Causal discovery offers a path for Machine Learning models to move beyond mere correlation to understand cause-and-effect relationships. However, its operational complexity remains a significant barrier for non-expert developers. This paper introduces Causal-Nest , a new, comprehensive framework that simplifies and operationalizes the causal analysis pipeline for non-experts. The contribution of the framework is demonstrated through an extensive, two-part evaluation. First, we validate its theoretical correctness and diagnostic power using the well-known Sachs et al. dataset and an extended benchmark of seven real-world datasets with known causal constraints. Our results show that the proposed framework effectively assesses causal graph integrity, using metrics, including the novel Knowledge Integrity Score (KIS), to pinpoint structural discovery failures and statistical unreliability quantitatively. Second, we demonstrate its practical utility in a common Machine Learning scenario where no such ground truth is available: telecommunications customer churn prediction. We show that using the proposed framework to guide a causality-driven feature engineering process yields a causal oversampled feature set that significantly improves the minority-class F1-Score of ensemble prediction models (e.g., CBC, GBMC). This work validates Causal-Nest as a single, dual-purpose tool that successfully bridges the gap between rigorous causal validation and practical, performance-driven machine learning applications. • Introduces Causal-Nest , a framework to simplify practical causal discovery. • Integrates 12 causal discovery algorithms with parallel execution. • Introduces Knowledge Integrity Score (KIS) and Priority Score metrics for model validation and prioritization. • Supports causal effect estimation and refutation with time limits. • Enhances ML feature engineering through causality-driven insights, validated by improved churn prediction metrics in a case study.

Churn is a significant problem for digital banks, given how easy it is for users to install and uninstall the mobile apps. This paper explores the application of deep learning techniques for churn prediction in the digital banking sector, focusing on privacy-preserving methods that rely solely on app installation and uninstallation data. We propose two models: AutoRNN, a hybrid model combining autoencoder RNNs and feed-forward neural networks, and an Encoder-only Transformer model adapted for non-sequential data. Both models are evaluated against a baseline Random Forest model optimized using TPOT. Our results demonstrate that the Transformer-based model consistently outperforms the baseline and AutoRNN across most metrics, particularly in accuracy and non-churn recall. AutoRNN shows promise in capturing temporal patterns but exhibits variability in performance across different datasets. The baseline model, while simpler, remains competitive, especially in churn precision. This work highlights the potential of privacy-centric, churn prediction as a viable and efficient approach.

This study presents a comprehensive analysis of predictive modeling approaches for Brazilian Real Estate Investment Funds (REIFs), evaluating the comparative performance of machine learning techniques against traditional time-series methods. Our methodological framework examines three critical dimensions: (1) fund categorization (paper, brick, and hybrid assets), (2) modeling architectures (individual fund-specific models versus a unified ensemble approach), and (3) temporal forecasting horizons (1-month versus 6-month predictions). The research introduces significant innovation through the systematic integration of property-level attributes into the predictive models, marking the first such application in REIF forecasting literature. Empirical results demonstrate that machine learning models achieve superior predictive accuracy, particularly for medium-term forecasts, while the incorporation of physical property characteristics emerges as a statistically significant factor in enhancing model performance. These findings contribute both methodological and practical insights for investors and fund managers operating in Brazil’s real estate investment sector.

Customer Lifetime Value (CLV) estimation plays a pivotal role in enhancing customer management and long-term business planning. However, current methodologies often suffer from significant constraints, including strong dependence on context-specific variables, substantial data requirements, and potential threats to user privacy. This work introduces a versatile and privacy-conscious machine learning approach to CLV prediction that relies solely on variables extracted from transaction timestamps and monetary amounts. The proposed model was evaluated across five distinct datasets, each with varying structures and behavioral patterns, and benchmarked against well-known models from the literature. Results demonstrate that the proposed solution delivers robust predictive accuracy, achieving up to a 34% reduction in CLV error in some cases, while offering adaptability across a wide range of business scenarios.

Human mobility analysis is critical for applications ranging from urban planning and epidemic modeling to personalized recommendation systems. However, existing approaches often neglect the semantic dimension of mobility which can reveal latent behavioral insights. In this work, we introduce semantic motifs to capture lifestyle shifts at both individual and collective levels. Using the proposed framework, we analyze 95.5 million GPS records from 65,402 mobile users in Brazil, comparing movement behaviors between 2021, strict COVID-19 restrictions period, and 2022, after easing restrictions. Our distributed pipeline, implemented on Apache Spark and Apache Sedona, scales efficiently to extract and cluster these motifs, enabling large-scale spatiotemporal analysis. Key findings reveal a transition from teleworking to hybrid lifestyles with greater variability in displacements.

Federated Learning (FL) enables devices to collaboratively train a shared model by exchanging parameters with a central server rather than raw data, thereby enhancing privacy, scalability, and efficiency. As modern devices are increasingly equipped with diverse sensors, supporting multiple tasks has become essential. Multi-model Federated Learning (MEFL) extends FL by allowing a single server to coordinate the training of multiple independent tasks, improving overall performance and resource utilization. In this work, we propose MultiFedAvg with Model-wise Data Heterogeneity Awareness (MultiFedAvg-MDH), the first approach to adapt multi-model orchestration in MEFL according to each model’s data heterogeneity. Our key insight is that, under highly non-IID conditions, prioritizing training intensity over frequency leads to more robust learning. Experimental results demonstrate that MultiFedAvg-MDH improves system performance by up to 16.49% while also reducing performance variability.

There is a shortage of mobility datasets - real or synthetic - available in the literature, limiting the development of new research. This generates demand for newer large-scale datasets, considering time, space, and population size. This work introduces a novel, two-step approach to generating large-scale mobility data considering the surrounding context, such as traffic. We validate our approach using two real mobility datasets, resulting in an enriched, large-scale dataset with more than 1 million origin-destination trips, which we share with the community to enable new research opportunities.

An event can be defined as something that happens at a particular place and time. A traffic event is a specific event kind that occurs on roads and affects the users’ mobility. Traffic events can be helpful in a variety of Intelligent Transportation System (ITS) applications, such as routing planning and emergency notifications. The detection of traffic events is not a trivial task, given the particularities of the environment and data availability. In this work, we propose a workflow that guides an ITS application designer on modeling how to detect events of interest, given the application’s requirements and the available data characteristics. As part of the workflow, we propose a decision component that selects the most appropriate event extraction algorithm for a particular scenario. An instance of the proposed model using two social networks as data sources and four machine learning algorithms was implemented as a case study. The results reveal that it was possible to extract a significant part of the expected events, all of them with complete what, where, and when information.

In recent years, we are facing a significant increase in the collection and availability of geospatial data. This type of data is paramount to help decision makers in different contexts, such as smart cities, mobile social networks, and e-commerce. In addition, the behavior of mobile users can be extracted and exploited to improve the quality of the services offered by mobile providers. However, raw location data needs to be semantically enriched to be useful, which is an arduous task because it requires queries on different data sources of different formats and complex joins. To mitigate this difficulty, this work proposes MoreData, a flexible and expandable framework for the semantic enrichment of geospatial data. The framework accepts different input formats, provides connectors to different sources (APIs, Relational Database, Fast Search Engine Tool, and Open Street Map), and is easy to extend to new sources. To evaluate the framework's benefits, we use it to enrich a real database comprised of thousands of mobile users' locations with data. We observe that the adoption of the framework leads to less time and effort in the data enrichment process, saving time for the more important tasks, such as building and analyzing models.

The feasibility of vehicular networks is directly related to the understanding of mobility patterns, which is a necessary knowledge for the elaboration and application of novel algorithms and technologies for such networks. Thereupon, mobility traces have an important role in the exploration of mobility patterns. Since mobility traces represent the movement of entities in real-life situations, their comprehensive study is a great opportunity to characterize mobility behaviors, which can then be applied to improve the design of mobile networks. In this work, we perform a thorough characterization of a taxi dataset regarding spatial and temporal characteristics, to better understand how these factors affect traffic conditions. To this end, we define a new metric, called path optimality, that represents the rate of trips that follow the shortest path between two points. We use this metric to understand what are the conditions that affect the mobility pattern of drivers on a large city.

The discovering of patterns regarding how, when, and where users interact with mobile applications reveals important insights for mobile service providers. In this work, we exploit for the first time a real and large-scale dataset representing the records of mobile application usage of 5,342 users during 2014. The data was collected by a software agent, installed at the users’ smartphones, which monitors detailed usage of applications. First, we look for general patterns of how users access some of the most popular mobile applications in terms of frequency, duration, diversity, and data traffic. Next, we mine the dataset looking for temporal patterns in terms of when and how often accesses occur. Finally, we exploit the location of each access to detect users’ points of interest and location-based communities. Based on the results, we derive a model to generate synthetic datasets of mobile application usage and evaluate solutions to predict the next application to be launched. We also discuss a series of implications of the findings regarding telecommunication services, mobile advertisements, and smart cities. This is the first time this dataset is used, and we also make it publicly available for other researchers.

The focus of this work is on the detection of incidents that have a direct impact on the traffic of vehicles in large cities, such as, accidents, road constructions-renovations and traffic jams using Online Social Networks(OSNs). The proposed model aims to find problems being reported, as well as information on the location of the event. The results obtained were significant in the task of categorizing the incident, reaching up to 94% accuracy and 98% of general hits in the task of determining usual traffic incidents, besides promising results in obtaining references to the points in the city where the incidents take place, with up to 58% recall.

Simulation is the most frequently adopted approach for evaluating protocols and algorithms for Vehicular Ad hoc Networks (VANETs) and Delay-Tolerant Networks (DTNs). Usually, simulation tools use mobility traces to build the network topology based on the existing contacts between mobile nodes. However, quality of the traces, in terms of spatial and temporal granularity of each entry in the logfile, is a key factor that impacts the network topology directly. Therefore, the reliability of the results depends strongly on the accurate representation of the real network topology by the vehicular mobility model. We show that five widely adopted existing real vehicular mobility traces present gaps, leading to fallible outcomes. In this work, we propose a solution to fill those gaps, leading to more fine-grained traces, which lead to more trustworthy simulation results. We propose and evaluate a data-based solution using clustering algorithms to fill the gaps of real-world traces. In addition, we also present the evaluation results that compare the communication graph of the original and the calibrated traces using network metrics. The results reveal that the gaps do indeed induce network topologies differing from reality, decreasing the quality of the evaluation results. To contribute to the research community, we have made the calibrated traces publicly available, so that other researchers may adopt them to improve their evaluation results.

Vehicular network applications are emerging to bring many benefits to the users during their journey. However, the design and deployment of such applications require studies that provide insightful information about the network formed by vehicles. To this end, in the last years researchers have been characterizing real mobility traces collected from taxis equipped with GPS devices. However, these traces present temporal and spatial gaps, as it was demonstrated in other studies. In this work, we compare the real original traces with their calibrated version (i.e., with no gaps) to show that the existing gaps expressively affect key metrics such as contact duration, inter-contact time, and network capacity. As result, we contribute to the research community by presenting important analysis of real vehicular network topology and by showing the importance of calibrating the traces before evaluating vehicular network applications.

Most Vehicular Ad-hoc Network (VANET) applications require the delivery of a variety of content to vehicles. Furthermore, content in such applications is usually geo-localized (i.e., designated to a specific region of interest). To help with the delivery of geo-localized content in VANET applications, content replication strategies can be exploited to keep content available where the interested vehicles are expected to be. To this end, we propose a Geo-Localized Origin-Destination-based Content Replication (GO-DCR) solution that relies on the vehicles' origin and destination points to select the vehicles most likely appropriate to replicate content. We compare GO-DCR to an existing solution through simulation. The results reveal that GO-DCR increases the content availability while reducing delivery cost.

Vehicular ad-hoc network (VANET) is a special type of mobile networks that has received significant attention from the academia and the industry in recent years. VANET applications are emerging as new opportunities for the automobile industry to provide advanced services to connected vehicles and their users. Most of those services, like information, safety, and entertainment systems, require a variety of content to be delivered to vehicles. An useful concept in this case is the one of content replication, which defines that content should be placed close to its potential clients. However, existing replicating solutions, as originally proposed for the Internet, are not suitable to VANETs because of their particular characteristics, such as highly dynamic topology, existence of push-based applications, information-centric nature, location- and time-dependent content, and intermittent connections. Therefore, we propose and evaluate in this work new techniques to replicate content in VANETs that rely on the vehicles' origin-destination (O-D) points. The results reveal that our solutions were able to increase the content availability in two case studies, named city-wide and region-wide.

Simulation is the approach most adopted to evaluate Vehicular Ad hoc Network (VANET) and Delay-Tolerant Network (DTN) solutions. Furthermore, the results' reliability depends fundamentally on mobility models used to represent the real network topology with high fidelity. Usually, simulation tools use mobility traces to build the corresponding network topology based on existing contacts established between mobile nodes. However, the traces' quality, in terms of spatial and temporal granularity, is a key factor that affects directly the network topology and, consequently, the evaluation results. In this work, we show that highly adopted existing real vehicular mobility traces present gaps, and propose a solution to fill those gaps, leading to more fine-grained traces. We propose and evaluate a cluster-based solution using clustering algorithms to fill the gaps. We apply our solution to calibrate three existing, widely adopted taxi traces. The results reveal that indeed the gaps lead to network topologies that differ from reality, affecting directly the performance of the evaluation results. To contribute to the research community, the calibrated traces are publicly available to other researchers that can adopt them to improve their evaluation results.

Information dissemination applications (e.g. weather, news, advertisements, traffic, and so on) are among the most promising ones for vehicular networks. However, disseminating information to all vehicles in vehicular networks is not a trivial task because of their specific characteristics like high dynamic topology and different density values along time and space. The proposals found in literature focused on improved flooding schemes that, in realistic scenarios, may cause unnecessary overhead and redundant data. This work addresses the dissemination to all vehicles problem by assuming that some vehicles are more appropriate to be good disseminators than others. This hypothesis was validated through a large-scale urban mobility scenario and two heuristics were proposed to select the appropriate disseminators. Results showed that the proposed heuristics were able to achieve high coverage (more than 97% of vehicles) with low redundant data.

Vehicular Ad-Hoc Network (VANET) applications are context-aware since they need environment and local contexts to operate properly and help their users on many aspects, like indicating the best route to take or warning about accidents and traffic congestion situations. These applications need a context provider system responsible for collecting, analyzing, reasoning and making contexts available to them. In this work we propose a smart context provisioning middleware, called ConProVA, that provides context information to applications, is able to infer logical context and deals with conflicts of interest. We evaluate through simulation an instance of ConProVA considering an accident detection and avoidance application. The results show that it is possible to improve the users' satisfaction by reducing total travel time and carbon dioxide emission when accidents occur.

Evaluating large scale vehicular networks solutions is a great challenge once real experiments are not feasible yet and there are no large scale testbeds available. Thus, the most used evaluation technique is simulation, which requires a realistic mobility model to provide accurate results. However, building large scale realistic vehicular mobility models is a hard task and many important characteristics have been neglected, like the origin-destination (O-D) macroscopic one. In this work it is proposed a trace-based O-D macroscopic model inferred by the adoption of statistical techniques in order to characterize a validated traffic model. This work contributes to the research community by providing a more realistic O-D macroscopic model, which can be considered a complement for many traffic generators available in literature.

Context-awareness allows ubiquitous systems to adapt their services according to the current users' characteristics and the surrounding environment. Most of the context-aware applications found in literature only consider external contexts like location and temperature. However, internal contexts, like feelings and physiological needs, could also be relevant to adaptation decisions. Nevertheless, these contexts are hard to be obtained because, usually, they can not be gathered by physical sensors. In this work it is proposed a system that infers feelings and physiological needs based on the semantic processing of sentences posted in Twitter by the users of ubiquitous context-aware applications. The preliminary qualitative results show that the inference system is accurate when the sentences are well-formed and do not present typos.

There are many approaches that can be applied to the management of wireless sensor networks (WSN) and nowadays the research in this field is only beginning. A management approach determines how the management monitor and control functions are performed by the network and network elements. Since a wireless sensor network presents severe resource constraints, a certain management approach must be chosen properly in order to optimize resource usage and at the same time satisfy the application requirements. In this paper we try to answer the following question: what is the tradeoff between to migrate management services or to adopt traditional approaches for WSN management? We have concluded from simulated and experimental results that the choice of the best approach to be adopted must consider the size of the network, the configuration of the nodes, the application goals and the management complexity, in terms of managed objects. We have also noticed that, despite the service migration approach being theoretically interesting for WSN management, for the sensor networks and sensor nodes hardware current technological situation, the approach would be applied with some negative impact on the performance of the network.

The goal of a wireless sensor network (WSN) management solution is to promote network resources productivity and quality of services. This paper presents a comparative study of client/server (CS) and mobile agent (MA) approaches used in distributed self-management solutions for a hierarchical heterogeneous WSN, in which the managers are embedded into the cluster head nodes. These approaches were simulated and evaluated according to energy consumption. Considering energy consumption with transmission, the results reveal that the MA approach is more scalable than the CS one, tending to be more useful when the number of network elements and managed objects increases. On the other hand, the CS approach presents more interesting results regarding energy consumption with processing. Besides, the simulations performed also show that the MA size is an important parameter that impacts directly on the results

A wireless sensor network aims to collect data and, sometimes, control an environment. This kind of network is composed of hundreds to thousands of devices that have the capability of sensing, processing and wireless communicating, called sensor nodes. The sensor nodes are projected with small dimensions (cm/sup 3/ or mm/sup 3/) and this size limitation ends up restraining the node resources, like energy, processor and transceiver capacity. The task of building and deploying management systems in environments where there will be tens of thousand of network elements with particular features and organization is very complex. This task becomes worse due to the physical restrictions of these unattended sensor nodes. In this paper we have implemented and evaluated some automatic services of configuration and performance management, proposed by a WSN management architecture called MANNA. This architecture is based on the paradigm of self-management, which contains the automatic functions and services of management using a minimum of human interference. This work aims to evaluate different WSN configurations considering an application of continuous data sensing and dissemination, and the effects of the management solution proposed for this network. The built application does temperature and carbon monoxide concentration level monitoring, in an urban area. The results show the cost-benefit relations of the different organizations and demonstrates that management can promote the productivity of the resources and control the quality of the provided services.