In the era of big data, there is a growing need to explore large, high-dimensional datasets in an unsupervised manner to identify clusters, detect outliers, and generate labels. Many existing tools and methods facilitate these tasks through visual exploration, but there remains a significant gap between the 'zoomed-in' view, which focuses on individual objects using tables and bar plots, and the 'zoomed-out' view, where the entire dataset is visualized in two dimensions using dimensionality reduction techniques like PCA or t-SNE. To bridge this gap, we propose VISAnt, a visual data analytics tool that employs Chernoff faces visualization. Alongside the standard two-dimensional dataset projection followed by a scatterplot, VISAnt represents groups of adjacent objects using Chernoff face glyphs, while allowing users to zoom in on individual instances within these groups. This approach not only aids in revealing local patterns along with the view of the whole structure, but also facilitates comparing the efficiency of different dimensionality reduction methods in a given dataset. We demonstrate the utility of VISAnt in a real-world case study involving patients with multiple sclerosis.
Many modern search domains comprise high-dimensional vectors of floating point numbers derived from neural networks, in the form of embeddings. Typical embeddings range in size from hundreds to thousands of dimensions, making the size of the embeddings, and the speed of comparison, a significant issue. Quantisation is a class of mechanism which replaces the floating point values with a smaller representation, for example a short integer. This gives an approximation of the embedding space in return for a smaller data representation and a faster comparison function. Here we take this idea almost to its extreme: we show how vectors of arbitrary-precision floating point values can be replaced by vectors whose elements are drawn from the set {-1, 0, 1}. This yields very significant savings in space and metric evaluation cost, while maintaining a strong correlation for similarity measurements. This is achieved by way of a class of convex polytopes which exist in the high-dimensional space. In this article we give an outline description of these objects, and show how they can be used for the basis of such radical quantisation while maintaining a surprising degree of accuracy.
Many applications need to quantify the similarity between two clusterings of a data set. For example, one good way to assess the quality of a clustering method is to compare its output to a known ground truth. This paper considers scenarios where it is desirable or even necessary to estimate the similarity between two clusterings. Dozens of clustering similarity measures have been invented, yet most of them are based on pair counting or on concepts of information theory. We investigate how to estimate both kinds of measures accurately and quickly, given a random sample of data point pairs or of individual data points. One of our main contributions is the mathematical analysis of the bias and variance of these estimators, in terms of the sample size t. They either are unbiased or have a bias of 𝒪(1/t) , and have a variance of 𝒪(1/t) . Our results cover virtually every known pair-counting index, and some of the most popular information-theoretic measures. Regarding computational complexity, the estimates can be obtained in 𝒪(t) time and space by using appropriate data structures. Sometimes these costs include an additional term of order c· c' , where c and c' are the respective number of clusters in the two clusterings being compared. Preliminary experimental results agree well with our theoretical findings.
Understanding the nature of hand-drawn sketches is challenging due to the wide variation in their creation. Federico et al. [10] demonstrated that recognizing complex structural patterns enhances both sketch recognition and generation. Building on this foundation, we explore how the extracted features can also be leveraged for hand-drawn sketch retrieval. Inthiswork, we extend ViSketch-GPT, a multi-scale context extraction model originally designed for classification and generation, to the task of retrieval. The model's ability to capture intricate details at multiple scales allows it to learn highly discriminative representations, making it well-suited for retrieval applications. Through extensive experiments on the QuickDraw and TU-Berlin datasets, we show that ViSketch-GPT surpasses state-of-the-art methods in sketch retrieval, achieving substantial improvements across multiple evaluation metrics. Our results show that the extracted feature representations, originally designed for classification and generation, are also highly effective for retrieval tasks. This highlights ViSketch-GPT as a versatile and high-powerful framework for various applications in computer vision and sketch analysis.
Measuring distance and similarity is one of the most fundamental problems in data science and statistics. Proximity functions are used almost everywhere, from clustering problems to outlier detection, database systems, and data privacy. One important challenge is how to handle heterogeneous data, which makes up a large part of real-world tabular data. Euclidean distance with dummy variables is only half a solution, and Gower's similarity, which is usually the go-to choice for similarity estimation in heterogeneous data, also lacks stability when it comes to combinations of categorical variables. In this paper, we propose Resample Exposure, as a novel heterogeneous measure that effectively captures the similarity of records based on resampling probability and distance with traversal penalty. Resample exposure integrates well with various distance- or similarity-based tasks; we show that the measure is competitive for nearest neighbour classification and viable for partitioning around medoids, making resample exposure a relevant measure for heterogeneous data applications.
Both 2025 SISAP Challenges require participants to develop memory-efficient indexing solutions. Proposed solutions are run in a Linux container limited to 8 virtual CPUs and 16 GiB of memory, where virtual memory must not be enabled. Challenge 1 is to query a data set of 24 million items, and Challenge 2 is to build a near-neighbour table of 3 million items, making the memory restriction a significant limitation. We have submitted solutions to both challenges using our newly-invented technique of 2-bit quantisation based on Equi-Voronoi Polytopes (EVPs). This gives two very clear advantages: the data is compressed to 2 bits per vector element, therefore requiring only around 6 b_2sp , which is highly parallelisable on SIMD processors. The compression maintains sufficient accuracy to give reasonable results over the original Euclidean space. For both challenges we build a near-neighbour (NN) table using the ultra-quantised data and the b_2sp metric. This is built using a variant of the NN-Descent algorithm, which has the advantage of using fixed-size memory. The combination of these techniques along with a parallel, lock-free Rust implementation allows us to build a reasonably accurate near-neighbour table very quickly. For Challenge 1, a 10NN table is constructed within the allowed memory in around 50 min, and for Challenge 2 the 15NN table is constructed in around 10 min.
In modern data science, high-dimensional datasets with heterogeneous features, outliers and distributed storage are increasingly common. Recent research suggests that alternative distance measures, such as cosine similarity in text mining or embeddings, can outperform traditional metric distances in similarity search tasks. However, most existing exact or approximate methods are designed for metric spaces or centralised environments, limiting their applicability. To bridge this gap, we propose PDASC (Parametrizable Distributed Approximate Similarity Search with Clustering), a distributed indexing algorithm specifically designed for Approximate Nearest Neighbour (ANN) search with arbitrary distance functions. It integrates a lightweight, data-driven pruning strategy based on the empirical cumulative distribution function (ECDF) of distances, combined with local neighbourhood insights from the index structure. This approach enables effective pruning even when global coordination is unfeasible and geometric assumptions do not hold. This project's broader goal is to explore the impact of using diverse dissimilarity measures in ANN search. As a long-term direction, we aim to investigate the hypothesis that combining multiple dissimilarity representations can yield significantly better performance than relying on a single distance, especially in problems where no clear optimal distance function exists or when each dissimilarity representation emphasises different features of the data.
We present a solution for Task 2 of the SISAP Indexing Challenge 2025. The task consists of building the k-NN graph (self-similarity join) under limited memory and storage resources. Our solution is based on an approximate algorithm called Root Join, which we combine with some pre-processing steps to improve its performance with large high-dimensional data. For the specific task, we require building the k-NN graph for k = 15 with vectors of 384-D and the dataset has a size of approximately 3 million vectors. Our solution is focused on working under the restricted execution conditions of the Challenge, which consists of: a Linux container with 8 virtual CPUs, 16GB of RAM, and a time limit of 12 h.
Relevance feedback is a well-established approach to refine search results based on user input, but its comparative evaluation across different methods remains limited in practice. This demonstration paper introduces an interactive platform that supports and compares four relevance feedback methods-Rocchio, PicHunter, Polyadic Search, and SVM-based active learning-under consistent conditions. The primary goal is to enhance the understanding of how different relevance feedback methods affect retrieval performance from both a technical and user-centric perspective. The source code is available at https://github.com/francescascotti16/Demo-Relevance-Feedback, while the demonstration can be found at http://relevance-feedback.isti.cnr.it/.
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.
As a fundamental problem of metric-space data processing, pivot selection determines pivots, or reference points, for given metric spaces, so that each data element can be represented by its distances to pivots. Commonly, one first estimates the intrinsic dimension of data and uses it as the number of pivots, and then selects pivots according to some heuristics. Existing methods estimating intrinsic dimension are usually inaccurate or unstable. We propose a variance-based idea to use the position of the elbow in the sequence of particular variance-related statistic to estimate the intrinsic dimension. Further, we propose to incrementally select the next pivot with the maximal variance to other data after removing the covariance to other pivots by Gram–Schmidt orthogonalization. Experimental results on a comprehensive dataset show that our method estimates the intrinsic dimensions accurately and stably. Further, our pivot selection algorithm significantly outperforms typical existing algorithms in almost all the cases, by up to 70
We present a theoretical and empirical study of bilinear similarity functions in neural IR, showing that they are strictly more expressive than dot-product and weighted dot-product (WDP) models under fixed embeddings. We prove this separation formally and illustrate it with the Structured Agreement Ranking Task, where a simple rank-2 bilinear model achieves 100 × gain. These results offer a principled and practical case for using low-rank bilinear models in dense retrieval. Code: https://github.com/shubham526/bilinear-projection-theory .
This paper presents a practical implementation of a fast and memory-efficient approximate nearest neighbor (ANN) search system based on a two-stage filtering strategy using short and long quantized projections. The proposed system builds on a method recently developed by our group and is tailored for Task 1 of the SISAP 2025 Indexing Challenge, utilizing publicly available datasets and evaluation protocols. The indexing method, referred to as double filtering, usestwo typesof quantized dimension-reduction projections: short binary sketches derived by 1-bit quantization, and longer projections for refined filtering. Both are obtained by applying dimension reduction to the original high-dimensional vectors and quantizing the results. To meet the challenge's hardware constraints (16 GB RAM and 8 CPU cores), we tuned parameters such as sketch width and projection dimensionality. Unlike systems optimized for batch processing and throughput, our method operates in an online setting, where each query is processed sequentially and independently. We focus on per-query response latency, measuring average time and variability, which makes the system well-suited for real-time or interactive applications under resource limitations.
Vector search systems typically rely on fixed indexing methods and standard similarity measures, such as Euclidean distance, assuming a universal notion of semantic similarity. This assumption overlooks the subjective and context-dependent nature of human similarity perception. Prior approaches proposed personalised querying via Mahalanobis distance parameterised by learned user-profile matrices, but did not address strategies for acquiring real user relevance feedback. We present an end-to-end pipeline designed to integrate semantic feedback directly from user interactions. We explore three feedback strategies, evaluating them across ten established metric learning models. Through an empirical analysis, we assess how these strategies affect key performance indicators, such as retrieval precision. We also outline three practical operational scenarios – balancing precision and computational costs, prioritising computational efficiency, and optimising for precision – and provide concrete model and system configuration recommendations. Experiments show that the recommended setups make search efficient while improving retrieval precision, showing the practicality of the approach.
Entity matching, also known as entity resolution or record linkage, is a critical process in data integration, wherein the objective is to identify and reconcile different records that correspond to the same real-world entity across heterogeneous and often noisy data sources. Traditional approaches often rely on structured schemas or attribute alignment, limiting their applicability in contexts where entity descriptions are unstructured or inconsistent. We present PUFFME (Probabilistic Unkeyed Feature Fusion for Matching Entities), a novel method designed to operate directly on unstructured, semantic-free, entity attributes represented as bag-of-words. PUFFME models similarity using probabilistic distributions over textual features, enabling robust and interpretable comparisons between entities without requiring schema alignment or manual feature engineering. Evaluated on standard entity matching benchmarks, PUFFME achieves state-of-the-art performance on some datasets, significantly surpassing established baselines in others. These results highlight PUFFME's effectiveness in tackling entity matching tasks under minimal assumptions and with high accuracy, offering a practical alternative for data integration tasks where traditional methods struggle.
Dimensionality reduction (DR) methods are often used to unveil important information from a dataset. Most advanced DR algorithms are non-linear, presenting better performance in the result at the cost of a lower interpretability of whether the reduced dataset resembles the original one. While new methods for dimensionality reduction arise, UMAP is still more used than any other, for its great performance and solid theoretical foundations. As an embedding operation, DR methods induce a loss of information. Evaluation measures can be used in order to quantify the (local) faithfulness of the reduction. In this paper we propose using fuzzy simplicial sets (the cornerstone of UMAP) to develop evaluation measures. We demonstrate the usefulness of these measures in detecting unfaithfulness by revealing structural distortions, and identifying when the reduction falls below the intrinsic dimensionality.
While information retrieval (IR) systems have made significant strides in high-resource languages, their application to low-resource languages like Vietnamese remains a challenge. This is primarily due to data scarcity, limited parallel corpora, and the absence of standardised datasets. Vietnamese also presents unique difficulties, including tonal variation, compound word formations, and ambiguous segmentation. These challenges hinder the effectiveness of traditional retrieval approaches. In this work, we propose a retrieval framework tailored specifically for Vietnamese. The framework combines two key innovations: (1) a contextualised late-interaction retriever, which computes fine-grained token-level semantic similarity and improves retrieval accuracy through contrastive alignment with minimal bilingual data; and (2) a knowledge-distilled cross-encoder reranker, which transfers semantic knowledge from a larger teacher model to a smaller, more efficient student model, preserving accuracy while reducing inference time. This enables fast inference while maintaining strong semantic discrimination. We evaluate our system on multiple Vietnamese retrieval benchmarks spanning news, Wikipedia, and special domain (legal) documents. The framework consistently achieves high retrieval quality, with recall@100 exceeding 90
Current exact single-linkage clustering algorithms have asymptotically quadratic complexity. We present algorithms for approximate single-linkage clustering with empirically near-linear scalability. We explore both graph index-based incremental nearest neighbor search and an iterative exploration scheme on the graph index approximating the MST of the reachability graph similar to Kruskal. As graph index, we use both the bottom layer and a combination of all layers of an HNSW as a stand-in for connected search graphs. We provide experiments comparing the clusterings to baselines such as exact single linkage implementation and an algorithm using metric tree-based searchers. We explore the impact of the HNSW hyperparameters on the performance in terms of running time and clustering quality and evaluate the empirical asymptotic complexity.
SCAPE is a novel password-guessing method that extends Probabilistic Context-Free Grammar (PCFG) based approaches by introducing context awareness through word embeddings and similarity search. Evaluated on real-world datasets, including RockYou and various forum leaks, SCAPE consistently outperforms state-of-the-art methods. Our results show that combining classic NLP techniques with semantic similarity search is a powerful and efficient strategy for password guessing in cybersecurity applications.
The SWOT mission opened a new chapter in understanding ocean dynamics, yielding constantly growing amounts of high-resolution Sea Surface Height (SSH) observations. However, this extremely valuable data source is still not fully exploited by researchers, due to technical reasons and novelty. In particular, long orbit cycles and KaRIn altimeter's limited spatial coverage impose challenges for eddy detection and tracking using SWOT. Conversely, traditional satellite altimeters served as the backbone of mesoscale research for over 30 years, providing wide spatio-temporal coverage. The integration of such satellites and SWOT could offer possible solutions to some of the aforementioned challenges. We aim to provide a system capable of overcoming the limitations of KaRIn altimetry data, leveraging the integration between SWOT and conventional SSH altimetry. Adopting an information retrieval approach, we identify and index high-resolution eddy signals similar to a low-resolution counterpart. To do so, the Eddy Hunter System (EHS) integrates the Mesoscale Eddy Trajectory Atlas (META) and SWOT Level-3 products, by leveraging the spatio-temporal similarities in the SSH signals. The EHS extracts a spatio-temporal Region Of Interest (ROI) from every eddy observation present in the META and computes the SWOT passages that could possibly contain related signals. The introduction of "Marine Data Science" topics in the similarity-search community can provide new problems and challenges regarding information retrieval and data mining. The deployment of our system could foster advancements both in oceanographic and applied ML/AI research, providing a powerful data mining tool to build high-resolution eddy repositories and unveil previously unseen ocean phenomena.
