Local intrinsic dimensionality (LID) provides insight into the behavior of individual training points in deep neural networks, with applications including adversarial detection, prevention of dimensional collapse in self-supervised learning, and identification of untruthful responses from large language models (LLMs). In such contexts, efficient LID estimation has depended on the use of mini-batches, due to the high cost of computing neighborhoods in latent space. However, estimation with respect to small subsets of the training data usually reflects the dimensionality of the global manifold structure rather than the intended local distribution around each point. In this paper, we propose the Nearest Distance Cache (NDC), a method that improves the locality of LID estimation by reusing nearest-neighbor distances observed in past mini-batches. This strategy faces two key challenges: representations evolve over time, and limited memory prevents storing all past distances. To address these, NDC maintains a compact cache of nearest distances per example and uses window-based change detection to discard outdated samples affected by distributional drift. We also evaluate NDC on two tasks: an autoencoder trained on synthetic data with known ground-truth LID, and a ResNet trained on CIFAR-10. Results show that NDC captures local properties of deep representations not revealed by single mini-batch estimates.
Large language models (LLMs) can implement row-level semantic transformations over tabular data, but a general-purpose model reserves the same accelerator resources even when every row uses one fixed prompt. IOLM-DB compiles that recurring prompt–column pair into an operator sized to the work it actually does. It samples the target column, constructs calibration sequences that represent both prompt and output behavior, evaluates specialized candidates, and selects an implementation under explicit memory and quality constraints. Across heterogeneous datasets and operator types, column-calibrated quantization reduces the resident footprint of the reference model by 1.8–2.8× while the 8-bit profile remains a near-lossless replacement and the 4-bit profile preserves behavior on categorical label-output operators. At fixed precision, column calibration consistently improves fidelity over generic GPTQ calibration, and output-aware calibration provides an additional benefit for generative operators. Deployment experiments reveal two complementary outcomes: weight-only compression does not raise per-row throughput in a compute-rich regime, where it primarily releases capacity, but the smaller weight stream also improves throughput when memory bandwidth becomes limiting. The compact artifacts enable deployment under a memory budget where the full-precision operator cannot start and allow three specialized operators to occupy roughly the footprint of one full-precision model. IOLM-DB therefore turns a recurring semantic transformation from an immutable call to a general-purpose backend into a compact, measurable, and hardware-aware operator.
This work is a companion reproducibility paper for “Data augmentation via context similarity: An application to biomedical Named Entity Recognition” (Bartolini et al., 2023). In the original paper, we presented COntext SImilarity-based data augmentation for NER (COSINER), a data augmentation method for Named Entity Recognition (NER) tasks that uses context similarity to generate plausible training data. In this paper, we provide a fully integrated reproducibility package consisting of scripts, configuration files, and execution workflows that automate the setup, execution, and evaluation of all COSINER experiments. Our goal is to give other researchers easy access to the code and enable full reproducibility of the experiments, addressing the common lack of reproducibility in published work. We also extend COSINER’s compatibility to multiple operating systems, including Windows and Linux. Finally, we present a new set of replicated experiments obtained with the updated codebase and analyze their similarities and differences with the original results, showing the validity of the old results.
Searching in highly homogeneous video domains is challenging, especially when relying solely on human memory. This difficulty arises because homogeneous content often requires domain-specific vocabulary to describe effectively, and a single text label might fit a substantial subset of the database. This paper investigates supplementing text queries with spatial location information to better address specific search intents—a strategy applicable when users possess strong visual memory of a target object, or have external knowledge of its position. To process this location information effectively, we formally define and evaluate static grid and dynamic segmentation strategies for video frame partitioning. Furthermore, we present a large-scale cognitive user study involving 220 participants, designed to simulate realistic memory constraints of Known-Item Search (KIS) tasks commonly found in interactive retrieval benchmarks. The study utilizes standard working memory interference techniques, comprising a target exposure phase, a distraction period, and subsequent query specification from memory. These queries are then evaluated against the proposed spatial retrieval models to determine how human spatial memory decay impacts retrieval effectiveness. Our results reveal that while dynamic segmentation models achieve the highest theoretical retrieval bounds under perfect conditions, their strict geometric boundaries are severely affected by human memory decay. Consequently, static overlapping grids demonstrate high robustness, outperforming dynamic models under more realistic, memory-driven search constraints. Finally, our experiments show that the observed spatial annotation perturbations can be modeled using a four-dimensional Kernel Density Estimation (KDE) method, enabling the simulation of realistic human memory decay on ideal bounding boxes.
The rapid growth of data repositories limits user’s awareness of their content and the understanding of interrelationships between diverse datasets, increasing the demand for effective data discovery tools that help analysts locate relevant data. Join discovery, the process of finding tables that can be joined based on common attributes, is one possible strategy. This paper introduces a novel join discovery approach tailored to data structures commonly used in analytics: multidimensional or analytic cubes. By leveraging the unique structure of analytic cubes, the approach utilizes the concept of profile, which refers to metadata providing information about data distribution. Specifically, we introduce a profile-based, query-driven join discovery model that allows users to identify alternative joinable cubes starting from a declarative query, as well as to rank and qualitatively assess alternatives. We validate the efficiency and effectiveness of our approach through extensive experimentation. Our dual approach, which combines schema-level and content-level information, enhances both the expressivity of join discovery and the clarity of join discovery results, making the analysis process more intuitive and interpretable.
In our primary work (Gagliardelli et al., 2024), we presented Generalized Supervised Meta-blocking, a novel approach that is based on the task of associating every pair of candidates with a probabilistic estimation of their matching likelihood. It formulates Meta-blocking as a probabilistic binary classification task, using a wide variety of weighting schemes as features. The resulting probability estimations can be combined with any pruning algorithm that improves blocking by retaining the best candidate comparisons, drastically reducing false positives without any significant impact on true positives. They can also be used by state-of-the-art progressive Entity Resolution methods to identify the most promising candidates as early as possible. We experimentally identified the best pruning algorithms, their optimal sets of features, and the minimum possible size of the training set. Our experiments demonstrate that the resulting approaches achieve excellent performance in several established benchmark datasets. This work introduces a companion reproducible paper of our previous work (Gagliardelli et al., 2024) to describe how to reproduce the entire experimental study and discuss how to extend it with new features, classification algorithms, and datasets. The presented reproducibility methodology is based on a Docker image and the NVIDIA CUDA Toolkit 12.0, which are evaluated on several Linux-based platforms equipped with at least 128 GB RAM, a GPU and at least 300 GB of free disk space. The reproducibility methodology leads to a weak reproducibility, with minor deviations from the experimental results reported in our primary work (Gagliardelli et al., 2024).
The object-centric paradigm has recently transformed research in Predictive Process Monitoring (PPM) thanks to new formalisms and techniques to handle ongoing process executions as interconnected and mutually influencing execution traces. Hence, the predictive objective of PPM methods has passed from delivering a prediction for a single object to delivering a prediction for all ongoing objects simultaneously. Building on this renewed goal of object-centric PPM, we introduce a natural language flow process representation to obtain the story telling of ongoing trace executions by preserving relationships between multiple objects involved in the same process execution, as well as relationships among the ongoing trace executions sharing the same objects. Based upon these representations, we introduce CICERONE a novel object-centric PPM method that uses a Large Language Model coupled with a global learning approach to deliver predictions for all objects involved in all ongoing process executions in a single step. Experiments across some Object-Centric Event Logs (OCELs) assess the effectiveness of CICERONE compared to competing object-centric PPM methods. In addition, we perform an occlusion-based analysis to examine the effect of object interactions regarding ongoing process executions on all the delivered predictions.
The integration of AI models into DBMSs has recently been proposed and implemented in tools such as PostgresAI. However, these solutions do not take full advantage of the multidimensional vision of information typically adopted in complex analytical tasks. At the same time, AI offers a chance to reconsider and improve analytical tasks based on the multidimensional model. In this work, we propose Cube+AI , a fresh vision of the multidimensional model aimed at extending the expressive power of cube queries to include semantically rich and multimodal operations on free text and images in addition to the analysis of categorical and numerical data. Cube+AI supports crossed semantic searches between text and images for filtering and grouping; moreover, it significantly extends aggregation by operating on text and images rather than on numeric data only. Thus, while the multidimensional model provides the structured context necessary to enhance AI execution, AI models expand the multidimensional paradigm by enabling the analysis and aggregation of unstructured data. The Cube+AI framework includes (i) a formal extension of the multidimensional model, which supports the dynamical definition of virtual categorical levels derived at query time from other levels via AI, as well as (ii) a text-to-SQL method where an LLM is leveraged to translate natural language queries into SQL. Specifically, we propose an implementation that relies on PostgresAI as a DBMS and on the Gemini LLM as a natural language querying interface. The paper is completed by the discussion of a set of experiments made with a sample workload of queries and by some robustness tests.
Deep Learning has revolutionized Entity Resolution (ER) by enabling high-accuracy matching through dense vector embeddings. However, this paradigm shift transforms the traditional blocking step into a high-dimensional similarity search problem, introducing a massive computational bottleneck that threatens scalability. While modern ER frameworks increasingly rely on blocking, the so-called indexing, to generate candidate pairs, the architectural design of this retrieval stage is often treated as a black box, with little attention paid to the resource implications of high-dimensional models. This paper conducts a rigorous design space exploration of blocking architectures, evaluating nine candidate generation strategies across graph-based, partition-based, and hash-based families. A primary contribution is a systematic analysis of Vector Quantization – specifically Product Quantization and Scalar Quantization – demonstrating that quantization is not merely a memory optimization, but a critical requirement for deploying state-of-the-art embeddings at scale. We evaluate these architectures on seven real-world datasets, measuring the interplay between embedding granularity, recall, and computational cost. Our results reveal that while graph-based methods offer peak precision, quantized partitioning methods reduce memory footprints by up to 96% with negligible loss in recall, effectively democratizing high-performance ER on commodity hardware. The study culminates in a decision framework for designing scalable, resource-aware blocking pipelines.
Entity Resolution (ER) constitutes a fundamental task in data management, involving the identification and linking of entities that refer to the same real-world object across different datasets. It is typically addressed through the Filtering-Verification framework, where the first step (also called Blocking) reduces the computational cost to the most similar pairs of entities, which are then analytically compared by the second step (also called Entity Matching). The latest advancements in Verification leverage Large Language Models (LLMs), which operate in an unsupervised manner and require no labelled training data. However, LLM-based matching is typically addressed as a classification task, independently of Filtering and without examining conflicts among the detected matches. Moreover, the substantial computational requirements of LLMs often limit their practical applicability in Entity Matching.In this work, we propose an open-source system called ChatMatcher, that implements end-to-end ER pipelines combining LLM-based Matching with Filtering and Entity Clustering. ChatMatcher provides a design space for matching prompts by combining different example orderings in few-shot prompts and several example selection strategies. To ensure that ChatMatcher runs on commodity hardware, it exclusively supports 7B-parameter LLMs with 4-bit quantization for Entity Matching. The performance of the resulting ER pipelines is evaluated on eight established ER datasets, with the experiments examining the relative performance of numerous configurations. In fact, our experimental study is the first to analyse how the precision–recall trade-off induced by Filtering affects LLM-based matching performance, and to demonstrate that Entity Clustering is essential to reconcile the inconsistent predictions produced by LLMs. The experimental results demonstrate that ChatMatcher can achieve state-of-the-art ER performance with resource-constrained models, despite hardware limitations.
Financial system digitization has intensified the tension between data analytics needs, privacy protection, and regulatory compliance. With annual money laundering exceeding $800 billion globally and financial institutions facing over 20,000 cyber incidents resulting in $12 billion in losses over two decades, the need for privacy-preserving computation techniques has become critical. Regulations including GDPR, PSD2, and CCPA impose stringent data protection requirements while demanding open data sharing for fraud detection, credit risk assessment, and regulatory oversight. Current approaches face a Privacy-Performance-Compliance Trilemma: no single technique can satisfy privacy guarantees, computational efficiency, and regulatory requirements at once. Current literature exhibits significant fragmentation, focusing on individual techniques without systematic quantitative comparison or practical deployment guidance. This survey provides a PRISMA-compliant review of privacy-preserving computation for financial applications, analyzing 90 carefully selected papers from 694 candidates spanning 2005-2025. We propose a hierarchical classification that organizes research into three main categories, i.e., Federated Learning, cryptographic primitives, and domain-specific applications—across six financial domains. We further evaluate representative methods on a common IBM anti-money laundering benchmark containing 5,078,345 transactions. By comparing centralized, local-only, federated, differentially private, secure aggregation, and encrypted aggregation settings, we show that cross-institutional collaboration improves anti-money laundering detection while exposing clear privacy-utility and efficiency trade-offs. Our findings suggest that hybrid solutions are necessary for building production-ready privacy-preserving financial systems.
Evaluating process discovery algorithms requires ground truth models against which discovered models can be compared. A critical but underexplored dimension of such evaluations concerns how algorithms perform when the underlying process model is not static but evolving. For example, processes can become more or less restrictive over time due to regulatory, organizational, or operational changes. Addressing this requires the ability to systematically construct sets of process models that are hierarchically related through specialization and generalization. While procedural ground truth model generation is well established, declarative process mining lacks comparable support, and the ability to systematically specialize or generalize generated models is largely absent. This gap is consequential given the growing importance of declarative models in dynamic business environments, where behavior is defined through constraints rather than explicit execution paths. This paper addresses this gap by presenting algorithms for the synthetic generation and specialization of Declare ground truth models, following the methodology of algorithm engineering. The proposed specialization algorithm is grounded in a theorem that characterizes when one declarative model constitutes a specialization of another, enabling controlled reduction of allowable behavior through constraint modification. Together, these techniques support the construction of hierarchically related model sets for use in process discovery and conformance checking evaluations across a spectrum of behavioral flexibility. We provide knowledge about the design of the algorithms by evaluating effectiveness and efficiency. Results show that model generation succeeds in over 80% of cases when models contain between 1 and 25 constraints and involve 11 to 35 activities. Increasing the number of constraints negatively impacts success rates and execution time, with the generator exhibiting exponential and the specializer linear time trends. Template choice also affects performance, with Response and Precedence improving success rates while Exclusive Choice reduces them. These findings offer practical guidance for generating declarative ground truth models that are both realistic and systematically varied, supporting more rigorous evaluation of declarative process mining algorithms.
The mobility sector plays a critical role in modern societies, encompassing diverse types of movement for both tourism and work-related purposes. Moreover, this sector is linked with the current ability to widely generate real-time data, providing valuable information that can be leveraged for analytics and informed decisionmaking. However, extracting actionable insights remains challenging due to the pronounced heterogeneity of data sources and the often unreliable nature of traditional data exchange environments. Data spaces have emerged as a solution to these challenges, providing secure and governed environments for inter-organizational data sharing while ensuring that data owners retain full sovereignty over what they provide. Specifically, data spaces enable secure collaborations between consumers, who need to exploit their data, and providers, who offer value-added services to support decision-making processes. Nonetheless, despite their potential, research on data spaces is still in its early stages, and there is no clear consensus on the core concepts that underpin their practical implementation. Therefore, we first conduct a comprehensive literature review to design an architecture for data spaces and then we present the implementation of this architecture in a mobility data space. Our case study demonstrates the practical applicability of data spaces in the mobility domain, illustrating how they can be used not only to activate datasets but also to enable value-added services built upon those data, such as interactive dashboards. The results show that data spaces offer a viable approach for effectively leveraging mobility data, supporting both operational optimization and strategic decision-making in transportation services.
Prescriptive Process Monitoring is an emerging area within Process Mining that focuses on recommending actions to optimize business outcomes. Most existing works focus on prescribing pre-defined interventions, that is, pre-defined (sets of) actions, on specific ongoing process executions, aimed at achieving a specific objective or Key Performance Indicator (KPI). In contrast, only a few approaches have explored the learning and evaluation of optimal behavioral policies, that is, general strategies that determine the best sequence of actions that constitute a process execution in order to maximize the desired KPI. In this paper, we address the problem of learning optimal behavioral policies by proposing an AI-based approach that learns an optimal policy directly from historical process executions using Reinforcement Learning (RL), with the goal of recommending the best actions to optimize a KPI of interest. To this end, we employ two distinct RL techniques. The first is a classical, model-based approach that extends previous work by the authors, overcoming its limitations, by constructing a Markov Decision Process (MDP) that captures the process behavior. The second is a model-free technique based on offline Deep RL, a rapidly advancing family of methods that have demonstrated strong performance across a variety of domains. Differently from state of the art work, we aim at building methods that minimize the usage of domain knowledge on the scenario at hand, and learn optimal policies directly from historical event data. In this way, we investigate whether, given a relevant KPI for the process under analysis, it is possible not only to learn when to apply an intervention, but also to discover which interventions are effective directly from data. Moreover, we aim at targeting complex scenarios, such as the ones modeling an interplay with customers or external actors, in which the process owner (and therefore the behavioral policy) may control only part of the process activities. Concerning the evaluation, we adopt an approach already used in the evaluation of pre-defined interventions of exploiting Business Process Simulation (BPS), but we adapt and customize it to the task of evaluating optimal behavioral policies. In particular, we build a data-driven BPS environment, to evaluate the discovered policies. Our results show that both methods consistently improve the targeted KPI with similar effectiveness, with the model-based approach outperforming offline Deep RL in terms of computational efficiency.
The field of process mining offers a range of techniques for evidence-based improvement of business processes. The quality of the process data used, as stored in so-called event logs, is paramount to the reliability and usefulness of the process mining outcomes. Due to the increased uptake, at scale and for more complex types of applications, the field of process mining has evolved and event logs now need to be object-centric rather than event-centric. To understand and manage the quality problems that can occur in object-centric event logs a systematic approach is required that is different from past investigations into event-centric logs. To this end, we adopt a pattern-based approach, a tried and tested method to characterise problems that are otherwise hard to capture. A new collection of patterns is presented for object-centric logs, where each pattern captures the nature of the problem, how its manifestation can be detected, and how the problem can be remedied. The pattern collection is validated through a multi-prong approach, i.e., evidence-based, literature-based, empirical, and user-based (with the process mining community). The results show that these patterns are perceived as important to identify and that they do occur in practical settings.
Temporal Knowledge Graphs (TKGs) incorporate temporal dimensions to accurately reflect evolving real-world knowledge. However, their inherent incompleteness necessitates Temporal Knowledge Graph Completion (TKGC) to predict missing facts. Existing tensor decomposition-based completion methods usually embed facts independently into a latent space, lacking modeling of explicit semantic correlations between related facts across different timestamps. Moreover, the semantics of the same entity or relation remain identical across different facts, leading to a lack of flexibility in semantic expression. To overcome these limitations, this study proposes a semantic collaborative adjustment strategy, which effectively integrates global semantic associations with local contexts through semantic modeling, enables semantic adjustment under temporal perspective, and supports integration with existing tensor factorization methods. This strategy captures semantic correlation across temporal facts by applying consistent semantic adjustment factors to facts sharing the same entity-relation combinations, while adapting the semantics of entities and relations based on specific contextual information. Furthermore, to enhance the quality of embedding representations, we introduce dropout-augmented contrastive learning as a complement to existing loss functions. By utilizing InfoNCE loss to maximize the similarity between positive pairs while minimizing the similarity with negative samples, the model is enabled to learn more discriminative entity representations. Experiments on the benchmark datasets show that the proposed strategy significantly improves performance across multiple metrics.
Contemporary information systems operate in heterogeneous and continuously evolving data environments, where representation choices and structural redesign decisions strongly influence system behavior. Existing benchmarking approaches, however, rely mostly on static datasets and fixed schemas, providing limited support for analyzing architectural trade-offs or guiding evolution in multi-model settings.This paper introduces TransforMMer, a framework for evolution-aware and representation-aware benchmark engineering in heterogeneous information systems. The approach treats benchmark construction as a systematic design process: starting from raw data, inferring structure, refining it conceptually, and generating comparable dataset variants across relational, document, and graph systems. The framework is grounded in a unified representation that enables explicit modeling of schemas and cross-model mappings and supports reproducible transformations across alternative representations.We position benchmarking as a system-design tool for evaluating architectural and representation-level decisions in evolving information systems, rather than as a static comparison of database engines. Through controlled benchmark construction scenarios on real-world datasets, we demonstrate how structural redesign steps—such as embedding, enrichment, and hybrid partitioning—affect observed query costs across systems. The results show that performance differences emerge primarily from the interaction between workload and representation design.By enabling systematic generation of structurally distinct yet semantically aligned dataset variants, the proposed approach connects conceptual data modeling with empirical system evaluation and supports reproducible, evolution-aware analysis of heterogeneous information systems.
Traditional Big Data systems operate in a passive fashion, responding to queries only when they are issued. Such an approach is increasingly inadequate for modern applications in which users demand continuous updates on newly arriving data, exploration of inter-data relationships, and enrichment with external context. Big Active Data (BAD) frameworks have been introduced to address this need by supporting large-scale data subscriptions and active analytics. However, as data volumes and subscriber populations grow, BAD platforms face fundamental scalability challenges, including redundant processing of similar subscriptions, unnecessary evaluation of irrelevant data, and delayed filtering of qualifying records. We present three optimizations that together enhance the scalability and efficiency of BAD: (i) subscription aggregation, which reduces redundant computation and communication; (ii) query plan modifications, which incorporate user preferences earlier in execution; and (iii) the BAD index, which performs early result filtering by indexing only records that satisfy fixed predicates. A comprehensive experimental evaluation, using both synthetic and real-world Twitter datasets on clusters of up to eight nodes, demonstrates substantial performance gains in execution time, broker efficiency, and scalability. The results establish BAD as an effective end-to-end framework for delivering continuous and enriched insights at Big Data scale.
