[KDD '26: Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining] [Code]
Md Khairul Islam, Zeyu Xia, Ryan Goudjil, Jialu Wang, Arya Farahi, Judy Fox
University of Virginia, Virginia, USA University of Texas at Austin, Austin, Texas, USA
Reconstructing the early Universe from the evolved present-day Universe is a challenging and computationally demanding problem in modern astrophysics. We devise a novel generative framework, Cosmo3DFlow, designed to address dimensionality and sparsity, the bottlenecks inherent in current methods for cosmological inference. By integrating 3D Discrete Wavelet Transform (DWT) with flow matching, we effectively represent high-dimensional cosmological structures. Using large-scale cosmological 𝑁 -body simulations, at 128^3 resolution, we achieve up to 50× faster sampling than diffusion models, combining a 10× reduction in integration steps with lower per-step computational cost from wavelet compression
Md Khairul Islam, Judy Fox
University of Virginia, USA
We present OmniSpectra, a native-resolution foundation model for astronomy spectra. Unlike traditional models, which are limited to fixed-length input sizes or configurations, OmniSpectra handles spectra of any length at their original size, without resampling or interpolation. OmniSpectra demonstrates excellent zero-shot generalization compared to methods tailored for specific tasks. This transfer learning capability makes this model the state-of-the-art across various astronomy tasks, including source classification, redshift estimation, and properties prediction for stars and galaxies.
Mills Staylor, Amirreza Dolatpour Fathkouhi, Md Khairul Islam, Kaleigh O’Hara, Ryan Ghiles Goudjil, Geoffrey Fox, Judy Fox
University of Virginia, USA
Current deep-learning models on large-scale astronomical image data demand substantial computational resources. We introduce the Cloud-based Astronomy Inference (CAI) framework to address these challenges. This scalable solution integrates pre-trained foundation models with serverless cloud infrastructure. Using a foundation model for redshift prediction as a case study, our extensive experiments cover user devices, HPC (High-Performance Computing) servers, and Cloud. CAI’s significant scalability improvement on large data sizes provides an accessible and effective tool for the astronomy community.
Md Khairul Islam, Ayush Karmacharya, Timothy Sue, Judy Fox
University of Virginia, USA
Existing interpretation methods are limited by focusing mostly on classification tasks, evaluating using custom baseline models, using simple synthetic datasets, and requiring training another model. We introduce a novel interpretation method called Windowed Temporal Saliency Rescaling (WinTSR) addressing these limitations. This captures temporal dependencies among the past time steps and efficiently scales the feature importance. We benchmark WinTSR against 10 recent interpretation techniques with 5 state-of-the-art models of different architectures, including a foundation model. We use 3 real-world datasets for classification and regression. WinTSR significantly outranks the other local interpretation methods in overall performance. Finally, we provide a novel and open-source framework to interpret the latest time series transformers and foundation models.
Md Khairul Islam, Ayush Karmacharya, Timothy Sue, Judy Fox
University of Virginia, USA
We use state-of-the-art time series models including pre-trained LLMs (GPT-2 as the backbone), transformers, and other models to demonstrate their ability to outperform traditional approaches with minimal (”few-shot”) or no fine-tuning (”zero-shot”). Our benchmark study with eight financial time series tasks, shows the potential of using LLMs for scarce financial datasets.
Md Khairul Islam, Tyler Valentine, Timothy Joowon Sue, Ayush Karmacharya, Luke Neil, Benham, Zhengguang Wang, Kingsley Kim, Judy Fox
University of Virginia, USA
We interpreted six latest time series transformer-based models with eight recent local interpretation methods. We showed how we can efficiently benchmark the interpretation performance of those methods. The primary dataset consists of daily COVID-19 infection cases collected from 3,142 US counties for three years and around 3.5 million sample instances. Then we showed the generability of our approach using two other popular time series datasets.
Md Khairul Islam, Judy Fox
University of Virginia, USA
Interpreting time series forecasting models faces unique challenges compared to image and text data. These challenges arise from the temporal dependencies between time steps and the evolving importance of input features over time. My thesis focuses on addressing these challenges by aiming for more precise explanations of feature interactions, uncovering spatiotemporal patterns, and demonstrating the practical applicability of these interpretability techniques using real-world datasets and state-of-the-art deep learning models.
Md Khairul Islam, Di Zhu, Yingzheng Liu, Andrej Erkelen, Nick Daniello, Aparna Marathe, and Judy Fox
University of Virginia, USA
We forecast US county-level COVID-19 infections using the Temporal Fusion Transformer (TFT). We focus on heterogeneous time-series deep learning model prediction while interpreting the complex spatiotemporal features learned from the data. We collected around 2.5 years of socioeconomic and health features for 3142 US counties. Our results show that the TFT model outperforms other baseline models in all evaluation metrics. We then interpreted the temporal and spatial patterns learned by the TFT model using the multi-head self-attention weights.
Md. Khairul Islam 1, Andrew Wang 1, Tianhao Wang 1 , Yangfeng Ji 1 , Judy Fox 1 , Jieyu Zhao 2
1 University of Virginia, 2 University of Maryland, College Park
In this work, we show the impact of DP on bias in LMs. Differential private training can increase the model bias against protected groups w.r.t AUC-based bias metrics. DP makes it more difficult for the model to differentiate between the positive and negative examples from the protected groups and other groups in the rest of the population.
Md. Khairul Islam 1, Andrew Wang 1, Tianhao Wang 1 , Yangfeng Ji 1 , Jieyu Zhao 2
1 University of Virginia, 2 University of Maryland, College Park
In this work, we show through empirical analysis the impact of DP on bias in LLMs. We find that differentially private training can increase the model bias against protected groups w.r.t AUC-based bias metrics. DP makes it more difficult for the model to differentiate between the positive and negative examples from the protected groups and other groups in the rest of the population.
Arup Kumar Sarker 1, Md Khairul Islam 1, Yuan Tian 2, Geoffrey Fox 1
1 University of Virginia, 2 University of California, Los Angeles, USA
This paper examines the vulnerabilities of the TrustZone extension of ARM Cortex-M processors and develops a threat model to carry out these attacks. After performing multi-variety attacks from different angles, it is found that TrustZone is susceptible to buffer overflow attacks that can compromise the security of other trusted apps. Finally, a trust model is proposed to address these vulnerabilities.
Khairul Islam 1, Toufique Ahmed 2, Rifat Shahriyar 1, Anindya Iqbal 1, and Gias Uddin 3
1 Bangladesh University of Engineering and Technology, 2 University of California, Davis, and 3 University of Calgary
In this work, we have presented a LightGBM classifier-based tool called PredCR, that can predict whether a code change would be merged or abandoned as soon as the code change request is submitted. We have mined 146,612 code changes from the code reviews of three large and popular open-source software. Using longitudinal ten-fold cross-validation, our tool achieves an 85% AUC score on average and relatively improves the state-of-the-art by 14-23%.
Md. Khairul Islam 1, Prithula Hridi 1, Md. Shohrab Hossain 1, Husnu S. Narman 2
1 Bangladesh University of Engineering and Technology, 2 Marshall University
In this paper, we have worked on a benchmark network anomaly detection dataset UNSW-NB15. We have used a machine learning classifier LightGBM to perform binary classification on this dataset and achieved state-of-the-art performance. Using ten-fold cross-validation on the train, test, and combined dataset, our model has achieved 97.21%, 98.33%, and 96.21% f1 scores, respectively. Also, the model fitted only on train data achieved a 92.96% f1 score on the separate test data.