arXiv subjects

Sandipan Sikdar

Publications and source records attributed to Sandipan Sikdar.

At least 19 recordsLinked to original sources

arXiv2026

When Safety Speaks a Language: A Mechanistic Analysis of Safety-Language Identity Entanglement in LLMs

Safety alignment of large language models (LLMs) degrades across languages, yet the internal mechanism driving this asymmetry remains poorly understood. Our work, therefore, presents a systematic mechanistic analysis of multilingual safety using sparse autoencoder (SAE) features, sparse interpretable directions in the residual stream associated with harmful and harmless model behavior across three instruction-tuned LLMs, eight languages, and all model layers. We observe that safety-relevant features are architecture-dependent in terms of where they are located and how they are distributed across layers. Additionally, they are geometrically entangled with language identity and exhibit cross-lingual sharing patterns, i.e., languages share safety features to varying degrees across model depths and architectures. This safety-language entanglement has direct consequences such that ablating safety features impacts not only harmful response rates but also target language, with the degree of intervention predicted by the relationship between safety and language features. Our findings qualify the language-universality of safety alignment as architecture-dependent and offer a mechanistic account of multilingual safety interventions.

cs.CL↗

arXiv2026

REMEDI: A Benchmark for Retention and Unlearning Evaluation in Multi-label Clinical Disease Inference

Language models trained for clinical disease inference are trained on patient data, which may include sensitive and private information, and data owners may request the removal of their data from a trained model due to privacy or copyright concerns. However, exactly unlearning patient-specific data is intractable, and retraining with minor data removal is resource-intensive. While there exists several machine unlearning methods that can be used, their utility is generally restricted to non-medical domains. Moreover, the existing benchmarks for evaluating such unlearning methods primarily utilize synthetically curated datasets, which are not truly representative of real-world systems. Hence, the effectiveness of these unlearning methods in the medical domain is largely unclear. To this end, we introduce REMEDI, an extensive benchmark for machine unlearning tailored to multi-label and multiclass clinical disease inference, where label correlations, longitudinal structure, and safety constraints make unlearning particularly challenging. Unlike the existing benchmarks, REMEDI considers: (1) a relevant application domain (medical), (2) comprehensive unlearning setups involving diverse sets of forget instances, (3) challenging unlearning scenarios including multi-label and multi-class classification tasks, and (4) evaluation metrics involving performance both in terms of utility and extent of unlearning achieved. REMEDI is developed using the MIMIC-III clinical database that contains comprehensive clinical data of patients. Experiments with existing unlearning methods indicate that there exists a trade-off between utility and unlearning performance. They are also largely unsuited to multi-label classification tasks. To facilitate reproducibility, we make our benchmark publicly available.

cs.LG↗

arXiv2026

Towards Sensitivity-Aware Language Models

With LLMs increasingly deployed in corporate data management, it is crucial to ensure that these models do not leak sensitive information. In the context of corporate data management, the concept of sensitivity awareness has been introduced, enabling LLMs to adhere to predefined access rights rules. However, it remains unclear how sensitivity awareness relates to established notions of privacy, such as differential privacy (DP), thereby making it difficult to deploy meaningfully in real-world applications. In this work, we formalize the notion of sensitivity awareness and theoretically establish its connection to DP. Additionally, we develop a supervised fine-tuning recipe to make existing, four-bit quantized LLMs more sensitivity-aware. With a performance boost of up to 21.7%, the finetuned LLMs not only substantially improve over their baseline but also outperform other full-precision open-source and commercial models of similar size in achieving sensitivity awareness, demonstrating the effectiveness of our proposed approach. At the same time, our method also largely preserves the models' performance on other tasks, such as general instruction-following, mathematical, and common-sense reasoning.

cs.CR↗

arXiv2025

ACCESS DENIED INC: The First Benchmark Environment for Sensitivity Awareness

Large language models (LLMs) are increasingly becoming valuable to corporate data management due to their ability to process text from various document formats and facilitate user interactions through natural language queries. However, LLMs must consider the sensitivity of information when communicating with employees, especially given access restrictions. Simple filtering based on user clearance levels can pose both performance and privacy challenges. To address this, we propose the concept of sensitivity awareness (SA), which enables LLMs to adhere to predefined access rights rules. In addition, we developed a benchmarking environment called ACCESS DENIED INC to evaluate SA. Our experimental findings reveal significant variations in model behavior, particularly in managing unauthorized data requests while effectively addressing legitimate queries. This work establishes a foundation for benchmarking sensitivity-aware language models and provides insights to enhance privacy-centric AI systems in corporate environments.

cs.CL↗

arXiv2025

Exploring Disparity-Accuracy Trade-offs in Face Recognition Systems: The Role of Datasets, Architectures, and Loss Functions

Automated Face Recognition Systems (FRSs), developed using deep learning models, are deployed worldwide for identity verification and facial attribute analysis. The performance of these models is determined by a complex interdependence among the model architecture, optimization/loss function and datasets. Although FRSs have surpassed human-level accuracy, they continue to be disparate against certain demographics. Due to the ubiquity of applications, it is extremely important to understand the impact of the three components -- model architecture, loss function and face image dataset on the accuracy-disparity trade-off to design better, unbiased platforms. In this work, we perform an in-depth analysis of three FRSs for the task of gender prediction, with various architectural modifications resulting in ten deep-learning models coupled with four loss functions and benchmark them on seven face datasets across 266 evaluation configurations. Our results show that all three components have an individual as well as a combined impact on both accuracy and disparity. We identify that datasets have an inherent property that causes them to perform similarly across models, independent of the choice of loss functions. Moreover, the choice of dataset determines the model's perceived bias -- the same model reports bias in opposite directions for three gender-balanced datasets of ``in-the-wild'' face images of popular individuals. Studying the facial embeddings shows that the models are unable to generalize a uniform definition of what constitutes a ``female face'' as opposed to a ``male face'', due to dataset diversity. We provide recommendations to model developers on using our study as a blueprint for model development and subsequent deployment.

cs.CV↗

arXiv2024

Recommendation Fairness in Social Networks Over Time

In social recommender systems, it is crucial that the recommendation models provide equitable visibility for different demographic groups, such as gender or race. Most existing research has addressed this problem by only studying individual static snapshots of networks that typically change over time. To address this gap, we study the evolution of recommendation fairness over time and its relation to dynamic network properties. We examine three real-world dynamic networks by evaluating the fairness of six recommendation algorithms and analyzing the association between fairness and network properties over time. We further study how interventions on network properties influence fairness by examining counterfactual scenarios with alternative evolution outcomes and differing network properties. Our results on empirical datasets suggest that recommendation fairness improves over time, regardless of the recommendation method. We also find that two network properties, minority ratio, and homophily ratio, exhibit stable correlations with fairness over time. Our counterfactual study further suggests that an extreme homophily ratio potentially contributes to unfair recommendations even with a balanced minority ratio. Our work provides insights into the evolution of fairness within dynamic networks in social science. We believe that our findings will help system operators and policymakers to better comprehend the implications of temporal changes and interventions targeting fairness in social networks.

cs.SI↗

arXiv2024

IVP-VAE: Modeling EHR Time Series with Initial Value Problem Solvers

Continuous-time models such as Neural ODEs and Neural Flows have shown promising results in analyzing irregularly sampled time series frequently encountered in electronic health records. Based on these models, time series are typically processed with a hybrid of an initial value problem (IVP) solver and a recurrent neural network within the variational autoencoder architecture. Sequentially solving IVPs makes such models computationally less efficient. In this paper, we propose to model time series purely with continuous processes whose state evolution can be approximated directly by IVPs. This eliminates the need for recurrent computation and enables multiple states to evolve in parallel. We further fuse the encoder and decoder with one IVP solver utilizing its invertibility, which leads to fewer parameters and faster convergence. Experiments on three real-world datasets show that the proposed method can systematically outperform its predecessors, achieve state-of-the-art results, and have significant advantages in terms of data efficiency.

cs.LG↗

arXiv2023

A Review of the Role of Causality in Developing Trustworthy AI Systems

State-of-the-art AI models largely lack an understanding of the cause-effect relationship that governs human understanding of the real world. Consequently, these models do not generalize to unseen data, often produce unfair results, and are difficult to interpret. This has led to efforts to improve the trustworthiness aspects of AI models. Recently, causal modeling and inference methods have emerged as powerful tools. This review aims to provide the reader with an overview of causal methods that have been developed to improve the trustworthiness of AI models. We hope that our contribution will motivate future research on causality-based solutions for trustworthy AI.

cs.AI↗

arXiv2023

SensePOLAR: Word sense aware interpretability for pre-trained contextual word embeddings

Adding interpretability to word embeddings represents an area of active research in text representation. Recent work has explored thepotential of embedding words via so-called polar dimensions (e.g. good vs. bad, correct vs. wrong). Examples of such recent approaches include SemAxis, POLAR, FrameAxis, and BiImp. Although these approaches provide interpretable dimensions for words, they have not been designed to deal with polysemy, i.e. they can not easily distinguish between different senses of words. To address this limitation, we present SensePOLAR, an extension of the original POLAR framework that enables word-sense aware interpretability for pre-trained contextual word embeddings. The resulting interpretable word embeddings achieve a level of performance that is comparable to original contextual word embeddings across a variety of natural language processing tasks including the GLUE and SQuAD benchmarks. Our work removes a fundamental limitation of existing approaches by offering users sense aware interpretations for contextual word embeddings.

cs.CL↗

arXiv2022

Properties of Group Fairness Metrics for Rankings

In recent years, several metrics have been developed for evaluating group fairness of rankings. Given that these metrics were developed with different application contexts and ranking algorithms in mind, it is not straightforward which metric to choose for a given scenario. In this paper, we perform a comprehensive comparative analysis of existing group fairness metrics developed in the context of fair ranking. By virtue of their diverse application contexts, we argue that such a comparative analysis is not straightforward. Hence, we take an axiomatic approach whereby we design a set of thirteen properties for group fairness metrics that consider different ranking settings. A metric can then be selected depending on whether it satisfies all or a subset of these properties. We apply these properties on eleven existing group fairness metrics, and through both empirical and theoretical results we demonstrate that most of these metrics only satisfy a small subset of the proposed properties. These findings highlight limitations of existing metrics, and provide insights into how to evaluate and interpret different fairness metrics in practical deployment. The proposed properties can also assist practitioners in selecting appropriate metrics for evaluating fairness in a specific application.

cs.LG↗

arXiv2022

Adversarial Inter-Group Link Injection Degrades the Fairness of Graph Neural Networks

We present evidence for the existence and effectiveness of adversarial attacks on graph neural networks (GNNs) that aim to degrade fairness. These attacks can disadvantage a particular subgroup of nodes in GNN-based node classification, where nodes of the underlying network have sensitive attributes, such as race or gender. We conduct qualitative and experimental analyses explaining how adversarial link injection impairs the fairness of GNN predictions. For example, an attacker can compromise the fairness of GNN-based node classification by injecting adversarial links between nodes belonging to opposite subgroups and opposite class labels. Our experiments on empirical datasets demonstrate that adversarial fairness attacks can significantly degrade the fairness of GNN predictions (attacks are effective) with a low perturbation rate (attacks are efficient) and without a significant drop in accuracy (attacks are deceptive). This work demonstrates the vulnerability of GNN models to adversarial fairness attacks. We hope our findings raise awareness about this issue in our community and lay a foundation for the future development of GNN models that are more robust to such attacks.

cs.LG↗

arXiv2020

The POLAR Framework: Polar Opposites Enable Interpretability of Pre-Trained Word Embeddings

We introduce POLAR - a framework that adds interpretability to pre-trained word embeddings via the adoption of semantic differentials. Semantic differentials are a psychometric construct for measuring the semantics of a word by analysing its position on a scale between two polar opposites (e.g., cold -- hot, soft -- hard). The core idea of our approach is to transform existing, pre-trained word embeddings via semantic differentials to a new "polar" space with interpretable dimensions defined by such polar opposites. Our framework also allows for selecting the most discriminative dimensions from a set of polar dimensions provided by an oracle, i.e., an external source. We demonstrate the effectiveness of our framework by deploying it to various downstream tasks, in which our interpretable word embeddings achieve a performance that is comparable to the original word embeddings. We also show that the interpretable dimensions selected by our framework align with human judgement. Together, these results demonstrate that interpretability can be added to word embeddings without compromising performance. Our work is relevant for researchers and engineers interested in interpreting pre-trained word embeddings.

cs.CL↗

arXiv2020

The Effects of Gender Signals and Performance in Online Product Reviews

This work quantifies the effects of signaling and performing gender on the success of reviews written on the popular amazon shopping platform. Highly rated reviews play an important role in e-commerce since they are prominently displayed below products. Differences in how gender-signaling and gender-performing review authors are received can lead to important biases in what content and perspectives are represented among top reviews. To investigate this, we extract signals of author gender from user names, distinguishing reviews where the author's likely gender can be inferred. Using reviews authored by these gender-signaling authors, we train a deep-learning classifier to quantify the gendered writing style or gendered performance of reviews written by authors who do not send clear gender signals via their user name. We contrast the effects of gender signaling and performance on review success using matching experiments. While we find no general trend that gendered signals or performances influence overall review success, we find strong context-specific effects. For example, reviews in product categories such as Electronics or Computers are perceived as less helpful when authors signal that they are likely woman, but are received as more helpful in categories such as Beauty or Clothing. In addition to these interesting findings, our work provides a general chain of tools for studying gender-specific effects across various social media platforms.

cs.CY↗

arXiv2019

StRE: Self Attentive Edit Quality Prediction in Wikipedia

Wikipedia can easily be justified as a behemoth, considering the sheer volume of content that is added or removed every minute to its several projects. This creates an immense scope, in the field of natural language processing towards developing automated tools for content moderation and review. In this paper we propose Self Attentive Revision Encoder (StRE) which leverages orthographic similarity of lexical units toward predicting the quality of new edits. In contrast to existing propositions which primarily employ features like page reputation, editor activity or rule based heuristics, we utilize the textual content of the edits which, we believe contains superior signatures of their quality. More specifically, we deploy deep encoders to generate representations of the edits from its text content, which we then leverage to infer quality. We further contribute a novel dataset containing 21M revisions across 32K Wikipedia pages and demonstrate that StRE outperforms existing methods by a significant margin at least 17% and at most 103%. Our pretrained model achieves such result after retraining on a set as small as 20% of the edits in a wikipage. This, to the best of our knowledge, is also the first attempt towards employing deep language models to the enormous domain of automated content moderation and review in Wikipedia.

cs.SI↗

arXiv2018

Using Core-Periphery Structure to Predict High Centrality Nodes in Time-Varying Networks

Vertices with high betweenness and closeness centrality represent influential entities in a network. An important problem for time varying networks is to know a-priori, using minimal computation, whether the influential vertices of the current time step will retain their high centrality, in the future time steps, as the network evolves. In this paper, based on empirical evidences from several large real world time varying networks, we discover a certain class of networks where the highly central vertices are part of the innermost core of the network and this property is maintained over time. As a key contribution of this work, we propose novel heuristics to identify these networks in an optimal fashion and also develop a two-step algorithm for predicting high centrality vertices. Consequently, we show for the first time that for such networks, expensive shortest path computations in each time step as the network changes can be completely avoided; instead we can use time series models (e.g., ARIMA as used here) to predict the overlap between the high centrality vertices in the current time step to the ones in the future time steps. Moreover, once the new network is available in time, we can find the high centrality vertices in the top core simply based on their high degree.

cs.SI↗

arXiv2018

ComPAS: Community Preserving Sampling for Streaming Graphs

In the era of big data, graph sampling is indispensable in many settings. Existing sampling methods are mostly designed for static graphs, and aim to preserve basic structural properties of the original graph (such as degree distribution, clustering coefficient etc.) in the sample. We argue that for any sampling method it is impossible to produce an universal representative sample which can preserve all the properties of the original graph; rather sampling should be application specific (such as preserving hubs - needed for information diffusion). Here we consider community detection as an application scenario. We propose ComPAS, a novel sampling strategy that unlike previous methods, is not only designed for streaming graphs (which is a more realistic representation of a real-world scenario) but also preserves the community structure of the original graph in the sample. Empirical results on both synthetic and different real-world graphs show that ComPAS is the best to preserve the underlying community structure with average performance reaching 73.2% of the most informed algorithm for static graphs.

cs.SI↗

arXiv2017

Influence of Reviewer Interaction Network on Long-term Citations: A Case Study of the Scientific Peer-Review System of the Journal of High Energy Physics

A `peer-review system' in the context of judging research contributions, is one of the prime steps undertaken to ensure the quality of the submissions received, a significant portion of the publishing budget is spent towards successful completion of the peer-review by the publication houses. Nevertheless, the scientific community is largely reaching a consensus that peer-review system, although indispensable, is nonetheless flawed. A very pertinent question therefore is "could this system be improved?". In this paper, we attempt to present an answer to this question by considering a massive dataset of around $29k$ papers with roughly $70k$ distinct review reports together consisting of $12m$ lines of review text from the Journal of High Energy Physics (JHEP) between 1997 and 2015. In specific, we introduce a novel \textit{reviewer-reviewer interaction network} (an edge exists between two reviewers if they were assigned by the same editor) and show that surprisingly the simple structural properties of this network such as degree, clustering coefficient, centrality (closeness, betweenness etc.) serve as strong predictors of the long-term citations (i.e., the overall scientific impact) of a submitted paper. These features, when plugged in a regression model, alone achieves a high $R^2$ of \0.79 and a low $RMSE$ of 0.496 in predicting the long-term citations. In addition, we also design a set of supporting features built from the basic characteristics of the submitted papers, the authors and the referees (e.g., the popularity of the submitting author, the acceptance rate history of a referee, the linguistic properties laden in the text of the review reports etc.), which further results in overall improvement with $R^2$ of 0.81 and $RMSE$ of 0.46.

cs.DL↗

arXiv2016

Anomalies in the peer-review system: A case study of the journal of High Energy Physics

Peer-review system has long been relied upon for bringing quality research to the notice of the scientific community and also preventing flawed research from entering into the literature. The need for the peer-review system has often been debated as in numerous cases it has failed in its task and in most of these cases editors and the reviewers were thought to be responsible for not being able to correctly judge the quality of the work. This raises a question "Can the peer-review system be improved?" Since editors and reviewers are the most important pillars of a reviewing system, we in this work, attempt to address a related question - given the editing/reviewing history of the editors or re- viewers "can we identify the under-performing ones?", with citations received by the edited/reviewed papers being used as proxy for quantifying performance. We term such review- ers and editors as anomalous and we believe identifying and removing them shall improve the performance of the peer- review system. Using a massive dataset of Journal of High Energy Physics (JHEP) consisting of 29k papers submitted between 1997 and 2015 with 95 editors and 4035 reviewers and their review history, we identify several factors which point to anomalous behavior of referees and editors. In fact the anomalous editors and reviewers account for 26.8% and 14.5% of the total editors and reviewers respectively and for most of these anomalous reviewers the performance degrades alarmingly over time.

cs.DL↗