respective settings. We analyse the BA or preferential attachment model, fitness-based models (additive, multiplicative, and nonlinear), and spatial models. For each type of network growth model, the project covers thorough theoretical proofs on how the visibility of influential nodes changes over time and supports these claims with the help of comprehensive experimental analysis.