Tools designed to extract main content from web pages require thorough evaluation, yet existing benchmarks disproportionately focus on English-language datasets. Consequently, previous studies have shown that while these extractors are well-optimized for English, their effectiveness partially or entirely diminishes in other languages. This study reproduces and extends recent benchmarks by incorporating multilingual datasets as a key factor. We analyze extractor performance across five languages-Greek, English, Polish, Russian, and Chinese-highlighting the need to adapt extraction models to linguistic variations. Our results show that while some extractors maintain stable performance, others suffer significant drops in precision and recall on non-English or structurally irregular pages.

The pioneering work of Belohlavek et al. established a compelling connection between Boolean matrix factorization (BMF) and formal concept analysis (FCA), demonstrating that formal concepts serve as optimal factors for decomposing binary matrices. However, identifying the size-optimal decomposition remains an NP-hard problem, posing significant computational challenges. In this paper, we present a novel reformulation of the Boolean rank computation problem using hypergraph theory. Specifically, we show that the Boolean rank of a matrix corresponds to the size of the minimum transversal of the hypergraph constructed from the intervals of its formal concepts. This reformulation provides a theoretical foundation for understanding the structure of optimal factorizations and offers a new perspective on the problem. To validate our approach, we conducted an extensive experimental study to evaluate the characteristics of the solutions computed by our algorithm. The results demonstrated that our method not only achieved optimal factorizations but also exhibited favorable properties in terms of stability and separation.

The reliability of sales forecasting is critical for an industrial decision support system dedicated to raw material retailers. However, it turned difficult to train and maintain a custom model dedicated to each of the numerous references. For every reference, it would be needed to select the most accurate algorithm together with its relevant features, then, to exhaustively test every relevant combination of parameters. This was the reason why we explored an approach based on auto parametrization of well-known predictive models, while adding specific seasonal features. From our experiments, the Dynamic Harmonic Regression (DHR) based on ARMA stood out as being the most effective model for popular products: it reached a fair accuracy while requiring a reasonable cost to train. However, when it came to more volatile products, a simple prediction like the average sales per week over a year often performed the best. Thus, YearlyMean saved computational resources that could then be used to exhaustively train DHR or LSTM models on some company key products, leading to a potential improvement of their forecasts. Then, one details the implementation of a smart computing machine learning process based on predictive scenarios that seek for a trade-off between the consumed resources and the predictive performances.

In recent years, big data has increasingly drawn the attention of the R&D community. With the advent of marine data, monitoring marine big data becomes a new trend that advocates for assessing human impact on marine data. Nevertheless, there is a lack of support for acoustic sounds classification in such environment, covering diverse data that can exist (i.e., fish sounds, human activities sounds and environmental sounds). In this paper, we cope with this gap by proposing a deep learning-based approach that enables to efficiently classify these acoustic sounds aiming at automating the support of marine sound analysis in big data architectures. A set of experiments have been conducted using a real marine dataset to demonstrate the feasibility and the effectiveness of our approach.

International audience

In data mining, enumerate the frequent or the closed patterns is often the first difficult task leading to the association rules discovery. The number of these patterns represents a great interest. The lower bound is known to be constant whereas the upper bound is exponential, but both situations correspond to pathological cases. For the first time, we give an average analysis of the number of frequent or closed patterns. Average analysis is often closer to real situations and gives more information about the role of the parameters. In this paper, two probabilistic models are studied: a Bernoulli and a Markovian. In both models and for large databases, we prove that the number of frequent patterns, for a fixed frequency threshold, is exponential in the number of items and polynomial in the number of transactions. On the other hand, for a proportional frequency threshold, the number of frequent patterns is polynomial in the number of items and does not involve the number of transactions. Finally, we prove in the Bernoulli model that the number of closed patterns, for a proportional frequency threshold, is polynomial in the number of items.