Machine Learning (ML) systems require critical decisions about when to retrain models and how much data to use, balancing predictive performance against computational cost. Current practices rely on static schedules or heuristics, leading to sub-optimal resource utilization. We formalize these decisions as a cost-aware dynamic programming problem and provide three algorithmic contributions: (1) an optimal offline algorithm, (2) a Bayesian optimization approach that approximates optimal solutions while overcoming computational bottlenecks, and (3) lightweight online policies derived from offline solutions. Experiments on real and synthetic datasets demonstrate that our algorithms effectively navigate the cost-performance trade-off, outperforming drift-detection methods and heuristic retraining strategies.

Learned indexes improve query performance by adapting search structures to data and workload distributions. Although many learned indexes have been proposed, their trade-offs remain insufficiently understood for spatial range queries, where performance depends not only on model accuracy but also on data and query skew, layout granularity, selectivity, and storage behavior. In this work, we perform an experimental study of learned indexes for spatial range queries. We examine a representative set of indexes and address seven fundamental questions: (1) How does block size influence query latency, and what configurations yield optimal performance under varying selectivities? (2) How do skewed data and query distributions impact index performance? (3) How do indexes balance refinement and scan costs, and which designs favor one over the other? (4) How do disk-based storage conditions alter optimal block size and latency trade-offs compared to in-memory settings? (5) What are the construction costs of different indexes, and under what query volumes are these costs amortized? (6) For a given data and query workload, which index is expected to perform best? (7) Do index-selection insights learned from synthetic data generalize to real-world data distributions? To enable the analysis, we use a framework with a common storage backend, standardized query execution pipelines, and controlled variations in data and query skew. Our experiments reveal critical insights into refinement vs. scan trade-offs, the impact of block size, and the interplay between selectivity and layout effectiveness. We synthesize these findings into a workload-based decision tree for index selection and validate it on real OpenStreetMap point sets with synthetic queries, confirming that its recommendations exhibit minimal decision regret and typically yield near-optimal query performance.

While digitized corpora have transformed the study of intellectual transmission, current methods rely heavily on lexical text reuse detection, capturing verbatim quotations but fundamentally missing paraphrases and complex implicit engagement. This paper evaluates semantic search in 18th-century intellectual history through the reception of John Locke's foundational work. Using expert annotation grounded in a semantic taxonomy, we examine whether an off-the-shelf semantic search pipeline can surface meaning-level correspondences overlooked by lexical methods. Our results demonstrate that semantic search retrieves substantially more implicit receptions than lexical baselines. However, linguistic diagnostics also reveal a "lexical gatekeeping" effect, where retrieval remains partially constrained by surface vocabulary overlap. These findings highlight both the potential and the limitations of semantic retrieval for analyzing the circulation of ideas in large historical corpora. The data is available at https://github.com/COMHIS/locke-sim-data.

Text reuse is of fundamental importance in humanities research, as near-verbatim pieces of text in different documents provide invaluable information about the historical spread, evolution of ideas and composition of cultural artifacts. Traditionally, scholars have studied text reuse at a very small scale, for example, when comparing the writings of two philosophers; however, modern digitized corpora spanning entire centuries promise to revolutionize humanities research through the detection of previously unobserved large-scale patterns. This paper presents insights from ReceptionReader, a system for large-scale text reuse analysis over almost all known 18th-century books, articles, and newspapers. The system implements a data management pipeline for billions of text reuse instances and supports analysis tasks based on database queries (e.g., retrieving the most reused quotes from queried documents). The paper describes the principled and extensive evaluations across different normalization levels, query execution engines, and queries of interest that led to an optimized system—and offers insights from the observed trade-offs and how they were resolved to fit specific requirements. In summary, the paper explains how, for our system, (1) the row-store engine (MariaDB Aria) with denormalized relations emerged as the optimal choice for front-end interfaces, while (2) big data processing (Apache Spark) proved irreplaceable for data preprocessing.

Retraining a machine learning (ML) model is essential for maintaining its performance as the data change over time. However, retraining is also costly, as it typically requires re-processing the entire dataset. As a result, a trade-off arises: on the one hand, retraining an ML model too frequently incurs unnecessary computing costs; on the other hand, not retraining frequently enough leads to stale ML models and incurs a cost in loss of accuracy. To resolve this trade-off, we envision ML systems that make automated and cost-optimal decisions about when to retrain an ML model.In this work, we study the decision problem of whether to retrain or keep an existing ML model based on the data, the model, and the predictive queries answered by the model. Crucially, we consider the costs associated with each decision and aim to optimize the trade-off. Our main contribution is a Cost-Aware Retraining Algorithm, CARA, which optimizes the trade-off over streams of data and queries. To explore the performance of CARA, we first analyze synthetic datasets and demonstrate that CARA can adapt to different data drifts and retraining costs while performing similarly to an optimal retrospective algorithm. Subsequently, we experiment with real-world datasets and demonstrate that CARA has better accuracy than drift detection baselines while making fewer retraining decisions, thus incurring lower total costs.

Text reuse is a methodological element of fundamental importance in humanities research: pieces of text that re-appear across different documents, verbatim or paraphrased, provide invaluable information about the historical spread and evolution of ideas. Large modern digitized corpora enable the joint analysis of text collections that span entire centuries and the detection of large-scale patterns, impossible to detect with traditional small-scale analysis. For this opportunity to materialize, it is necessary to develop efficient data science systems that perform the corresponding analysis tasks. In this paper, we share insights from ReceptionReader, a system for analyzing text reuse in large historical corpora. The system is built upon billions of instances of text reuses from large digitized corpora of 18th-century texts. Its main functionality is to perform downstream text reuse analysis tasks, such as finding reuses that stem from a given article or identifying the most reused quotes from a set of documents, with each task expressed as a database query. For the purposes of the paper, we discuss the related design choices including various database normalization levels and query execution frameworks, such as distributed data processing (Apache Spark), indexed row store engine (MariaDB Aria), and compressed column store engine (MariaDB Columnstore). Moreover, we present an extensive evaluation with various metrics of interest (latency, storage size, and computing costs) for varying workloads, and we offer insights from the trade-offs we observed and the choices that emerged as optimal in our setting. In summary, our results show that (1) for the workloads that are most relevant to text-reuse analysis, the MariaDB Aria framework emerges as the overall optimal choice, (2) big data processing (Apache Spark) is irreplaceable for all processing stages of the system's pipeline.

Learned indexes fit machine learning (ML) models to the data and use them to make query operations more time and space-efficient. Recent works propose using learned spatial indexes to improve spatial query performance by optimizing the storage layout or internal search structures according to the data distribution. However, only a few learned indexes exploit the query workload distribution to enhance their performance. In addition, building and updating learned spatial indexes are often costly on large datasets due to the inefficiency of (re)training ML models. In this paper, we present WaZI, a learned and workload-aware variant of the Z-index, which jointly optimizes the storage layout and search structures, as a viable solution for the above challenges of spatial indexing. Specifically, we first formulate a cost function to measure the performance of a Z-index on a dataset for a range-query workload. Then, we optimize the Z-index structure by minimizing the cost function through adaptive partitioning and ordering for index construction. Moreover, we design a novel page-skipping mechanism to improve the query performance of WaZI by reducing access to irrelevant data pages. Our extensive experiments show that the WaZI index improves range query time by 40% on average over the baselines while always performing better or comparably to state-of-the-art spatial indexes. Additionally, it also maintains good point query performance. Generally, WaZI provides favorable tradeoffs among query latency, construction time, and index size.

The k-center clustering problem is of fundamental importance for a broad range of machine learning and data science applications. In this paper, we study the deletion-robust version of the problem. Specifically, we aim to extract a small subset of a given data set, referred to as a coreset, that contains a provably good set of k centers even after an adversary deletes up to z arbitrarily chosen points from the data set. We propose a 4-approximation algorithm that provides a coreset of size O(kz). To our knowledge, this is the first algorithm for deletion-robust k-center clustering with a theoretical guarantee. Moreover, we accompany our theoretical results with extensive experiments, demonstrating that our algorithm achieves significantly better robustness than non-trivial baselines against three heuristic gray-box and white-box adversarial deletion attacks.

Learned indexes fit machine learning (ML) models to the data and use them to make query operations more time and space-efficient. Recent works propose using learned spatial indexes to improve spatial query performance by optimizing the storage layout or internal search structures according to the data distribution. However, only a few learned indexes exploit the query workload distribution to enhance their performance. In addition, building and updating learned spatial indexes are often costly on large datasets due to the inefficiency of (re)training ML models. In this paper, we present WaZI, a learned and workload-aware variant of the Z-index, which jointly optimizes the storage layout and search structures, as a viable solution for the above challenges of spatial indexing. Specifically, we first formulate a cost function to measure the performance of a Z-index on a dataset for a range-query workload. Then, we optimize the Z-index structure by minimizing the cost function through adaptive partitioning and ordering for index construction. Moreover, we design a novel page-skipping mechanism to improve the query performance of WaZI by reducing access to irrelevant data pages. Our extensive experiments show that the WaZI index improves range query time by 40% on average over the baselines while always performing better or comparably to state-of-the-art spatial indexes. Additionally, it also maintains good point query performance. Generally, WaZI provides favorable tradeoffs among query latency, construction time, and index size.

Diversity maximization is a fundamental problem with broad applications in data summarization, web search, and recommender systems. Given a set X of n elements, the problem asks for a subset S of k≪n elements with maximum diversity, as quantified by the dissimilarities among the elements in S. In this paper, we study diversity maximization with fairness constraints in streaming and sliding-window models. Specifically, we focus on the max–min diversity maximization problem, which selects a subset S that maximizes the minimum distance (dissimilarity) between any pair of distinct elements within it. Assuming that the set X is partitioned into m disjoint groups by a specific sensitive attribute, e.g., sex or race, ensuring fairness requires that the selected subset S contains ki elements from each group i∈[m]. Although diversity maximization has been extensively studied, existing algorithms for fair max–min diversity maximization are inefficient for data streams. To address the problem, we first design efficient approximation algorithms for this problem in the (insert-only) streaming model, where data arrive one element at a time, and a solution should be computed based on the elements observed in one pass. Furthermore, we propose approximation algorithms for this problem in the sliding-window model, where only the latest w elements in the stream are considered for computation to capture the recency of the data. Experimental results on real-world and synthetic datasets show that our algorithms provide solutions of comparable quality to the state-of-the-art offline algorithms while running several orders of magnitude faster in the streaming and sliding-window settings.

Graph summarization via node grouping is a popular method to build concise graph representations by grouping nodes from the original graph into supernodes and encoding edges into superedges such that the loss of adjacency information is minimized. Such summaries have immense applications in large-scale graph analytics due to their small size and high query processing efficiency. In this paper, we reformulate the loss minimization problem for summarization into an equivalent integer maximization problem. By initially allowing relaxed (fractional) solutions for integer maximization, we analytically expose the underlying connections to the spectral properties of the adjacency matrix. Consequently, we design an algorithm called SpecSumm that consists of two phases. In the first phase, motivated by spectral graph theory, we apply k-means clustering on the k largest (in magnitude) eigenvectors of the adjacency matrix to assign nodes to supernodes. In the second phase, we propose a greedy heuristic that updates the initial assignment to further improve summary quality. Finally, via extensive experiments on 11 datasets, we show that SpecSumm efficiently produces high-quality summaries compared to state-of-the-art summarization algorithms and scales to graphs with millions of nodes.

Recommender systems typically suggest to users content similar to what they consumed in the past. If a user happens to be exposed to strongly polarized content, she might subsequently receive recommendations which may steer her towards more and more radicalized content, eventually being trapped in what we call a "radicalization pathway". In this paper, we study the problem of mitigating radicalization pathways using a graph-based approach. Specifically, we model the set of recommendations of a "what-to-watch-next" recommender as a d-regular directed graph where nodes correspond to content items, links to recommendations, and paths to possible user sessions. We measure the "segregation" score of a node representing radicalized content as the expected length of a random walk from that node to any node representing non-radicalized content. High segregation scores are associated to larger chances to get users trapped in radicalization pathways. Hence, we define the problem of reducing the prevalence of radicalization pathways by selecting a small number of edges to "rewire", so to minimize the maximum of segregation scores among all radicalized nodes, while maintaining the relevance of the recommendations. We prove that the problem of finding the optimal set of recommendations to rewire is NP-hard and NP-hard to approximate within any factor. Therefore, we turn our attention to heuristics, and propose an efficient yet effective greedy algorithm based on the absorbing random walk theory. Our experiments on real-world datasets in the context of video and news recommendations confirm the effectiveness of our proposal.

Diversity maximization aims to select a diverse and representative subset of items from a large dataset. It is a fundamental optimization task that finds applications in data summarization, feature selection, web search, recommender systems, and elsewhere. However, in a setting where data items are associated with different groups according to sensitive attributes like sex or race, it is possible that algorithmic solutions for this task, if left unchecked, will under- or over-represent some of the groups. Therefore, we are motivated to address the problem of \emph{max-min diversification with fairness constraints}, aiming to select $k$ items to maximize the minimum distance between any pair of selected items while ensuring that the number of items selected from each group falls within predefined lower and upper bounds. In this work, we propose an exact algorithm based on integer linear programming that is suitable for small datasets as well as a $\frac{1-\varepsilon}{5}$-approximation algorithm for any $\varepsilon \in (0, 1)$ that scales to large datasets. Extensive experiments on real-world datasets demonstrate the superior performance of our proposed algorithms over existing ones.

It is important to retrain a machine learning (ML) model in order to maintain its performance as the data changes over time. However, this can be costly as it usually requires processing the entire dataset again. This creates a trade-off between retraining too frequently, which leads to unnecessary computing costs, and not retraining often enough, which results in stale and inaccurate ML models. To address this challenge, we propose ML systems that make automated and cost-effective decisions about when to retrain an ML model. We aim to optimize the trade-off by considering the costs associated with each decision. Our research focuses on determining whether to retrain or keep an existing ML model based on various factors, including the data, the model, and the predictive queries answered by the model. Our main contribution is a Cost-Aware Retraining Algorithm called Cara, which optimizes the trade-off over streams of data and queries. To evaluate the performance of Cara, we analyzed synthetic datasets and demonstrated that Cara can adapt to different data drifts and retraining costs while performing similarly to an optimal retrospective algorithm. We also conducted experiments with real-world datasets and showed that Cara achieves better accuracy than drift detection baselines while making fewer retraining decisions, ultimately resulting in lower total costs.

Diversity maximization is a fundamental problem with wide applications in data summarization, web search, and recommender systems. Given a set $X$ of $n$ elements, it asks to select a subset $S$ of $k\ll n$ elements with maximum diversity, as quantified by the dissimilarities among the elements in S. In this paper, we focus on the diversity maximization problem with fairness constraints in the streaming setting. Specifically, we consider the max-min diversity objective, which selects a subset $S$ that maximizes the minimum distance (dissimilarity) between any pair of distinct elements within it. Assuming that the set $X$ is partitioned into $m$ disjoint groups by some sensitive attribute, e.g., sex or race, ensuring fairness requires that the selected subset $S$ contains k i elements from each group i є [1, m]. A streaming algorithm should process $X$ sequentially in one pass and return a subset with maximum diversity while guaranteeing the fairness constraint. Although diversity maximization has been extensively studied, the only known algorithms that can work with the max-min diversity objective and fairness constraints are very inefficient for data streams. Since diversity maximization is NP-hard in general, we propose two approximation algorithms for fair diversity maximization in data streams, the first of which is $\frac{1-\varepsilon}{4}$ -approximate and specific for m = 2, where є E (0,1), and the second of which achieves a $\frac{1-\varepsilon}{3m+2}$ -approximation for an arbitrary $m$ . Experimental results on real-world and synthetic datasets show that both algorithms provide solutions of comparable quality to the state-of-the-art algorithms while running several orders of magnitude faster in the streaming setting.

Distance oracles answer shortest-path queries between any pair of nodes in a graph. They are often built using succinct graph representations such as spanners, sketches, and compressors to minimize oracle size and query answering latency. Node embeddings, in particular, offer graph representations that place adjacent nodes nearby each other in a low-rank space. However, their use in the design of distance oracles has not been sufficiently studied. In this paper, we empirically compare exact distance oracles constructed based on a variety of node embeddings and other succinct representations. We evaluate twelve such oracles along three measures of efficiency: construction time, memory requirements, and query-processing time over fourteen real datasets and four synthetic graphs. We show that distances between embedding vectors are excellent estimators of graph distances when graphs are well-structured, but less so for more unstructured graphs. Overall, our findings suggest that exact oracles based on embeddings can be constructed faster than multi-dimensional scaling (MDS) but slower than compressed adjacency indexes, require less memory than landmark oracles but more than sparsifiers or indexes, can answer queries faster than indexes but slower than MDS, and are exact more often with a smaller additive error than spanners (that have multiplicative error) while not being lossless like adjacency lists. Finally, while the exactness of such oracles is infeasible to maintain for huge graphs even under large amounts of resources, we empirically demonstrate that approximate oracles based on GOSH embeddings can efficiently scale to graphs of 100M+ nodes with only small additive errors in distance estimations.

The task of node classification concerns a network where nodes are associated with labels, but labels are known only for some of the nodes. The task consists of inferring the unknown labels given the known node labels, the structure of the network, and other known node attributes. Common node classification approaches are based on the assumption that adjacent nodes have similar attributes and, therefore, that a node's label can be predicted from the labels of its neighbors. While such an assumption is often valid (e.g., for political affiliation in social networks), it may not hold in some cases. In fact, nodes that share the same label may be adjacent but differ in their attributes, or may not be adjacent but have similar attributes. In this work, we present JANE (Jointly using Attributes and Node Embeddings), a novel and principled approach to node classification that flexibly adapts to a range of settings wherein unknown labels may be predicted from known labels of adjacent nodes in the network, other node attributes, or both. Our experiments on synthetic data highlight the limitations of benchmark algorithms and the versatility of JANE. Further, our experiments on seven real datasets of sizes ranging from 2.5K to 1.5M nodes and edge homophily ranging from 0.86 to 0.29 show that JANE scales well to large networks while also demonstrating an up to 20% improvement in accuracy compared to strong baseline algorithms.

We study university admissions under a centralized system that uses grades and standardized test scores to match applicants to university programs. In the context of this system, we explore affirmative action policies that seek to narrow the gap between the admission rates of different socio-demographic groups while still accepting students with high scores. Since there is uncertainty about the score distribution of the students who will apply to each program, it is unclear what policy would have the desired effect on the admission rates of different groups. We address this challenge by using a predictive model trained on historical data to help optimize the parameters of such policies. We find that a learned predictive model does significantly better than relying on the ideal parameters for the last year. At the same time, we also find that a large pool of historical data yields similar results as our predictive approach for our data. Due to the more complex nature of the predictive approach, we conclude that a simpler approach should be preferred if enough data is available (e.g., long-standing, traditional university programs), but not for newer programs and other cases in which our predictive strategy can prove helpful.

Linear classifiers are well-known to be vulnerable to adversarial attacks: they may predict incorrect labels for input data that are adversarially modified with small perturbations. However, this phenomenon has not been properly understood in the context of sketch-based linear classifiers, typically used in memory-constrained paradigms, which rely on random projections of the features for model compression. In this paper, we propose novel Fast-Gradient-Sign Method (FGSM) attacks for sketched classifiers in full, partial, and black-box information settings with regards to their internal parameters. We perform extensive experiments on the MNIST dataset to characterize their robustness as a function of perturbation budget. Our results suggest that, in the full-information setting, these classifiers are less accurate on unaltered input than their uncompressed counterparts but just as susceptible to adversarial attacks. But in more realistic partial and black-box information settings, sketching improves robustness while having lower memory footprint.

Machine unlearning is the task of updating machine learning (ML) models after a subset of the training data they were trained on is deleted. Methods for the task are desired to combine effectiveness and efficiency (i.e., they should effectively "unlearn" deleted data, but in a way that does not require excessive computational effort (e.g., a full retraining) for a small amount of deletions). Such a combination is typically achieved by tolerating some amount of approximation in the unlearning. In addition, laws and regulations in the spirit of "the right to be forgotten" have given rise to requirements for certifiability (i.e., the ability to demonstrate that the deleted data has indeed been unlearned by the ML model). In this paper, we present an experimental study of the three state-of-the-art approximate unlearning methods for logistic regression and demonstrate the trade-offs between efficiency, effectiveness and certifiability offered by each method. In implementing this study, we extend some of the existing works and describe a common unlearning pipeline to compare and evaluate the unlearning methods on six real-world datasets and a variety of settings. We provide insights into the effect of the quantity and distribution of the deleted data on ML models and the performance of each unlearning method in different settings. We also propose a practical online strategy to determine when the accumulated error from approximate unlearning is large enough to warrant a full retraining of the ML model.