Copy Detection system aims to identify if a query image is an edited/manipulated copy of an image from a large reference database with millions of images. While global image descriptors can retrieve visually similar images, they struggle to differentiate near-duplicates from semantically similar instances. We propose a dual-triplet metric learning (DTML) technique to learn global image features that group near-duplicates closer than visually similar images while maintaining the semantic structure of the embedding space. On the DISC2021 copy detection benchmark, DTML outperforms DINO and SSCD descriptors significantly. Additionally, we design an ensemble-based local feature matching strategy that operates on retrieved candidates to detect duplicates with high precision by effectively aggregating matched keypoints from individual methods. However, match scores from individual methods can be noisy and sensitive, making automated decisioning challenging. To address this, we develop a decisioning engine that consumes match scores from different keypoint detectors to make accurate decisions, demonstrating significant improvement over baselines across multiple datasets. We study the performance of our method under multiple datasets and demonstrate significant improvement over competitive baselines.

As e-commerce stores broaden their reach into new regions and introduce new products within established markets, the development of effective machine learning models becomes increasingly challenging due to the scarcity of labeled data. Traditional transfer learning methods typically require some labeled data from the target domain and often face computational bottlenecks. Despite the availability of a few transfer learning techniques, most are primarily developed for vision and text applications, making them unsuitable for other types of data. In many industries, however, tabular data is a predominant and crucial data type. Our work introduces XCapsUTL, a novel unsupervised transfer learning framework specifically designed for tabular data, aiming to fill this significant gap. Our approach leverages Capsule Neural Networks (CapsNet) to extract domain-agnostic knowledge. This knowledge is then refined using a constrained fine-tuning process, ensuring adaptability to the target task while preserving learned representations. XCapsUTL's unique feature encapsulation capabilities within CapsNet promote effective knowledge transfer without the need for designing effective feature-wise interaction approaches to capture higher-level semantics. Extensive experiments demonstrate the robustness and generalization capabilities of XCapsUTL across multiple domains and datasets, highlighting its practical significance and utility in addressing the unique challenges of tabular data in industry settings.

Multi-task problems frequently arise in machine learning when there are multiple target variables, which share a common synergy while being sufficiently different that optimizing on any of the task does not necessarily imply an optimum for the others. In this work, we develop PEMBOT, a novel Pareto-based multi-task classification framework using a gradient boosted tree architecture. The proposed methodology involves a) generating multiple instances of Pareto optimal trees, b) diverse subset selection using a determinantal point process (DPP) model, and c) ensembling of diverse Pareto optimal trees to yield the final output. We tested our framework on a problem from an e-commerce domain wherein the task is to predict at order placement time the different adverse scenarios in the order shipment journey such as the package getting lost or damaged during shipment. This model enables us to take preemptive measures to prevent these scenarios from happening resulting in significant operational cost savings. Further, to show the generality of our approach, we demonstrate the performance of our algorithm on a publicly available wine quality prediction dataset and compare against state-of-the-art baselines.