This study explores dataset complexity estimation and its relationship with classification performance. Widely-used measures like the k-Disagreeing Neighbors (kDN) are limited by their reliance on a fixed number of neighbors ( k ) and their discrete output, which can lead to coarse estimations. To address these issues, we introduce Dynamic Disagreeing Neighbors (DDN), a measure that incorporates dynamic, density-aware neighborhoods and distance-based weighting to provide a smoother, more flexible complexity estimation. We conduct a large-scale empirical study on 65 binary datasets, comparing DDN against kDN and other state-of-the-art measures, analyzing the impact of the k parameter and class imbalance on the alignment with classifier performance. Our results show that DDN provides a more robust and stable correspondence with performance. Furthermore, we demonstrate that in a predictive modeling task, a regression model using DDN as a feature is significantly more accurate at estimating dataset performance than models using other complexity measures. These findings not only deepen our understanding of complexity estimation but also pave the way for more informed classifier selection and data preprocessing strategies.

Recent advances in Natural Language Processing (NLP) are highly based on black-box score-based models that provide only final predictions along with their score. This opacity impedes the comprehension of their internal decision-making processes, complicating the identification of potential weaknesses. A powerful strategy for analyzing model vulnerabilities is the generation of text adversarial examples. These attacks introduce subtle text perturbations that cause victim models to make incorrect predictions while preserving the original semantic meaning. This paper presents a novel method for generating text adversarial examples through Large Language Models (LLMs). The proposed method uses the outstanding text generation capabilities of LLMs to modify the original input text at multiple granularities: character, word, and sentence level. First, sentence-level perturbations are introduced by generating paraphrases with an LLM instruction prompt. Next, further character- and word-level perturbations are introduced to words that most affect predictions using another set of LLM instruction prompts. In particular, vulnerable words are perturbed by replacing them with their synonyms or misspelled variants, or by inserting additional neutral words adjacent to them. Experiments were conducted to assess the proposal’s viability on two sentiment classification tasks: sentence-level reviews and full-length reviews. The proposal demonstrates an advantage over many well-known approaches based on LLMs. It preserves the original semantics to a similar extent, while increasing the deception of victim models by 29 − 85 % over the best-analyzed state-of-the-art methods.

Abstract Data augmentation is a widely used strategy to enhance the predictive power of machine learning (ML) models. This is the case of intent classification problems, where the end goal of a utterance needs to be categorized using text mining techniques. Nevertheless, recent augmentation methods based on general off-the-shelf Large Language Models (LLMs) have room for improvement. They can struggle to effectively capture the nuances associated with domain-specific scenarios. This paper presents a novel utterance augmentation method that uses LLMs and word embedding models to address the issue, particularly in domain-specific problems. The proposed method starts from a given reference set. Then, paraphrases are generated using LLMs to obtain new utterances that are semantically similar to the original ones, but with different word choices and syntax. Next, synonym replacement is accomplished using a previously trained domain-specific word embedding model. This entails the incorporation of relevant vocabulary to a particular topic into the final augmented dataset, effectively capturing the nuances of domain-specific problems. Experiments were conducted to assess the quality of the proposal in two intent classification problems related to financial trading compliance. The proposal has an advantage over many well-known approaches in the first problem, comprising eight reference utterances and 375 challenging examples in general language. In the second problem, with 13 reference utterances and 525 challenging examples in general and financial trading vocabulary, the proposal outperforms the best-analyzed state-of-the-art methods by up to 15%.

Medical professionals are often overwhelmed by the amount of patients they have to care for, leaving little time available to keep up to date in their respective specialities. They usually find it challenging to keep up with the vast amount of medical literature and identify the most relevant articles for their practice, especially those related to their patient’s specific conditions. Therefore, a system that proactively supports healthcare professionals in selecting relevant articles related to the characteristics of the patients is crucial. This paper presents Medical Expert Linguist for Evaluating Nosology and Diagnosis Information (MELENDI) to tackle this issue. It is a recommendation system that effectively and efficiently recommends pertinent medical articles to healthcare professionals based on their patients’ diagnoses. It combines a semantic similarity model generated using the content of discharge summaries, with a relevance estimator produced by analysing scientific publications. To test the system, 1 , 000 , 000 abstracts were obtained from PubMed and 10 discharge reports from ’Medical Information Mart for Intensive Care (MIMIC-III) were used. A group of 5 medical specialists has been involved in the system’s evaluation. These evaluations demonstrated good overall performance, supporting the implementation of the system in a real-world environment, such as a hospital information system .

Abstract In the scientific community, there is no single, objective, and precise metric for ranking the work of researchers based on their scientific merit. Most existing metrics are based on the number of publications associated with an author along with the number of citations received by those publications. However, there is no standard metric officially used to evaluate the researchers’ careers. In this paper, the Framework for Reputation Estimation of Scientific Authors (FRESA) to address this issue is depicted. It is a system able to estimate the reputation of a researcher focusing on the achieved publications. It calculates two indexes making use of the relevance and the novelty concepts in the scientific domain. The system can depict the scientific trajectories of the researchers through the proposed indexes to illustrate their evolution over time. FRESA uses web information sources and applies similarity measures, text mining techniques, and clustering algorithms to also rank and group the researchers. The presented work is experimental, rendering promising results.

Abstract Complexity measures aim to characterize the underlying complexity of supervised data. These measures tackle factors hindering the performance of Machine Learning (ML) classifiers like overlap, density, linearity, etc. The state-of-the-art has mainly focused on the dataset perspective of complexity, i.e., offering an estimation of the complexity of the whole dataset. Recently, the instance perspective has also been addressed. In this paper, the hostility measure, a complexity measure offering a multi-level (instance, class, and dataset) perspective of data complexity is proposed. The proposal is built by estimating the novel notion of hostility: the difficulty of correctly classifying a point, a class, or a whole dataset given their corresponding neighborhoods. The proposed measure is estimated at the instance level by applying the k-means algorithm in a recursive and hierarchical way, which allows to analyze how points from different classes are naturally grouped together across partitions. The instance information is aggregated to provide complexity knowledge at the class and the dataset levels. The validity of the proposal is evaluated through a variety of experiments dealing with the three perspectives and the corresponding comparative with the state-of-the-art measures. Throughout the experiments, the hostility measure has shown promising results and to be competitive, stable, and robust.

Similarity search based on a distance function in metric spaces is a fundamental problem for many applications. Queries for similar objects lead to the well-known machine learning task of nearest-neighbours identification. Many data indexing strategies, collectively known as Metric Access Methods (MAM), have been proposed to speed up these queries. Moreover, since exact approaches to solving similarity queries can be complex and time-consuming, alternative options have emerged to reduce query execution time, such as returning approximate results or resorting to distributed computing platforms. In this paper, we introduce MASK (Multilevel Approximate Similarity search with k-means), an unconventional application of the k-means algorithm as the foundation of a multilevel index structure for approximate similarity search suitable for metric spaces. We show that this method leverages inherent properties of k-means for this purpose, like representing high-density data areas with fewer prototypes. An implementation of this new indexing procedure is evaluated using a synthetic dataset and two real-world datasets in high-dimensional and high-sparsity spaces. Experimental tests show that MASK performs better than alternative algorithms for approximate similarity search. Results are promising and underpin the applicability of this novel indexing method in multiple domains.

Customer acquisition is an issue that continues to receive attention from companies worldwide.Various marketing campaigns using psychological methodologies have been designed to address this issue.However, once a campaign is launched, it is highly complicated to detect which sets of customers are most likely to purchase an offered product.This fact is key since it allows companies to focus their efforts on specific clients and discard others.Several selection techniques have been implemented, but most of them are usually very demanding in terms of time and human resources for the companies.Artificial Intelligence techniques appear to help to simplify the process.Thus, companies have started to use Machine Learning (ML) models trained to efficiently detect those clients with certain proneness to purchase.Toward this goal, this paper presents a novel purchase propensity detection ML system based on Sentiment Analysis techniques able to consider customer comments regarding the offered products.The tourist domain was selected for the case study, where the obtained product was successfully embedded in an initial prototype.

Abstract Several performance metrics are currently available to evaluate the performance of Machine Learning (ML) models in classification problems. ML models are usually assessed using a single measure because it facilitates the comparison between several models. However, there is no silver bullet since each performance metric emphasizes a different aspect of the classification. Thus, the choice depends on the particular requirements and characteristics of the problem. An additional problem arises in multi-class classification problems, since most of the well-known metrics are only directly applicable to binary classification problems. In this paper, we propose the General Performance Score (GPS) , a methodological approach to build performance metrics for binary and multi-class classification problems. The basic idea behind GPS is to combine a set of individual metrics, penalising low values in any of them. Thus, users can combine several performance metrics that are relevant in the particular problem based on their preferences obtaining a conservative combination. Different GPS -based performance metrics are compared with alternatives in classification problems using real and simulated datasets. The metrics built using the proposed method improve the stability and explainability of the usual performance metrics. Finally, the GPS brings benefits in both new research lines and practical usage, where performance metrics tailored for each particular problem are considered.

The scientific and business communities are proposing new authentication methods more robust than traditional solutions relying on a single security point such as passwords (i.e. “something you know”). User and Entity Behavior Analysis (UEBA) has postulated as an excellent solution to improve authentication systems by performing continuous authentication to extend the authentication process over time. UEBA is based on detecting anomalies in the intrinsic behaviour of each user or entity (i.e. it is based on “something you are/do”). This paper presents a method for performing continuous authentication using UEBA techniques that allows combining information from multiple sources at the feature level. This combination is achieved through a novel Symbolic Aggregate approximation (SAX) using Random Trees Embeddings for each information source, producing a sequence of symbols. Then, these sequences of symbols are combined into a single sequence using temporal information. The resulting sequence of symbols feeds a density-based clustering model that uses a distance based on DNA sequence alignment techniques to extract behavioural cores. Finally, new samples are compared against these cores to detect anomalies using a risk model that evaluates if a behaviour is anomalous (suspected user impersonation). The model has been extensively tested and evaluated against well-known state-of-the-art datasets.

User and Entity Behaviour Analytics (UEBA) mechanisms rely on statistical techniques and Machine Learning to determine when a significant deviation from patterns or trends established as a standard for users and entities is occurring. These mechanisms are beneficial within cybersecurity contexts because they allow managers and administrators to have early alerts warning about potential security incidents. This paper proposes the utilisation of UEBA to improve the security of Federated Identity Management (FIM) solutions. The proposed UEBA workflow allows Relying Parties within identity federations to build a session fingerprint characterising each user’s behaviour from available information. Furthermore, it enables anomaly detection based on this fingerprint, integrating raised alerts within current identity management specifications. The proposed workflow is validated and evaluated in a real use case based on a web chat application using OpenID Connect for identity management.

Voice transmission is no longer the main usage of mobile phones. Data transmissions, in particular Internet access, are very common actions that we might perform with these devices. However, the spectacular growth of the mobile data demand in 5 G mobile communication systems leads to a reduction of the resources assigned to each device. Therefore, to avoid situations in which the Quality of Experience (QoE) would be negatively affected, an automated system for degradation detection of video streaming is proposed. This approach is named QoE Management for Mobile Users (QoEMU). QoEMU is composed of several modules to perform a real-time analysis of the network traffic, select a mitigation action according to the information of the traffic and to some predefined policies, and apply these actions. In order to perform such tasks, the best Key Performance Indicators (KPIs) for a given set of video traces are selected. A QoE Model is trained to define a global QoE for the set of traces. When an alert regarding degradation in the quality appears, a proper mitigation plan is activated to mitigate this situation. The performance of QoEMU has been evaluated over a degradation situation experiments with different video users.

Planteamiento inicial del estudio y objetivos Las últimas tecnologías están dando pie a nuevos tipos de sistemas ABC en los cruces de frontera. Unos de estos tipos son, los sistemas ABC con dos etapas segregadas. Estos sistemas, separan en dos dispositivos los procesos clave del cruce de fronteras, el registro y la validación. La separación de las etapas tiene como ventaja que, los viajeros pueden registrarse con antelación al viaje, agilizando así el cruce de fronteras. Pero también tiene alguna desventaja ya que estos sistemas tienen dos subsistemas biométricos, con dos verificaciones faciales, lo que incrementa su vulnerabilidad. Al requerir dos capturas biométricas hay dos puntos en los que el sistema puede ser atacado mediante ataques de presentación. Este estudio analiza los subsistemas biométricos de los sistemas ABC Segregados, evalúa su rendimiento y propone un sistema PAD adaptado a la topología de estos sistemas. Para analizar en profundidad los sistemas ABC segregados, fue posible acceder a sistemas reales de este tipo durante la implantación de los pilotos del proyecto europeo ABC4EU. Los sistemas ABC4EU son sistemas segregados que se ajustan a las nuevas leyes establecidas para la zona Schengen. Las pruebas con los pilotos ABC4EU se llevaron a cabo en un cruce de fronteras real, en la terminal T4-S (satélite T4) del aeropuerto Adolfo Suárez de Madrid-Barajas. Metodología y herramientas utilizadas Se evalúan las verificaciones faciales en las dos etapas del sistema con dos reconocedores faciales de alto rendimiento, uno open-source (FaceNet) y otro COTS. Además de evaluar las verificaciones con presentaciones bona-fide, se probaron también, presentaciones de ataque con distintos PAI. Se evaluó la detección de ataques en las dos etapas del sistema, considerando dos escenarios de ataque, ataque sólo en la etapa de validación (VPA), y ataque en la etapa de registro y en la de validación (EPA+VPA). El ataque en validación consiste en suplantar a un viajero que se ha registrado previamente, y el ataque en ambas etapas, consiste en suplantar a un viajero al registrarse y continuar suplantando su identidad en el cruce de fronteras.

A fast automated biometric solution has been proposed to satisfy the future border control needs of airports resulting from the rapid growth in the number of passengers worldwide. Automated border control (ABC) systems handle the problems caused by this growth, such as congestion at electronic gates (e-gates) or delays in the planned arrival schedules. Different modalities, such as face, fingerprint, or iris recognition, will be used in most of the ABC systems located at airports in the European/Schengen areas. Because facial recognition is the modality that travelers consider most acceptable, it was decided to include this modality in all second generation passports. Face recognition systems, installed in small kiosks inside the e-gates, require high quality facial images to allow high performance and efficiency. Accurate face recognition algorithms, which should be invariant to non-idealities, such as changes in pose and expression, occlusions, and changes in lighting, are also required for these systems. In this paper, a review of the most important face recognition algorithms described in the literature that are invariant to these non-idealities and that can be used in ABC e-gates is presented. A comparative analysis of the most common ABC e-gates located at the different airports is provided. In addition, the results of an experimental evaluation of a face recognition system when halogen, white LEDs, near infra-red, or fluorescence illumination was used, which was conducted in order to determine which type of illumination is optimal for use in ABC e-gates, are presented. To conclude, improvements that could be implemented in the near future in ABC face recognition systems are described.

In recent years, there has been a growing interest in dynamic vision sensors due to their incredible advantages in speed, computational cost and power consumption. These new vision sensors have been inspired from biological retinae and use asynchronous address-event representation for visual information instead of a series of snapshots taken from traditional frame-based devices. Spiking neurons are biologically-plausible artificial neurons that process information in sequences of time events and are particularly suited for processing address-event information. A novel and refined biologically-inspired Gabor feature approach based on spiking neural networks is presented here. This approach utilises the retina-inspired data from dynamic vision sensors with Gabor edge detection in a hierarchical structure that has been populated with Leaky-Integrate and Fire neurons that have been trained via the Remote Supervision Method. The number of active spiking neurons at each time instance depends on the number of time events. This idea provides a flexible approach that avoids unnecessary computations and complexity. The biologically-inspired model developed for this preliminary work has shown promising results and has laid the foundation for a rapid parallel object recognition model designed for the new retina-like address-event representation sensors.

This paper presents a novelty system for the detection of driving-risk situations based on the knowledge acquired from traffic safety experts. A complete methodology to generate a driving-risk reference signal has been developed. A set of driving sessions was executed in a very realistic truck simulator, where several measures and visual information from the vehicle, the driver, and the road were collected. Two kinds of experiments were designed, i.e., controlled driving sessions (where several risk situations were induced) and natural driving sessions (where no risk situations were induced and a natural driving behavior was expected). A group of traffic safety experts from the Royal Automobile Club of Spain was consulted to evaluate the driving risk in each simulated session. The information acquired from the traffic safety experts was used to develop a methodology to combine their evaluations. The risks detected with the proposed methodology were analyzed to determine the most common human factors related with the generation of driving-risk situations.

The results and comparative analysis of two face recognition methods are presented in this article. The Permutation Coding Neural Classifier (PCNC) and Support Vector Machine (SVM) methods were selected. The main idea is to improve the image recognition rate. For this purpose we increase the training set by including distortions of initial images. Different numbers of distortions can increase the training image set and improve the quality of the classifier. The goal is to investigate the influence of the number of distortions on the PCNC recognition rate. Using distortions it is possible to improve the PCNC recognition rate. Sometimes it is possible to decrease the number of errors by 10 times or more. The PCNC with 12 distortions outperforms the results of SVM.

This paper presents a method for the selection of the optimal combination of experts' knowledge needed for the generation of a reliable driving risk ground truth. The driving risk of a controlled driving session, recorded in a highly realistic truck simulator, was evaluated by a large number of traffic safety experts. The risk evaluations were grouped in several clusters in order to find experts with high agreement. Next, a method for the selection of the optimal experts' evaluations is proposed. We found, through the experiments performed in this study, that a low number of experts are sufficient for the properly detection of driving risks. In addition, we show some of the advantages of the consideration of traffic safety experts' knowledge for the generation of a driving risk ground truth.

In this paper, a novel accident reproduction system for the identification of the main human factors involved on traffic accidents is presented. The system is based on a wireless in-vehicle Electronic Data Recorder that could be easily installed in any vehicle's cabin for the monitoring of the three basic elements of traffic safety: driver, road and vehicle. The system has been tested in a highly realistic truck simulator with a group of professional drivers. The data, collected with the system at the moments before traffic accidents, were used to generate a novel database that was carefully analyzed by a group of traffic safety experts. The validation process shows the reliability of the developed system as a tool for the identification of the main causes of the monitored traffic accidents.

A novel multidisciplinary system for the automatic driving risk level classification is presented. The data considered involves the three basic traffic safety elements (driver, road, and vehicle), as well as knowledge from traffic experts. The driving experiments were conducted in a truck cabin simulator handled by a professional driver, considering the most common real-world enviroments. Each traffic expert evaluate the driving risk on a 0 to 100 visual analogue scale. The driver, road and vehicle information was used to train five different data mining algorithms in order to predict the driving risk level. The benefits of the completeness of the data considered in our system are presented and discussed.

In this paper a novel method of information fusion at classifier level is applied to face verification. Three complementary kinds of facial data have been considered: texture, range data and curvature images. Three different kernels have been defined from each representation and finally a combined kernel has been developed. The resulting kernel has been used to train a classifier based on Support Vector Machines and it has been applied to face verification. The method has been deeply tested using the Face Recognition Grand Challenge database. The experiments show that in all cases the combined proposed classifier improves individual classifiers.