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.

In data exploration, executing queries over a large database can be time-consuming. Previous work has proposed approximate query processing as a way to speed up aggregate queries in this context, but do not address non-aggregate queries. Our paper introduces a novel holistic approach to handle both types of queries by finding an optimized subset of data, referred to as an approximation set. The goal is to maximize query result quality while using a smaller set of data, thereby significantly reducing the query execution time. We formalize this problem as Holistic Approximate Query Processing and establish its NP-completeness. To tackle this, we propose an approximate solution using Reinforcement Learning , termed HARLM. While HARLM does not provide theoretical guarantees due to its reliance on Reinforcement Learning, it effectively overcomes challenges related to the large action space and the need for generalization beyond a known query workload. Experimental results on both non-aggregate and aggregate benchmarks show that HARLM significantly outperforms the baselines both in terms of accuracy (30% improvement) and efficiency (10–35X).

Logs produced by software applications are invaluable for spotting deviations from expected system behavior. However, automatically detecting anomalies from log data is challenging due to the volume, semi-structured nature, lack of standard formatting, and potential evolution of log records over time. In this work, we approach log-based anomaly detection as a semantic similarity problem. We generate pairwise similarity scores using a general-purpose pre-trained language model and further augment them with ground-truth binary labels. The generated similarity labels supervise an encoder trained for semantic similarity. At inference time, anomalies are detected based on the cosine similarity between the encoded query sequence and the average normal encoding. Our method outperforms contemporary techniques on multiple benchmarks without template extraction or a fixed vocabulary and achieves competitive performance even when provided with limited abnormal examples.

Tabular embedding methods have become increasingly popular due to their effectiveness in improving the results of various tasks, including classic databases tasks and machine learning predictions. However, most current methods treat these embedding models as "black boxes" making it difficult to understand the insights captured by the models. Our research proposes a novel approach to interpret these models, aiming to provide local and global explanations for the original data and detect potential flaws in the embedding models. The proposed solution is appropriate for every tabular embedding algorithm, as it fits the black box view of the embedding model. Furthermore, we propose methods for comparing different embedding models, which can help identify data biases that might impact the models' credibility without the user's knowledge. Our approach is evaluated on multiple datasets and multiple embeddings, demonstrating that our proposed explanations provide valuable insights into the behavior of tabular embedding methods. By making these models more transparent, we believe our research will contribute to the development of more effective and reliable embedding methods for a wide range of applications.

We present TabEE, Tab ular E mbedding E xplanations, a framework designed to generate explanations for interpreting tabular embedding models. Our framework aims to furnish both local and global explanations for the original data, facilitating the detection of potential flaws in embedding models. TabEE is versatile and compatible with any tabular embedding algorithm, as it adheres to the black box perspective of embedding models. The generated explanations also enable comparisons between multiple embedding models. This demonstration illustrates the effectiveness of TabEE in providing interpretable insights into tabular embedding models, contributing to improved model understanding and credibility assessment.

In the era of data-driven decision-making, efficient machine learning model training is crucial. We present a novel algorithm for constructing tabular data coresets using datamaps created for Gradient Boosting Decision Trees models. The resulting coresets, computed within minutes, consistently outperform other baselines and match or exceed the performance of models trained on the entire dataset. Additionally, a training enhancement method leveraging datamap insights during the inference phase improves performance with mathematical guarantees, given a defined property holds. An explainability layer and tools for coreset size optimization further enhance the efficiency of training tabular machine learning models.

Data scientists frequently examine the raw content of large tables when exploring an unknown dataset. In such cases, small subsets of the full tables (sub-tables) that accurately capture table contents are useful. We present a framework which, given a large data table T, creates 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?Our first contribution is an informativeness metric for sub-tables with two complementary dimensions: cell coverage, which measures how well the sub-table captures prominent data patterns in T, and diversity. We use association rules as the patterns captured by sub-tables, and show that computing optimal sub-tables directly using this metric is infeasible. We then develop an efficient algorithm that indirectly accounts for association rules using table embedding. The resulting framework produces sub-tables for the full table as well as for the results of queries over the table, enabling the user to quickly understand results and determine subsequent queries. Experimental results show that high-quality sub-tables can be efficiently computed, and verify the soundness of our metrics as well as the usefulness of selected sub-tables through user studies.

Search mechanisms over large assortments of items are central to the operation of many platforms. As users commonly express filtering conditions based on item properties that are not initially stored, companies must derive the missing information by training and applying binary classifiers. Choosing which classifiers to construct is however not trivial, since classifiers differ in construction costs and range of applicability. Previous work has considered the problem of selecting a classifier set of minimum construction cost, but this has been done under the (often unrealistic) assumption that the available budget is unlimited and allows to support all search queries. In practice, budget constraints require prioritizing some queries over others. To capture this consideration, we study in this work a more general model that allows assigning to each search query a score that models how important it is to compute its result set and examine the optimization problem of selecting a classifier set, whose cost is within the budget, that maximizes the overall score of the queries it can answer. We show that this generalization is likely much harder to approximate complexity-wise, even assuming limited special cases. Nevertheless, we devise a heuristic algorithm, whose effectiveness is demonstrated in our experimental study over real-world data, consisting of a public dataset and datasets provided by a large e-commerce company that include costs and scores derived by business analysts. Finally, we show that our methods are applicable also for related problems in practical settings where there is some flexibility in determining the budget.

We demonstrate SubTab, a framework for creating small, informative sub-tables of large data tables to speed up data exploration. Given a table with n rows and m columns where n and m are large, SubTab creates a sub-table T_sub with k<n rows and l

We present WiClean, a Wikipedia plug-in that supports the identification and cleaning of certain types of errors in Wikipedia interlinks. The system mines update patterns in Wikipedia revision logs, identifies the common time frames in which they occur, and employs them to signal incomplete/inconsistent updates and suggests corrections. We demonstrate the effectiveness of WiClean in identifying actual errors in a variety of Wikipedia entity types, interactively employing the VLDB'19 audience as editors to correct the identified errors.