Md Omar Faruk Rokon

Shasvat Desai, Md Omar Faruk Rokon, Jhalak Nilesh Acharya et al.

Modern search systems rely on a fast first stage retriever to fetch relevant items from a massive catalog of items. Deployed search systems often use user engagement signals to supervise bi-encoder retriever training at scale, because these signals are continuously logged from real traffic and require no additional annotation effort. However, engagement is an imperfect proxy for semantic relevance. Items may receive interactions due to popularity, promotion, attractive visuals, titles, or price, despite weak query-item relevance. These limitations are further accentuated in Walmart's e-commerce sponsored search. User engagement on ad items is often structurally sparse because the frequency with which an ad is shown depends on factors beyond relevance such as whether the advertiser is currently running that ad, the outcome of the auction for available ad slots, bid competitiveness, and advertiser budget. Thus, even highly relevant query ad pairs can have limited engagement signals simply due to limited impressions. We propose a bi-encoder training framework for Walmart's sponsored search retrieval in e-commerce that uses semantic relevance as the primary supervision signal, with engagement used only as a preference signal among relevant items. Concretely, we construct a context-rich training target by combining 1. graded relevance labels from a cascade of cross-encoder teacher models, 2. a multichannel retrieval prior score derived from the rank positions and cross-channel agreement of retrieval systems running in production, and 3. user engagement applied only to semantically relevant items to refine preferences. Our approach outperforms the current production system in both offline evaluation and online AB tests, yielding consistent gains in average relevance and NDCG.

Md Omar Faruk Rokon, Shasvat Desai, Hong Yao et al.

How can we generate high-quality relevance annotations at scale without the cost and delays of human labeling? Relevance annotations are the backbone of search ranking systems which is needed for training data preparation, NDCG evaluation, and root cause analysis. However, human annotation is slow and off-the-shelf LLMs suffer from accuracy on domain-specific tasks. We propose a calibrated model cascade, a systematic approach for cost-efficient offline relevance annotation by routing queries through progressively larger fine-tuned classifiers. Our central insight is that accuracy and cost are orthogonal optimizations: domain-specific fine-tuning drives accuracy, cascading drives cost, and per-class isotonic calibration adds a small but reliable gain on top. Our contribution is threefold: (a) we decompose the gains and show that fine-tuning contributes 20 accuracy points while cascading is approximately accuracy-neutral but halves compute cost, (b) we introduce per-class isotonic calibration as one component of the cascade, contributing a small but statistically significant gain (+0.6 points over the strongest calibration baseline), and (c) we validate the system in production across six offline use cases, processing 150M+ annotations and enabling faster experimentation cycles. Our work is a building block for scalable, high-quality offline annotation pipelines in search and advertising systems.

Md Omar Faruk Rokon, Shasvat Desai, Jhalak Nilesh Acharya et al.

How can we generate high-quality training data for dense retrieval models at production scale, without relying on click signals or manual annotation? This question is critical for e-commerce sponsored search, where click-based training suffers from position bias and tail-query sparsity, and manual labeling at the scale of hundreds of millions of query-item pairs is economically infeasible. Our work is driven by the following insight: heterogeneous retrieval systems disagree on most items they retrieve, and this disagreement creates a natural source of structured training signal -- easy positives where all systems agree, hard positives that only lexical systems find, and hard negatives that fool exactly one system. As our key novelty, we combine three ideas into an end-to-end pipeline: (a) multi-channel retrieval mining with rank metadata from three production systems, (b) graded-relevance annotation by a calibrated three-model cascade ) that reaches 89.1% agreement with trained human annotators, and (c) three-stage progressive curriculum training that organizes 240M+ training examples across five difficulty levels. We deploy the trained two-tower BERT model on Walmart's sponsored search and evaluate it against 30K queries labeled by trained third-party human annotators. First, we show that the system achieves +5.1% NDCG@10 over the click-trained production baseline, with the largest gain on tail queries . Second, we show that embarrassing retrievals (rating 0) drop from 8.7% to 3.5%. Third, a two-week online A/B test with tens of millions of ad requests per arm confirms +2.80% ad spend, +1.4% CTR, +2.8% eCPM, and +2.9% click conversion rate. Overall, our work provides a practical and scalable blueprint for replacing click-based training with structured LLM-annotated supervision in production retrieval systems.

Md Omar Faruk Rokon, Jhalak Nilesh Acharya, Shasvat Desai et al.

How can we build a single relevance model that handles six different entity pair relationship types in e commerce from query product matching to product type similarity when each task has different data volumes, different semantic requirements, and potentially conflicting learning signals? This question is important because current industry practice relies on separate models for each task, preventing knowledge transfer and producing inconsistent relevance signals. Our work is driven by the following insight: encoder based and decoder only models encode task identity through different mechanisms, so the choice of task routing architecture how task identity is communicated to the shared model affects these two families in asymmetric ways. As our key novelty, we combine three ideas: (a) a unified multi task framework that jointly trains on six entity pair tasks under a shared three point relevance scale, (b) a systematic comparison of three task routing architectures (text prefix routing, multi head classification, and multihead with private transformer layers) across both LoRA adapted LLMs and fully finetuned cross encoders, and (c) a majority vote ensemble that exploits the diversity induced by private layer routing. First, we show that the MHP Ensemble (multi head with private layers) achieves 89.96% accuracy on 453K test examples the highest across all configurations . Second, we show that removing text prefixes without private layers causes severe degradation for decoder only LLMs while cross encoders remain robust , suggesting an encoder decoder asymmetry in task identity encoding. Third, we show that multi task training yields up to 14% improvement on low resource tasks over single task baselines.

Jianghong Zhou, Weizhi Du, Md Omar Faruk Rokon et al.

The rapid proliferation of e-commerce platforms accentuates the need for advanced search and retrieval systems to foster a superior user experience. Central to this endeavor is the precise extraction of product attributes from customer queries, enabling refined search, comparison, and other crucial e-commerce functionalities. Unlike traditional Named Entity Recognition (NER) tasks, e-commerce queries present a unique challenge owing to the intrinsic decorative relationship between product types and attributes. In this study, we propose a pioneering framework that integrates BERT for classification, a Conditional Random Fields (CRFs) layer for attribute value extraction, and Large Language Models (LLMs) for data annotation, significantly advancing attribute recognition from customer inquiries. Our approach capitalizes on the robust representation learning of BERT, synergized with the sequence decoding prowess of CRFs, to adeptly identify and extract attribute values. We introduce a novel decorative relation correction mechanism to further refine the extraction process based on the nuanced relationships between product types and attributes inherent in e-commerce data. Employing LLMs, we annotate additional data to expand the model's grasp and coverage of diverse attributes. Our methodology is rigorously validated on various datasets, including Walmart, BestBuy's e-commerce NER dataset, and the CoNLL dataset, demonstrating substantial improvements in attribute recognition performance. Particularly, the model showcased promising results during a two-month deployment in Walmart's Sponsor Product Search, underscoring its practical utility and effectiveness.

Md Omar Faruk Rokon, Pei Yan, Risul Islam et al.

How can we identify similar repositories and clusters among a large online archive, such as GitHub? Determiningrepository similarity is an essential building block in studying the dynamics and the evolution of such software ecosystems. The key challenge is to determine the right representation for the diverse repository features in a way that: (a) it captures all aspects of the available information, and (b) it is readily usable by MLalgorithms. We propose Repo2Vec, a comprehensive embedding approach to represent a repository as a distributed vector by combining features from three types of information sources. As our key novelty, we consider three types of information: (a)metadata, (b) the structure of the repository, and (c) the source code. We also introduce a series of embedding approaches to represent and combine these information types into a single embedding. We evaluate our method with two real datasets from GitHub for a combined 1013 repositories. First, we show that our method outperforms previous methods in terms of precision (93%vs 78%), with nearly twice as many Strongly Similar repositories and 30% fewer False Positives. Second, we show how Repo2Vecprovides a solid basis for: (a) distinguishing between malware and benign repositories, and (b) identifying a meaningful hierarchical clustering. For example, we achieve 98% precision and 96%recall in distinguishing malware and benign repositories. Overall, our work is a fundamental building block for enabling many repository analysis functions such as repository categorization by target platform or intention, detecting code-reuse and clones, and identifying lineage and evolution.

Risul Islam, Md Omar Faruk Rokon, Evangelos E. Papalexakis et al.

How can we expand the tensor decomposition to reveal a hierarchical structure of the multi-modal data in a self-adaptive way? Current tensor decomposition provides only a single layer of clusters. We argue that with the abundance of multimodal data and time-evolving networks nowadays, the ability to identify emerging hierarchies is important. To this effect, we propose RecTen, a recursive hierarchical soft clustering approach based on tensor decomposition. Our approach enables us to: (a) recursively decompose clusters identified in the previous step, and (b) identify the right conditions for terminating this process. In the absence of proper ground truth, we evaluate our approach with synthetic data and test its sensitivity to different parameters. We also apply RecTen on five real datasets which involve the activities of users in online discussion platforms, such as security forums. This analysis helps us reveal clusters of users with interesting behaviors, including but not limited to early detection of some real events like ransomware outbreaks, the emergence of a blackmarket of decryption tools, and romance scamming. To maximize the usefulness of our approach, we develop a tool which can help the data analysts and overall research community by identifying hierarchical structures. RecTen is an unsupervised approach which can be used to take the pulse of the large multi-modal data and let the data discover its own hidden structures by itself.

Risul Islam, Md Omar Faruk Rokon, Evangelos E. Papalexakis et al.

How can we get a security forum to "tell" us its activities and events of interest? We take a unique angle: we want to identify these activities without any a priori knowledge, which is a key difference compared to most of the previous problem formulations. Despite some recent efforts, mining security forums to extract useful information has received relatively little attention, while most of them are usually searching for specific information. We propose TenFor, an unsupervised tensor-based approach, to systematically identify important events in a three-dimensional space: (a) user, (b) thread, and (c) time. Our method consists of three high-level steps: (a) a tensor-based clustering across the three dimensions, (b) an extensive cluster profiling that uses both content and behavioral features, and (c) a deeper investigation, where we identify key users and threads within the events of interest. In addition, we implement our approach as a powerful and easy-to-use platform for practitioners. In our evaluation, we find that 83% of our clusters capture meaningful events and we find more meaningful clusters compared to previous approaches. Our approach and our platform constitute an important step towards detecting activities of interest from a forum in an unsupervised learning fashion in practice.

Risul Islam, Md Omar Faruk Rokon, Ahmad Darki et al.

Authors of malicious software are not hiding as much as one would assume: they have a visible online footprint. Apart from online forums, this footprint appears in software development platforms, where authors create publicly-accessible malware repositories to share and collaborate. With the exception of a few recent efforts, the existence and the dynamics of this community has received surprisingly limited attention. The goal of our work is to analyze this ecosystem of hackers in order to: (a) understand their collaborative patterns, and (b) identify and profile its most influential authors. We develop HackerScope, a systematic approach for analyzing the dynamics of this hacker ecosystem. Leveraging our targeted data collection, we conduct an extensive study of 7389 authors of malware repositories on GitHub, which we combine with their activity on four security forums. From a modeling point of view, we study the ecosystem using three network representations: (a) the author-author network, (b) the author-repository network, and (c) cross-platform egonets. Our analysis leads to the following key observations: (a) the ecosystem is growing at an accelerating rate as the number of new malware authors per year triples every 2 years, (b) it is highly collaborative, more so than the rest of GitHub authors, and (c) it includes influential and professional hackers. We find 30 authors maintain an online "brand" across GitHub and our security forums. Our study is a significant step towards using public online information for understanding the malicious hacker community.

Md Omar Faruk Rokon, Risul Islam, Ahmad Darki et al.

Where can we find malware source code? This question is motivated by a real need: there is a dearth of malware source code, which impedes various types of security research. Our work is driven by the following insight: public archives, like GitHub, have a surprising number of malware repositories. Capitalizing on this opportunity, we propose, SourceFinder, a supervised-learning approach to identify repositories of malware source code efficiently. We evaluate and apply our approach using 97K repositories from GitHub. First, we show that our approach identifies malware repositories with 89% precision and 86% recall using a labeled dataset. Second, we use SourceFinder to identify 7504 malware source code repositories, which arguably constitutes the largest malware source code database. Finally, we study the fundamental properties and trends of the malware repositories and their authors. The number of such repositories appears to be growing by an order of magnitude every 4 years, and 18 malware authors seem to be "professionals" with well-established online reputation. We argue that our approach and our large repository of malware source code can be a catalyst for research studies, which are currently not possible.