http://repository.iiitd.edu.in/xmlui/handle/123456789/1794 Year-2026 2026-09-01T23:21:40Z http://repository.iiitd.edu.in/xmlui/handle/123456789/2048 On recovering fair and accurate classifiers with imperfect distributions Sharma, Mohit; Shah, Rajiv Ratn (Advisor); Deshpande, Amit (Advisor) In the context of classification, the Bayes optimal classifier represents the best possible classification rule with respect to the 0-1 loss, for a given data distribution. In the setting of binary classification, it can often be expressed as a thresholding rule over the instance-dependent class posterior probability. Prior research works have studied the best classification performance and the corresponding Bayes optimal classification rule under fairness constraints, which can again be expressed as a group/instance- dependent thresholding rule. However, it is often observed that fairness and accuracy of a model are often at odds and exhibit a tradeoff. This tradeoff depends on the distribution and the fairness metric of interest, and understanding its cause and effects is crucial for real-world deployments. In this thesis, we study this tradeoff from several angles. One of the most important tools for all of our studies involves exploring the Bayes optimal fair classifiers to study the effects of data biases and derive theoretical guarantees. We first re-examine the critical stance on the fairness-accuracy tradeoff and instead study under what conditions fairness constraints can eventually help recover the uncon- strained Bayes optimal classifier, especially when we can characterize the factors of data bias in our given distribution. Next, we highlight how the Bayes optimal fair classification rule can be implemented at various stages of classification: Either as a pre-processing of the given distribution, a weighted risk minimization, or as a post-processing (thresholding) of the class posterior probability. We then show, via simulations on widely used fair classifiers, that with varying amounts of data bias, the theoretical equivalence does not translate into practice and, in fact, sometimes even fails to mitigate unfairness. Re-examining the tradeoff phenomena, we then ask whether it is possible to steer distributions most minimally and efficiently towards an ideal distribution, where the Bayes optimal classifiers are always fair by default. We mathematically characterize what an ideal distribution may look like when the distribution can be expressed with some parametric family (e.g., Gaussian distributions). Finally, we estimate the fairness-accuracy tradeoff using imperfect data access, i.e., in a model-free setting, only using class posterior probabilities and without access to features. In the spirit of prior works that use Bayes error estimation techniques to benchmark classification performance, we study how we can estimate and characterize the fairness-accuracy tradeoff for any distribution only using soft labels (class posterior probabilities). This allows us to compare various fairness-accuracy tradeoff estimation techniques in the literature against the theoretically best and serves as an effective bench- marking framework for future work in this space. 2026-08-01T00:00:00Z http://repository.iiitd.edu.in/xmlui/handle/123456789/2009 Lightweight optimization strategies for modern video applications over wireless networks Chaudhary, Shubham; Bhattacharya, Arani (Advisor) The various video applications running over wireless networks face a common set of challenges, such as inconsistent bandwidth, high network variability, and sudden latency spikes. Considering such constraints, this thesis explores the possible strategies and design choices for developing data-intensive video applications such as real-time traffic surveillance, live video streaming, and cloud-assisted autonomous driving. Our focus is primarily on two techniques across three distinct video-streaming applications. The first technique is the intelligent use of tiled encoding available in modern video codecs, where the encoded video has independent rectangular spatial regions that can be manipulated in real time without re-encoding. The second technique is to develop data filtering strategies to minimize ingestion costs by pruning extraneous information. In traffic surveillance, cameras stream video to servers for computer vision algorithms, consuming significant bandwidth. To reduce bandwidth usage, we use tile sampling to select spatial regions of frames that contain only moving objects, since the rest of the frame mostly has static backgrounds, such as the sky and buildings. To select such tiles, we propose an adaptive tile selection algorithm that samples only tiles with moving objects by leveraging their correlation with tile bitrates. Our evaluations across different lighting, weather, and traffic conditions, both using benchmark videos and a live deployment, show improved accuracy, reduced bandwidth usage, and lower overhead than existing systems. For live streaming, we propose using the network interface of a nearby helper device to get an aggregated bandwidth. To stream videos, we design a tile-level aggregation strategy that partitions tiles into two subsets and schedules them independently across available network interfaces, based on their importance. We demonstrate, through extensive experiments, that such an aggregation strategy provides better QoE than conventional multipath strategies. Lastly, we address the high data ingestion cost in self-driving cars. Autonomous vehicles rely on compressed models whose accuracy degrades over time as the distribution of real-world test data changes, requiring frequent retraining on a server. This necessitates limiting the number of selected training frames to minimize the end-to-end delay in obtaining the retrained model without impairing post-training accuracy. We sample only the most useful frames and adaptively encode them into a video based on the network bandwidth. Our lightweight sampling strategy seamlessly integrates with the existing workflow, achieving superior accuracy while minimizing update delay compared to baseline strategies across different network conditions 2026-07-01T00:00:00Z http://repository.iiitd.edu.in/xmlui/handle/123456789/2008 The shape of learning: topological and geometric perspectives on neural representations Suresh, Suryaka; Abrol, Vinayak (Advisor) This thesis presents a unified framework for analysing the “expressivity” of Deep Neural Networks by combining factor analysis with topological data analysis (TDA). The central idea is to study the learning in Neural Networks as a geometric and topological object that encodes the representational capacity of the model. Towards the same, the geometric structure of Archetypal Analysis is exploited to study the structural manifold underlying the neural embeddings. The archetypal subspace provides a scalable foundation for applying tools from computational topology, enabling the characterisation of topological transitions that reflect changes in representational structure and network complexity. Building upon this methodological foundation, the framework is applied to the model selection problem in the context of transfer learning, where the objective is to identify an optimal model from a collection of pre-trained architectures. By quantifying the topological and geometric signatures of the embedding space, the approach provides an interpretable measure of expressivity that correlates with the generalization behavior of the Neural Network. Furthermore, the study extends the analysis of expressivity to generative learning, through the topological inference of their generative manifolds forging a relationship between manifold geometry, data diversity, and generative performance. Overall, this research contributes a topology-guided and geometrically interpretable framework for understanding, comparing, and selecting neural architectures. It advances the theoretical and empirical interpretations of deep representation learning and establishes a scalable approach to bridging topological structure, expressivity, and generalisation in neural architecture. 2026-06-01T00:00:00Z http://repository.iiitd.edu.in/xmlui/handle/123456789/1998 A data-centric lens for representation learning in graphs via semantic and structural enhancement Goyal, Karan; Goyal, Vikram (Advisor); Mohania, Mukesh (Advisor) While progress in graph representation learning has often been model-centric: focusing on the design of more powerful architectures, this thesis presents a data-centric lens: a framework viewing the graph not as a static input, but as an object to be actively and intelligently enhanced. We argue and then show that superior performance can be unlocked through a deliberate progression, beginning with specialised node-level semantic enhancement and culminating in generic graph-level structural enhancement. Our investigation begins with the specialised task of citation recommendation in scholarly networks, where we first explore semantic enhancement. Our formative framework, SymTax, enriches node data by modelling human exploratory behaviour via symbiotic relationship and by fusing taxonomies into hyperbolic space. Building upon this, we address critical issues of scalability and real-world applicability by introducing Profiler, a highly efficient module that creates a rich Public Profile for each paper by statically aggregating signals from its inward citations, which is then carefully integrated with our designed DAVINCI reranker. This advancement is validated under our proposed Inductive evaluation protocol, establishing a more rigorous, robust and realistic standard for the field. Having established the power of semantic enrichment, we then widen the horizon of our data-centric lens to structural enhancement for the generalised challenge of learning on heterophilic graphs. We propose PathLens, a model-agnostic technique that fundamentally augments the graph’s topology. By strategically adding supernodes derived from spectral clustering, PathLens creates new information pathways, effectively reducing the average shortest path length between similar nodes thereby increasing the graph’s homophily. This directly addresses the core impediment of heterophily, significantly boosting the performance of a wide range of standard GNNs as well as heterophilic GNNs by providing them with a more learnable graph structure. Collectively, this body of work validates the efficacy of the data-centric lens. By first enriching the semantic content of nodes and then augmenting the structural fabric of the graph, we demonstrate a powerful and generalisable paradigm for building high- performance systems for learning on graphs. This thesis contributes not only a suite of novel techniques but also a guiding philosophy that prioritises the enhancement of the data itself as a primary driver of success for representation learning in graphs. 2026-07-17T00:00:00Z