Abstract Recommender systems assist users in navigating information-rich environments by delivering personalized content. While model-based collaborative filtering approaches, such as matrix factorization (MF) and graph neural networks (GNN), are widely adopted, the inherent uncertainty in user preferences and the sparsity of data can lead to unreliable predictions. Confidence estimation has emerged as a strategy to quantify prediction reliability, yet its integration remains unexplored in GNN-based models, and prior methods often degrade accuracy or suffer from convergence issues. This study benchmarks four prominent confidence-aware models—OrdRec, Confidence-aware Probabilistic Matrix Factorization, Confidence-aware Bayesian Probabilistic Matrix Factorization, and Lightweight Beta Distribution across three public datasets: Amazon Movies and TVs, Jester Joke, and Movie Lens. We evaluate these models in terms of rating accuracy, ranking quality, and confidence estimate quality. In addition, we propose a novel confidence-integrated model based on a deep graph attention network architecture. Experimental results reveal that while distribution-based confidence methods are highly sensitive to dataset characteristics and harm accuracy, the proposed method demonstrates consistent performance across all datasets and metrics, outperforming prior distribution-based models. Nevertheless, challenges remain in aligning confidence estimates with prediction error.

• Multiple deep reinforcement learning agents coordinate home automation devices. • User interactions are modeled as implicit feedback to generate agent rewards • The recommender system successfully adapts to changes in users’ daily routines. • Two domain-specific action selection strategies adapted for reinforcement learning systems in environments with the possibility of routine change. • A new metric based on the Hamming Score for evaluating predictions for smart devices with a composite state. The rapid expansion of the Internet of Things (IoT) has provided the technological foundation for smart homes, in which interconnected devices enable the environment to adapt to residents’ needs. Yet, this proliferation of IoT-enabled devices also introduces a significant challenge: managing and coordinating their multiple operational states, which extend far beyond a simple “switch on/switch off”. Aiming to automate the action process and anticipate user actions, this study proposes a state recommender system for actuator devices in smart homes, developed using deep reinforcement learning algorithms integrated with implicit feedback. This approach enables composite state encoding and coordination among multiple agents, thereby anticipating user needs and adapting to dynamic routines. In addition to using a dataset captured from a real-world IoT environment, we use a smart home simulator to generate two datasets based on three different routines and perform experiments with two reinforcement learning algorithms, Deep Q-Learning and Differential Semi-gradient n-step SARSA, benchmarking their performance against an online Supervised Learning baseline based on Behavior Cloning. Additionally, we tested these algorithms using two-state approaches: simple and composite. The results confirm the system’s successful adaptation to user routines across both approaches, underscoring its potential to enhance personalization in smart home environments.

Abstract Video surveillance has benefited greatly from advances in artificial intelligence, particularly in computer vision, and the number of monitored environments has consequently increased, creating new challenges in extracting relevant information. When multiple cameras cover adjacent areas, overlapping fields of view can cause the same subject to be detected across cameras, introducing duplicate counts. Although the literature offers established solutions, many rely on high-quality recordings and complex, computationally intensive methods, limiting their use on resource-constrained devices. In this work, we address these challenges with a solution that leverages minimal information about target subjects and uses a transductive strategy to detect duplicates based on interactions within overlapping fields of view. Experiments in real-world settings show that our approach suppresses duplicates effectively, making it suitable for deployment on resource-limited hardware. We evaluate state-of-the-art lightweight models with high inference speed, as well as classical re-identification methods, in scenarios with low-quality video and constrained devices. The results underscore the effectiveness of the proposed approach and motivate exploration of re-ID with domain adaptation, as well as anchor-free methods with weak or semi-supervised learning.

Due to limited hardware, consumer-grade surveillance cameras usually rely on cloud-based computer vision models to detect people in video footage. However, this approach introduces a recurring cost, as users must continuously pay for cloud processing. One possible solution is to use mobile devices for person detection, as they can be found in most households. Yet, experimental evaluations on the impact of different computer vision models on mobile device resource usage are limited. This study examines the efficiency of YOLOv8 models on Android devices, assessing detection performance, inference time, memory consumption, and energy efficiency. The models were tested using two machine learning frameworks, LiteRT (formerly TensorFlow Lite) and ONNX Runtime, to determine the most suitable approach for mobile inference. Experimental results confirm that compact models, such as YOLOv8n and YOLOv8s, offer the best trade-off between computational efficiency and detection accuracy, while LiteRT outperforms ONNX in all evaluated metrics.

Smart homes are transforming domestic environments by integrating connected devices and sensors, enabling lighting, temperature, and security automation. While these systems enhance comfort and efficiency, they often rely on predefined settings or manual input due to the absence of adaptive recommendation systems. AI-driven recommendation systems personalize actions by learning from user behavior and environmental data, improving the smart home experience. However, they also introduce cybersecurity risks, particularly data poisoning attacks, where manipulated data disrupts system functionality. This paper exposes and examines vulnerabilities in smart home recommendation systems, categorizing data poisoning attacks and analyzing their impact. Through a literature review and attack vector analysis, we identify key weaknesses and propose mitigation strategies to enhance security. Our goal is to contribute to developing robust smart home technologies that protect user privacy, ensure reliability, and withstand adversarial threats.

ABSTRACT Context Smart home devices have become increasingly popular in modern households, powered by the Internet of Things (IoT) advances. The data generated by smart devices can provide valuable insights into users' behavior and preferences. By analyzing the data, one can understand how people interact with their homes, thus creating a “smart home profile”. To comprehend the complete IoT ecosystem dynamics of an intelligent environment, it is necessary to learn from each IoT device to predict its status in the future time. Nevertheless, dealing with real‐world IoT data structure requires considerable preprocessing tasks and the employment of classifiers that can learn multiple IoT inputs from a single IoT message. Objective Aware of these challenges, this paper proposes a novel methodology to process multi‐label IoT data and provide a comprehensive comparison of multi‐label classifiers for forecasting the status of smart devices, considering their efficiency and accuracy. Method We propose a data transformation method to preprocess the IoT data to be used by multi‐label classifiers. This method is based on real data structure. Results We evaluate our proposal in two real‐world scenarios and various multi‐label classifiers. The promising findings indicate that efficient classifiers can generate many correct predictions for a comprehensive IoT ecosystem in a small fraction of a second. Conclusions Our proposed data transformation can fit the context of prediction to smart homes and work with multi‐label classifiers to understand user behavior.

The essence of calibration in recommender systems is to generate recommendations that match the distribution of a given user’s past preferences regarding certain item features—e.g., in terms of preferred genres in the case of movies—while preserving relevance. The user’s past preference distribution is usually derived by considering the features of all items that the user previously liked. However, the most common approach in the literature to derive this distribution has certain limitations. First, it does not consider that user preferences may change over time. Second, there are domains where the relevant item features are set-valued, e.g., a movie can have several genres. In such cases, existing calibration approaches may represent the true user’s preference distribution in a suboptimal way. In this work, we, therefore, propose two novel approaches to derive the preference distributions of users for the purpose of calibration. The first method allows us to decrease the relevance of possibly outdated preference information. The second method is an entropy-based approach, which aims to capture better the user’s true preferences toward certain item features. Extensive experimental evaluations on four distinct datasets confirm that the proposed techniques are more effective in reducing the level of miscalibration than the common state-of-the-art calibration approach.

Recommendation systems are used to suggest items that users can be interested in. These systems are based on the user preference historic to create a recommendation list with items that have the higher similarity with the user's interests, in order to achieve the best possible user's satisfaction, which is usually measured as recommendation precision. However, the search for the best precision can cause some side effects such as overspecialization, few diversity and miscalibration of genres, classes and niches. Calibration provides fairer recommendations, which respect the genre proportionality on the user's preferences, avoiding overspecialization. This article aims to explore ways to balance the trade-off weight between precision and calibration based on divergence measures, as well as to propose metrics to evaluate the calibration in the suggested list. The proposed system works in a post-processing step and does not depend on a specific recommender algorithm or workflow. For this purpose, we evaluate six recommender algorithms applied in the movie domain, analyzing variations of three fairness measures, two personalized trade-off weights and eleven constant weights. To understand the results we use the precision, the reciprocal rank and two proposed metrics. The results indicate that the trade-off formulation of personalized weights obtains better results when used to compare the recommendation lists using matrix factorization-based approaches on Movielens dataset. In addition, the calibration also impacts the precision and fairness of all considered algorithms used in evaluation.

Recommendation strategies are used in several contexts in order to bring potential users closer to products with a strong probability of interest.When recomendations focus on niche items, they are called recommendations in the long tail.In these cases, they also look for less popular items and try to find your target custumer, niche market.This paper proposes a long tail recommendation approach that prioritizes relevance, diversity and popularity of recommended items.For that, a hybrid approach based on two techniques are used.The first is clustering with dynamic parameters that adapt from according to the dataset used and the second is a type of Markov chains for to calculate the distance of interest of a user to an item of relevance for this user.The results show that the techniques used have a better relevance indexes at the same time more diverse and less popular recommendations.

A vast amount of data that can be easily read by machines have been published in freely accessible and interconnected datasets, creating the so-called Linked Open Data cloud. This phenomenon has opened opportunities for the development of semantic applications, including recommender systems. In this paper, we propose Personalized Linked Data Semantic Distance (PLDSD), a novel similarity measure for linked data that personalizes the RDF graph by adding weights to the edges, based on previous user's choices. Thus, our approach has the purpose of minimizing the sparsity problem by ranking the best features for a particular user, and also, of solving the item cold-start problem, since the feature ranking task is based on features shared between old items and the new item. We evaluate PLDSD in the context of a LOD-based Recommender System using mixed data from DBpedia and MovieLens, and the experimental results indicate better accuracy of recommendations compared to a non-personalized baseline similarity method.

It is becoming more common to publish data in a way that accords with the Linked Data principles. In an effort to improve the human exploitation of this data, we propose a Linked Data browser that is enhanced with recommendation functionality. Based on a user profile, also represented as Linked Data, we propose a technique that we call LDRec that chooses in a personalized way which of the resources that lie within a certain neighbourhood in a Linked Data graph to recommend to the user. The recommendation technique, which is novel, is inspired by a collective classifier known as the Iterative Classification Algorithm. We evaluate LDRec using both an off-line experiment and a user trial. In the off-line experiment, we obtain higher hit rates than we obtain using a simpler classifier. In the user trial, comparing against the same simpler classifier, participants report significantly higher levels of overall satisfaction for LDRec.

The variety of interaction paradigms on the Web, such as clicking, commenting or rating are important sources that help recommender systems to gather accurate information about users' preferences.Ensemble methods can be used to combine all these pieces of information in a post-processing step to generate recommendations that are more relevant.In this paper, we review the application of existing ensemble methods to improve ranking recommendations in the multimodal interactions context.We compared four ensemble strategies, ranging from simple to complex algorithms including Gradient Descent and Genetic Algorithm to find optimal weights.The evaluation using the HetRec 2011 MovieLens 2k dataset with three different types of interactions shows that a considerable 7% improvement in the Mean Average Precision can be achieved using ensembles when compared to the most performant single interaction.

The recommendation systems aim to minimize information overload by helping user's in searching desired information. Faced with this scenario, we investigate the use of cloud factors able to have a positive influence on generating recommendations. Thus, we present a new, simple model based on cloud features which is associated with the content-based technique of recommendation. The practical applicability of data storage environments in the cloud provides the best use of cloud resources and meets user's preferences.

Content sharing services have become immensel popular on the Web. More than 1 billion people use this kind o services to communicate with friends and exchange all sorts o information. In this new context, privacy guarantees are essential guarantees about the potential release of data to unintended recipients and the use of user data by the service provider Although the general public is concerned about privacy question related to unintended audiences, data usage by service provider is still misunderstood. In order to further explore this level o misunderstanding, this work presents the results of a surve: conducted among 900 people with the aim of discovering hov people care about the use of their personal data by servic providers in terms of social media. From the results, we found that: (i) in general people do not read license terms and do no know very much about service policies, and when presented with these policies people do not agree with them; (ii) a good number of people would support alternative models such as paying for privacy or selling their personal data; and (iii) there are some differences between generations in relation to how they care about their data.

Collaborative tagging has emerged as a useful means to organize and share resources on the Web. Recommender systems have been utilized tags for identifying similar resources and generate personalized recommendations. In this paper, we analyze social and behavioral aspects of a tag-based recommender system which suggests similar Web pages based on the similarity of their tags. Tagging behavior and language anomalies in tagging activities are some aspects examined from an experiment involving 38 people from 12 countries.