Research

My research mainly focuses on Hyperbolic Graph Machine Learning and Privacy Preserving Distributed Graph learning.

Hyperbolic Graph Machine Learning

Euclidean space, while effective for graph-related learning tasks, has inherent limitations in encoding complex patterns due to its polynomially expanding capacity. Despite efforts with nonlinear techniques, complex graph patterns may still require computationally intractable embedding dimensionality. Recent research indicates that many complex datasets exhibit non-Euclidean underlying structures. For instance, tree-like structures are prevalent in various real-world networks, such as hypernym structures in natural languages, subordinate structures in knowledge graphs, organizational structures in financial fraud, and power-law distributions in recommender systems. In such cases, Euclidean space fails to provide optimal geometric representations. Hyperbolic space has emerged as a promising alternative for processing tree-like graph data. Hyperbolic space possesses a unique geometric property where its volume increases exponentially with its radius, unlike the polynomial growth of Euclidean space. This trait offers two key advantages for handling tree-like graph data. Firstly, hyperbolic space minimizes distortion and aligns well with hierarchies, closely matching the growth rate of tree-like data, which Euclidean space cannot achieve. Secondly, even in low-embedding dimension spaces, hyperbolic models can generate high-quality representations, making them particularly advantageous in low-memory and low-storage scenarios.

Federated Learning

The growing complexity of machine learning models, as well as the amount of data required to train them, has led to the development of distributed learning. The Federated learning (FL) framework is becoming a powerful approach for creating shared machine learning models from a bunch of participants. Potential applications of FL are endless including tasks such as learning the activities of mobile phone users, ensuring the personalized quality of services, Internet of Things (IoT), communication and networking, robotics, video streaming, transportation and cab services, autonomous vehicles, health care, legal technology, earth sciences, etc.

FL enables mobile phones to collaboratively learn a shared prediction model while keeping all the training data on device, decoupling the ability to do machine learning from the need to store the data in the cloud.

Graph structured data: social network, where nodes are individuals and edges are associations and friendships.

Why do we need Graphs?

In research and engineering, graph models serve as an efficient means of representing data. Examples include social graphs, biological graphs, financial graphs, transportation graphs, sensor graphs, and neural networks. Datasets are often simply a collection of unstructured samples in various applications, and the underlying graph information (showing the connections between the samples) is often not readily available. Hence, the first step in obtaining pertinent information from any given dataset is to obtain a graph-based representation of the dataset. It is required to deduce the graph’s structure from the data itself when a natural choice of the graph is not immediately available. Graph-based machine learning is emerging as a powerful tool for handling diverse and complex problems with highly irregular data in domains such as drug discovery, genetics, cancer research, network medicine, neuroscience, finance, and material science.

Highlights.

> Presented our work SENSE: SENsing Similarity SEeing Structure at Affinity Workshop-WiML @ NeurIPS 2025 Oral Presentation.

Attended and presented at NeurIPS 2025; it was a great experience. You can check it here on this link. Also sharing some pictures from the conference this year!

>Presented our work Learning to Collaborate: Personlized Federated GNNs, at LOG (Learning on Graphs) Conference 2025.

Attended and presented our poster at LOG 2025. You can check out the paper here at this link. Also sharing some pictures from the conference this year!

> "𝗛𝘆𝗽𝗲𝗿𝗗𝗲𝗳𝗲𝗻𝗱𝗲𝗿: 𝗔 𝗥𝗼𝗯𝘂𝘀𝘁 𝗙𝗿𝗮𝗺𝗲𝘄𝗼𝗿𝗸 𝗳𝗼𝗿 𝗛𝘆𝗽𝗲𝗿𝗯𝗼𝗹𝗶𝗰 𝗚𝗡𝗡𝘀" has been accepted at 𝗔𝗔𝗔𝗜 𝟮𝟬𝟮𝟱.
Hyperbolic spaces have emerged as a powerful tool for embedding hierarchical data due to their exponential growth properties. While several Hy-GNN architectures have been developed to leverage hyperbolic geometry for high-fidelity representation of hierarchical datasets, their robustness to adversarial perturbations remains an open challenge.

The Problem. 𝘌𝘹𝘪𝘴𝘵𝘪𝘯𝘨 𝘏𝘺-𝘎𝘕𝘕 𝘮𝘰𝘥𝘦𝘭𝘴, 𝘥𝘦𝘴𝘱𝘪𝘵𝘦 𝘵𝘩𝘦𝘪𝘳 𝘱𝘳𝘰𝘮𝘪𝘴𝘪𝘯𝘨 𝘱𝘦𝘳𝘧𝘰𝘳𝘮𝘢𝘯𝘤𝘦, 𝘢𝘳𝘦 𝘴𝘶𝘴𝘤𝘦𝘱𝘵𝘪𝘣𝘭𝘦 𝘵𝘰 𝘢𝘥𝘷𝘦𝘳𝘴𝘢𝘳𝘪𝘢𝘭 𝘢𝘵𝘵𝘢𝘤𝘬𝘴 𝘢𝘯𝘥 𝘯𝘰𝘪𝘴𝘦. 𝘛𝘩𝘪𝘴 𝘷𝘶𝘭𝘯𝘦𝘳𝘢𝘣𝘪𝘭𝘪𝘵𝘺 𝘤𝘰𝘮𝘱𝘳𝘰𝘮𝘪𝘴𝘦𝘴 𝘵𝘩𝘦 𝘩𝘪𝘦𝘳𝘢𝘳𝘤𝘩𝘪𝘤𝘢𝘭 𝘴𝘵𝘳𝘶𝘤𝘵𝘶𝘳𝘦 𝘰𝘧 𝘥𝘢𝘵𝘢 𝘪𝘯 𝘤𝘳𝘪𝘵𝘪𝘤𝘢𝘭 𝘢𝘱𝘱𝘭𝘪𝘤𝘢𝘵𝘪𝘰𝘯𝘴, 𝘮𝘰𝘵𝘪𝘷𝘢𝘵𝘪𝘯𝘨 𝘰𝘶𝘳 𝘸𝘰𝘳𝘬 𝘵𝘰 𝘣𝘳𝘪𝘥𝘨𝘦 𝘵𝘩𝘪𝘴 𝘨𝘢𝘱.

Solution. We introduce HyperDefender, a first-of-its-kind framework that 1️) strengthens Hy-GNNs against adversarial attacks and noise. 2️) Restores hierarchical integrity in compromised datasets, ensuring reliable predictions. Our research also examines the limitations of traditional GNNs and Hy-GNNs under adversarial conditions, revealing the inadequacy of existing defense mechanisms for hierarchical data.

Discover more about our work here and our lab MISN here.

> "PPDA: A Privacy-Preserving Framework for Distributed Graph Learning" has been accepted at the ICONIP 2024 Conference.

PPDA is a novel framework for distributed graph learning that respects privacy while enabling effective graph-based machine learning.

Graph-based learning has immense applications in domains like social networks, robotics, communication, and medicine. However, the sensitive nature of many datasets often limits their usability due to privacy concerns. Existing methods attempt to preserve privacy by preprocessing data to extract user-side features, but these approaches are vulnerable to adversarial attacks that can infer private attributes. Our work introduces PPDA, a novel framework for distributed graph learning that respects privacy while enabling effective graph-based machine learning. By preserving structural properties without requiring access to raw features, our approach supports feature learning and distance computation directly on the server side. Importantly, it’s highly adaptable and can be integrated with various distance approximation methods and graph learning techniques. Through extensive experiments on synthetic and real-world datasets, we demonstrate that PPDA achieves performance comparable to traditional methods while maintaining strict privacy guarantees—making it the first privacy-preserving distributed graph learning framework of its kind.

The ICONIP 2024 proceedings will be published in Springer’s Lecture Notes in Computer Science (LNCS) and Communications in Computer and Information Science (CCIS) series.

The conference was an incredible experience, filled with insightful talks, tutorials, workshops, and cultural events celebrating the beauty of New Zealand. Sharing some pictures below from my presentation and the event!

Learn more about PPDA here.

Gallery

Successfully conducted a Workshop on Machine Learning at IIT Delhi for the students of DSEU Data Analytics.

Conducted workshop on Machine Learning with Python at Hindu College (University of Delhi)

I work in MISN Lab at IIT Delhi

Machine Learning Workshop at Delhi University.

Teaching Students of Delhi University..

Pictures from ICONIP 2024 Conference Gathering and Paper Presentation at Auckland University of Technology, New Zealand

Publications:

N. Malik and S. Majumdar, "Multiview Human Gait Analysis Using the First and Third Person Data," 2021 Emerging Trends in Industry 4.0 (ETI 4.0), 2021, pp. 1-5, doi: https://ieeexplore.ieee.org/document/9619340
Nikita Malik and Sudipta Majumdar, "Multiview running and walking gait analysis using the first and third person data", 2021 J. Phys.: Conf. Ser. 2070 012138, doi: https://iopscience.iop.org/article/10.1088/1742-6596/2070/1/012138

Next project

Page updated

Google Sites

Report abuse