arXiv subjects

Paolo Ferragina

Publications and source records attributed to Paolo Ferragina.

At least 19 recordsLinked to original sources

arXiv2026

Diffuse to Compress: Leveraging Diffusion LMs for Lossless Compression

We study the problem of lossless text compression, motivated by the rapid growth in the collection and storage of digital textual data - including plain text, source code, and structured formats such as XML - and by recent advances in neural language model-based compression. In particular, recent LLM-based approaches, whether built on symbol-ranking pipelines or paired with a statistical compressor, have demonstrated compression ratios significantly superior to general-purpose compressors such as zstd, gzip, or bzip on text and code. However, these neural approaches suffer from severe throughput limitations, making them not yet practically usable. For the first time in the context of lossless neural text compression, we introduce Diffusion Language Models (DLMs) as an alternative inference paradigm to autoregressive LLM-based approaches. We argue that replacing autoregressive LLMs with DLMs within the same compression framework could overcome the throughput bottleneck caused by their one-symbol-per-step limitation. However, achieving these improvements requires addressing algorithmic challenges introduced by applying DLMs to lossless compression, where the architecture allows the number and positions of symbols encoded at each forward pass to be decided independently. We design efficient and effective strategies to solve these challenges and evaluate them experimentally against LLM-based and general-purpose compressors on enwik8, a well-established textual benchmark. Our results show that the newly proposed DLM-based framework advances the state of the art in lossless text compression. Moreover, as DLMs are still a relatively young paradigm, recent advances toward increasingly capable and efficient models suggest substantial room for further improvements.

cs.CL↗

arXiv2026

Extended Depth-First Representations of $k^2$-trees

In this paper, we study static, computation-friendly, lossless compression formats for graphs, focusing on memory locality and operational efficiency of $k^2$-trees. We observe that their traditional level-wise layouts suffer from poor cache performance due to weak locality, especially in operations such as matrix-vector and matrix-matrix operations. To address this limitation, we propose four depth-first representations of $k^2$-trees: a plain depth-first layout (EDF-1), a balanced-parenthesis representation (BP), and their compressed variants (CEDF and CBP). We further introduce a linear-time compression method based on suffix and LCP arrays to identify and compress identical subtrees. We experimentally evaluate the execution time, the disk space, and the peak-memory usage of our approaches against classical level-wise $k^2$-trees and DFUDS-based representations across two real and one synthetic dataset (i.e., Web Graphs, Wikidata, and random adjacency matrices) over the above linear-algebra operations. Results show that our depth-first layouts are competitive and often superior than known approaches: CEDF achieves the best compression in most settings, EDF-1 and CEDF reduce the peak memory usage consistently, and performance varies by workload, with different layouts excelling in different operations and data regimes. Overall, this work demonstrates that depth-first layouts of $k^2$-trees provide a practical and efficient alternative to traditional layouts, improving both compression and computational performance in matrix operations.

cs.DS↗

arXiv2026

LLM-based Source Code Compression via Thresholded Symbol Ranking

We study the problem of lossless compression of source code, motivated by the storage demands of large-scale software archives, such as Software Heritage (https://www.softwareheritage.org/). General-purpose compressors (e.g., zstd, bzip2) offer a good trade-off between compression ratio and speed, but fail to exploit all special regularities inherent in source code. Recent approaches leverage Large Language Models (LLMs) within Shannon's symbol-ranking framework, relying on a scheme in which the predicted rank can grow arbitrarily. While effective at reducing space, this setting incurs significant throughput degradation, and leaves open the question whether it is necessary to explicitly encode all ranks. In this work, we introduce LLM-based compressors deploying two novel symbol-ranking variants that bound predictions to the top-$T$ ranks ($T=1$ or $63$), with out-of-threshold symbols stored as exceptions and compressed jointly with the rank stream via general-purpose compressors. We conduct the first large-scale evaluation of LLM-based source code compression across 30 LLMs, including general-domain, code-specialized, and quantized models. Our $T$-bounded approach outperforms prior LLM-based compressors both in compression ratio (up to 37% relative improvement) and compression throughput (40% faster). Compared to general-purpose compressors (e.g., zstd, bzip2), we obtain up to 82% relative compression gain but at a lower speed, thus offering a new trade-off point in the compression-speed spectrum. We also show that these gains are stronger on source code than on natural language, suggesting an interesting indication, namely that source code exposes regularities captured by LLMs but missed by general-purpose exact-match-based compressors. We conclude by commenting on open problems that offer theoretical and practical avenues of research.

cs.IT↗

arXiv2026

Recall Before Rerank: Benchmarking Deep Learning Models for Large-Scale Code-to-Code Retrieval

Semantic code search and clone detection are essential for software development, maintenance, and reuse. This paper evaluates the effectiveness, efficiency, and scalability of contemporary deep learning models for first-stage recall in large-scale code-to-code search engines. Benchmarking across multiple programming languages and datasets reveals critical limits in the precision and scalability of these models on Terabyte-scale source-code collections. We present LLM-based code normalisation and query-rewriting schemes that yield significant gains in precision for lower-performing models. Our results question the sustainability of resource-constrained deployment and the assumed robustness of current code-specialised LLMs across datasets. We conclude with actionable insights for building scalable, efficient code-retrieval systems.

cs.SE↗

arXiv2026

The Energy-Throughput Trade-off in Lossless-Compressed Source Code Storage

Retrieving data from large-scale source code archives is vital for AI training, neural-based software analysis, and information retrieval, to cite a few. This paper studies and experiments with the design of a compressed key-value store for the indexing of large-scale source code datasets, evaluating its trade-off among three primary computational resources: (compressed) space occupancy, time, and energy efficiency. Extensive experiments on a national high-performance computing infrastructure demonstrate that different compression configurations yield distinct trade-offs, with high compression ratios and order-of-magnitude gains in retrieval throughput and energy efficiency. We also study data parallelism and show that, while it significantly improves speed, scaling energy efficiency is more difficult, reflecting the known non-energy-proportionality of modern hardware and challenging the assumption of a direct time-energy correlation. This work streamlines automation in energy-aware configuration tuning and standardized green benchmarking deployable in CI/CD pipelines, thus empowering system architects with a spectrum of Pareto-optimal energy-compression-throughput trade-offs and actionable guidelines for building sustainable, efficient storage backends for massive open-source code archival.

cs.DS↗

arXiv2025

Compressibility Measures and Succinct Data Structures for Piecewise Linear Approximations

We study the problem of deriving compressibility measures for Piecewise Linear Approximations (PLAs), i.e., error-bounded approximations of a set of two-dimensional increasing data points using a sequence of segments. Such approximations are widely used tools in implementing many learned data structures, which mix learning models with traditional algorithmic design blocks to exploit regularities in the underlying data distribution, providing novel and effective space-time trade-offs. We introduce the first lower bounds to the cost of storing PLAs in two settings, namely compression and indexing. We then compare these compressibility measures to known data structures, and show that they are asymptotically optimal up to a constant factor from the space lower bounds. Finally, we design the first data structures for the aforementioned settings that achieve the space lower bounds plus small additive terms, which turn out to be succinct in most practical cases. Our data structures support the efficient retrieval and evaluation of a segment in the (compressed) PLA for a given $x$-value, which is a core operation in any learned data structure relying on PLAs. As a result, our paper offers the first theoretical analysis of the maximum compressibility achievable by PLA-based learned data structures, and provides novel storage schemes for PLAs offering strong theoretical guarantees while also suggesting simple and efficient practical implementations.

cs.DS↗

arXiv2024

Learned Compression of Nonlinear Time Series With Random Access

Time series play a crucial role in many fields, including finance, healthcare, industry, and environmental monitoring. The storage and retrieval of time series can be challenging due to their unstoppable growth. In fact, these applications often sacrifice precious historical data to make room for new data. General-purpose compressors can mitigate this problem with their good compression ratios, but they lack efficient random access on compressed data, thus preventing real-time analyses. Ad-hoc streaming solutions, instead, typically optimise only for compression and decompression speed, while giving up compression effectiveness and random access functionality. Furthermore, all these methods lack awareness of certain special regularities of time series, whose trends over time can often be described by some linear and nonlinear functions. To address these issues, we introduce NeaTS, a randomly-accessible compression scheme that approximates the time series with a sequence of nonlinear functions of different kinds and shapes, carefully selected and placed by a partitioning algorithm to minimise the space. The approximation residuals are bounded, which allows storing them in little space and thus recovering the original data losslessly, or simply discarding them to obtain a lossy time series representation with maximum error guarantees. Our experiments show that NeaTS improves the compression ratio of the state-of-the-art lossy compressors that use linear or nonlinear functions (or both) by up to 14%. Compared to lossless compressors, NeaTS emerges as the only approach to date providing, simultaneously, compression ratios close to or better than the best existing compressors, a much faster decompression speed, and orders of magnitude more efficient random access, thus enabling the storage and real-time analysis of massive and ever-growing amounts of (historical) time series data.

cs.LG↗

arXiv2024

Toward Greener Matrix Operations by Lossless Compressed Formats

Sparse matrix-vector multiplication (SpMV) is a fundamental operation in machine learning, scientific computing, and graph algorithms. In this paper, we investigate the space, time, and energy efficiency of SpMV using various compressed formats for large sparse matrices, focusing specifically on Boolean matrices and real-valued vectors. Through extensive analysis and experiments conducted on server and edge devices, we found that different matrix compression formats offer distinct trade-offs among space usage, execution time, and energy consumption. Notably, by employing the appropriate compressed format, we can reduce energy consumption by an order of magnitude on both server and single-board computers. Furthermore, our experiments indicate that while data parallelism can enhance execution speed and energy efficiency, achieving simultaneous time and energy efficiency presents partially distinct challenges. Specifically, we show that for certain compression schemes, the optimal degree of parallelism for time does not align with that for energy, thereby challenging prevailing assumptions about a straightforward linear correlation between execution time and energy consumption. Our results have significant implications for software engineers in all domains where SpMV operations are prevalent. They also suggest that similar studies exploring the trade-offs between time, space, and energy for other compressed data structures can substantially contribute to designing more energy-efficient software components.

cs.DS↗

arXiv2024

Balanced Learned Sort: a new learned model for fast and balanced item bucketing

This paper aims to better understand the strengths and limitations of adopting learned-based approaches in sequential sorting numerical data, via two main research steps. First, we study different learned models for distribution-based sorting, starting from some known ones (i.e., two-layer RMI or simple linear models) and then introducing some novel models that either improve the two-layer RMI or are fully new in their algorithmic structure thus resulting space efficient, monotonic, and very fast in building balanced buckets. We test those models over 11 synthetic datasets drawn from different distributions of 200M 64-bit floating-point items, so deriving hints about their ultimate performance and usefulness in designing a sorting algorithm. Based on these findings, we select and plug the best models from above in a new learned-based algorithmic scheme and devise three new sorters that we will test against other 6 sequential sorters (5 classic and 1 learned, known and new ones) over 33 datasets (11 synthetic and 22 real), whose size will be up to 800M items. Our experimental figures will show that our learned sorters achieve superior performance on 31 out of all 33 datasets (synthetic and real). In conclusion, these experimental results provide, on the one hand, a comprehensive answer to the main question: Which algorithmic structure for distribution-based sorting is suited to leverage a learned model in order to achieve efficient performance? and, on the other hand, they leave open several other research and engineering questions about the design of a highly performing sequential sorter that is robust over different input distributions.

cs.DS↗

arXiv2024

Grafite: Taming Adversarial Queries with Optimal Range Filters

Range filters allow checking whether a query range intersects a given set of keys with a chance of returning a false positive answer, thus generalising the functionality of Bloom filters from point to range queries. Existing practical range filters have addressed this problem heuristically, resulting in high false positive rates and query times when dealing with adversarial inputs, such as in the common scenario where queries are correlated with the keys. We introduce Grafite, a novel range filter that solves these issues with a simple design and clear theoretical guarantees that hold regardless of the input data and query distribution: given a fixed space budget of $B$ bits per key, the query time is $O(1)$, and the false positive probability is upper bounded by $\ell/2^{B-2}$, where $\ell$ is the query range size. Our experimental evaluation shows that Grafite is the only range filter to date to achieve robust and predictable false positive rates across all combinations of datasets, query workloads, and range sizes, while providing faster queries and construction times, and dominating all competitors in the case of correlated queries. As a further contribution, we introduce a very simple heuristic range filter whose performance on uncorrelated queries is very close to or better than the one achieved by the best heuristic range filters proposed in the literature so far.

cs.DS↗

arXiv2023

On nonlinear compression costs: when Shannon meets Rényi

Shannon entropy is the shortest average codeword length a lossless compressor can achieve by encoding i.i.d. symbols. However, there are cases in which the objective is to minimize the \textit{exponential} average codeword length, i.e. when the cost of encoding/decoding scales exponentially with the length of codewords. The optimum is reached by all strategies that map each symbol $x_i$ generated with probability $p_i$ into a codeword of length $\ell^{(q)}_D(i)=-\log_D\frac{p_i^q}{\sum_{j=1}^Np_j^q}$. This leads to the minimum exponential average codeword length, which equals the Rényi, rather than Shannon, entropy of the source distribution. We generalize the established Arithmetic Coding (AC) compressor to this framework. We analytically show that our generalized algorithm provides an exponential average length which is arbitrarily close to the Rényi entropy, if the symbols to encode are i.i.d.. We then apply our algorithm to both simulated (i.i.d. generated) and real (a piece of Wikipedia text) datasets. While, as expected, we find that the application to i.i.d. data confirms our analytical results, we also find that, when applied to the real dataset (composed by highly correlated symbols), our algorithm is still able to significantly reduce the exponential average codeword length with respect to the classical `Shannonian' one. Moreover, we provide another justification of the use of the exponential average: namely, we show that by minimizing the exponential average length it is possible to minimize the probability that codewords exceed a certain threshold length. This relation relies on the connection between the exponential average and the cumulant generating function of the source distribution, which is in turn related to the probability of large deviations. We test and confirm our results again on both simulated and real datasets.

cs.IT↗

arXiv2023

Learned Monotone Minimal Perfect Hashing

A Monotone Minimal Perfect Hash Function (MMPHF) constructed on a set S of keys is a function that maps each key in S to its rank. On keys not in S, the function returns an arbitrary value. Applications range from databases, search engines, data encryption, to pattern-matching algorithms. In this paper, we describe LeMonHash, a new technique for constructing MMPHFs for integers. The core idea of LeMonHash is surprisingly simple and effective: we learn a monotone mapping from keys to their rank via an error-bounded piecewise linear model (the PGM-index), and then we solve the collisions that might arise among keys mapping to the same rank estimate by associating small integers with them in a retrieval data structure (BuRR). On synthetic random datasets, LeMonHash needs 34% less space than the next larger competitor, while achieving about 16 times faster queries. On real-world datasets, the space usage is very close to or much better than the best competitors, while achieving up to 19 times faster queries than the next larger competitor. As far as the construction of LeMonHash is concerned, we get an improvement by a factor of up to 2, compared to the competitor with the next best space usage. We also investigate the case of keys being variable-length strings, introducing the so-called LeMonHash-VL: it needs space within 13% of the best competitors while achieving up to 3 times faster queries than the next larger competitor.

cs.DS↗

arXiv2022

Improving Matrix-vector Multiplication via Lossless Grammar-Compressed Matrices

As nowadays Machine Learning (ML) techniques are generating huge data collections, the problem of how to efficiently engineer their storage and operations is becoming of paramount importance. In this article we propose a new lossless compression scheme for real-valued matrices which achieves efficient performance in terms of compression ratio and time for linear-algebra operations. Experiments show that, as a compressor, our tool is clearly superior to gzip and it is usually within 20% of xz in terms of compression ratio. In addition, our compressed format supports matrix-vector multiplications in time and space proportional to the size of the compressed representation, unlike gzip and xz that require the full decompression of the compressed matrix. To our knowledge our lossless compressor is the first one achieving time and space complexities which match the theoretical limit expressed by the $k$-th order statistical entropy of the input. To achieve further time/space reductions, we propose column-reordering algorithms hinging on a novel column-similarity score. Our experiments on various data sets of ML matrices show that, with a modest preprocessing time, our column reordering can yield a further reduction of up to 16% in the peak memory usage during matrix-vector multiplication. Finally, we compare our proposal against the state-of-the-art Compressed Linear Algebra (CLA) approach showing that ours runs always at least twice faster (in a multi-thread setting) and achieves better compressed space occupancy for most of the tested data sets. This experimentally confirms the provably effective theoretical bounds we show for our compressed-matrix approach.

cs.DS↗

arXiv2021

An interactive dashboard for searching and comparing soccer performance scores

The performance of soccer players is one of most discussed aspects by many actors in the soccer industry: from supporters to journalists, from coaches to talent scouts. Unfortunately, the dashboards available online provide no effective way to compare the evolution of the performance of players or to find players behaving similarly on the field. This paper describes the design of a web dashboard that interacts via APIs with a performance evaluation algorithm and provides graphical tools that allow the user to perform many tasks, such as to search or compare players by age, role or trend of growth in their performance, find similar players based on their pitching behavior, change the algorithm's parameters to obtain customized performance scores. We also describe an example of how a talent scout can interact with the dashboard to find young, promising talents.

cs.HC↗

arXiv2020

Give more data, awareness and control to individual citizens, and they will help COVID-19 containment

The rapid dynamics of COVID-19 calls for quick and effective tracking of virus transmission chains and early detection of outbreaks, especially in the phase 2 of the pandemic, when lockdown and other restriction measures are progressively withdrawn, in order to avoid or minimize contagion resurgence. For this purpose, contact-tracing apps are being proposed for large scale adoption by many countries. A centralized approach, where data sensed by the app are all sent to a nation-wide server, raises concerns about citizens' privacy and needlessly strong digital surveillance, thus alerting us to the need to minimize personal data collection and avoiding location tracking. We advocate the conceptual advantage of a decentralized approach, where both contact and location data are collected exclusively in individual citizens' "personal data stores", to be shared separately and selectively, voluntarily, only when the citizen has tested positive for COVID-19, and with a privacy preserving level of granularity. This approach better protects the personal sphere of citizens and affords multiple benefits: it allows for detailed information gathering for infected people in a privacy-preserving fashion; and, in turn this enables both contact tracing, and, the early detection of outbreak hotspots on more finely-granulated geographic scale. Our recommendation is two-fold. First to extend existing decentralized architectures with a light touch, in order to manage the collection of location data locally on the device, and allow the user to share spatio-temporal aggregates - if and when they want, for specific aims - with health authorities, for instance. Second, we favour a longer-term pursuit of realizing a Personal Data Store vision, giving users the opportunity to contribute to collective good in the measure they want, enhancing self-awareness, and cultivating collective efforts for rebuilding society.

cs.CY↗

arXiv2019

The PGM-index: a multicriteria, compressed and learned approach to data indexing

The recent introduction of learned indexes has shaken the foundations of the decades-old field of indexing data structures. Combining, or even replacing, classic design elements such as B-tree nodes with machine learning models has proven to give outstanding improvements in the space footprint and time efficiency of data systems. However, these novel approaches are based on heuristics, thus they lack any guarantees both in their time and space requirements. We propose the Piecewise Geometric Model index (shortly, PGM-index), which achieves guaranteed I/O-optimality in query operations, learns an optimal number of linear models, and its peculiar recursive construction makes it a purely learned data structure, rather than a hybrid of traditional and learned indexes (such as RMI and FITing-tree). We show that the PGM-index improves the space of the FITing-tree by 63.3% and of the B-tree by more than four orders of magnitude, while achieving their same or even better query time efficiency. We complement this result by proposing three variants of the PGM-index. First, we design a compressed PGM-index that further reduces its space footprint by exploiting the repetitiveness at the level of the learned linear models it is composed of. Second, we design a PGM-index that adapts itself to the distribution of the queries, thus resulting in the first known distribution-aware learned index to date. Finally, given its flexibility in the offered space-time trade-offs, we propose the multicriteria PGM-index that efficiently auto-tune itself in a few seconds over hundreds of millions of keys to the possibly evolving space-time constraints imposed by the application of use. We remark to the reader that this paper is an extended and improved version of our previous paper titled "Superseding traditional indexes by orchestrating learning and geometry" (arXiv:1903.00507).

cs.DS↗

arXiv2019

WISER: A Semantic Approach for Expert Finding in Academia based on Entity Linking

We present WISER, a new semantic search engine for expert finding in academia. Our system is unsupervised and it jointly combines classical language modeling techniques, based on text evidences, with the Wikipedia Knowledge Graph, via entity linking. WISER indexes each academic author through a novel profiling technique which models her expertise with a small, labeled and weighted graph drawn from Wikipedia. Nodes in this graph are the Wikipedia entities mentioned in the author's publications, whereas the weighted edges express the semantic relatedness among these entities computed via textual and graph-based relatedness functions. Every node is also labeled with a relevance score which models the pertinence of the corresponding entity to author's expertise, and is computed by means of a proper random-walk calculation over that graph; and with a latent vector representation which is learned via entity and other kinds of structural embeddings derived from Wikipedia. At query time, experts are retrieved by combining classic document-centric approaches, which exploit the occurrences of query terms in the author's documents, with a novel set of profile-centric scoring strategies, which compute the semantic relatedness between the author's expertise and the query topic via the above graph-based profiles. The effectiveness of our system is established over a large-scale experimental test on a standard dataset for this task. We show that WISER achieves better performance than all the other competitors, thus proving the effectiveness of modelling author's profile via our "semantic" graph of entities. Finally, we comment on the use of WISER for indexing and profiling the whole research community within the University of Pisa, and its application to technology transfer in our University.

cs.IR↗

arXiv2019

SWAT: A System for Detecting Salient Wikipedia Entities in Texts

We study the problem of entity salience by proposing the design and implementation of SWAT, a system that identifies the salient Wikipedia entities occurring in an input document. SWAT consists of several modules that are able to detect and classify on-the-fly Wikipedia entities as salient or not, based on a large number of syntactic, semantic and latent features properly extracted via a supervised process which has been trained over millions of examples drawn from the New York Times corpus. The validation process is performed through a large experimental assessment, eventually showing that SWAT improves known solutions over all publicly available datasets. We release SWAT via an API that we describe and comment in the paper in order to ease its use in other software.

cs.IR↗