arXiv subjects

Kathy Razmadze

Publications and source records attributed to Kathy Razmadze.

4 recordsLinked to original sources

arXiv2026

Quantifying the Pre-training Dividend: Generative versus Latent Self-Supervised Learning for Time Series Foundation Models

The success of self-supervised learning (SSL) in vision and NLP has motivated its rapid adoption for time series. However, research has focused primarily on Generative paradigms and forecasting tasks, leaving the broader utility of learned representations unquantified. We establish a controlled framework to evaluate the "pre-training dividend": the value added by SSL across diverse temporal tasks. We systematically compare Generative paradigms against Latent Alignment architectures, introducing adaptations of LeJEPA and DINO for time series. These adaptations utilize Discrete Wavelet Transform (DWT) augmentations to enforce invariance to local fluctuations. Our analysis reveals that the pre-training dividend is highly asymmetric: SSL yields gains of up to 375% for anomaly detection and classification, yet remains marginal for forecasting. We demonstrate that representational utility is non-universal, governed by a precision-invariance trade-off where the specific signal resolution required by the task must align with the objective. Finally, we show that representation quality is largely independent of data origin and saturates at moderate architectural depths, suggesting a path to scaling via massive synthetic generation. Our code is available at: https://github.com/noammajor/Models

cs.LG↗

arXiv2025

Differentially Private Explanations for Clusters

The dire need to protect sensitive data has led to various flavors of privacy definitions. Among these, Differential privacy (DP) is considered one of the most rigorous and secure notions of privacy, enabling data analysis while preserving the privacy of data contributors. One of the fundamental tasks of data analysis is clustering , which is meant to unravel hidden patterns within complex datasets. However, interpreting clustering results poses significant challenges, and often necessitates an extensive analytical process. Interpreting clustering results under DP is even more challenging, as analysts are provided with noisy responses to queries, and longer, manual exploration sessions require additional noise to meet privacy constraints. While increasing attention has been given to clustering explanation frameworks that aim at assisting analysts by automatically uncovering the characteristics of each cluster, such frameworks may also disclose sensitive information within the dataset, leading to a breach in privacy. To address these challenges, we present DPClustX, a framework that provides explanations for black-box clustering results while satisfying DP. DPClustX takes as input the sensitive dataset alongside privately computed clustering labels, and outputs a global explanation, emphasizing prominent characteristics of each cluster while guaranteeing DP. We perform an extensive experimental analysis of DPClustX on real data, showing that it provides insightful and accurate explanations even under tight privacy constraints.

cs.CR↗

arXiv2024

Learning Approximation Sets for Exploratory Queries

In data exploration, executing complex non-aggregate queries over large databases can be time-consuming. Our paper introduces a novel approach to address this challenge, focusing on finding an optimized subset of data, referred to as the approximation set, for query execution. The goal is to maximize query result quality while minimizing execution time. We formalize this problem as Approximate Non-Aggregates Query Processing (ANAQP) and establish its NP-completeness. To tackle this, we propose an approximate solution using advanced Reinforcement Learning architecture, termed ASQP-RL. This approach overcomes challenges related to the large action space and the need for generalization beyond a known query workload. Experimental results on two benchmarks demonstrate the superior performance of ASQP-RL, outperforming baselines by 30% in accuracy and achieving efficiency gains of 10-35X. Our research sheds light on the potential of reinforcement learning techniques for advancing data management tasks. Experimental results on two benchmarks show that ASQP-RL significantly outperforms the baselines both in terms of accuracy (30% better) and efficiency (10-35X). This research provides valuable insights into the potential of RL techniques for future advancements in data management tasks.

cs.DB↗

arXiv2022

Selecting Sub-tables for Data Exploration

We present a framework for creating small, informative sub-tables of large data tables to facilitate the first step of data science: data exploration. Given a large data table table T, the goal is to create a sub-table of small, fixed dimensions, by selecting a subset of T's rows and projecting them over a subset of T's columns. The question is: which rows and columns should be selected to yield an informative sub-table? We formalize the notion of "informativeness" based on two complementary metrics: cell coverage, which measures how well the sub-table captures prominent association rules in T, and diversity. Since computing optimal sub-tables using these metrics is shown to be infeasible, we give an efficient algorithm which indirectly accounts for association rules using table embedding. The resulting framework can be used for visualizing the complete sub-table, as well as for displaying the results of queries over the sub-table, enabling the user to quickly understand the results and determine subsequent queries. Experimental results show that we can efficiently compute high-quality sub-tables as measured by our metrics, as well as by feedback from user-studies.

cs.DB↗