Abstract This work examines the Halpern’s iterative method as a means to create new neural network optimizers that could surpass many current existing approaches. We introduce HalpernSGD, an innovative network optimizer that leverages Halpern’s technique to enhance the rate of convergence of the Stochastic Gradient Descent (SGD), leading to reduced carbon emissions in neural network training processes. The combination of Halpern’s iterative method and Gradient Descent (GD) has led to an algorithm with a quadratic rate of convergence compared to the simple GD. Experimental comparisons between their stochastic versions show that HalpernSGD achieves greater efficiency than SGD by requiring fewer training epochs, thus reducing energy consumption and carbon footprint while maintaining model accuracy. We also compare HalpernSGD with ADAM, identifying potential improvements to ADAM’s approach in terms of stability and convergence, and suggesting a future direction for the development of combined optimizers. The implementation for HalpernSGD can be accessed at: https://github.com/EttoreRitacco/HalpernSGD.git

Research in neural generative models for dynamic networks is constantly evolving, and sophisticated solutions have been exploited to characterize the long-term evolution of temporal graphs. Despite the efforts in the literature, state-of-the-art models face the problem of handling changes in the graph structure by relying on prior knowledge, compromising the model’s flexibility. In this paper, we propose a graph-size invariant probabilistic generative model, named $$\textrm{FuDGE}$$ , Fully Dynamic Graph Evolution, for predicting the graph evolution through step-wise changes in the graph structure. $$\textrm{FuDGE}$$ can generate evolving graphs by exploring the whole node space, thus ensuring fast and effective generation. We evaluate $$\textrm{FuDGE}$$ on real and synthetic benchmark datasets and compare its performance against state-of-the-art competitors. The results demonstrate that our approach offers a competitive advantage in generation and prediction quality compared to existing literature. The code is publicly available at https://github.com/FuDGE2023/fudge .

Simulating a recommendation system in a controlled environment, to identify specific behaviors and user preferences, requires highly flexible synthetic data generation models capable of mimicking the patterns and trends of real datasets.In this context, we propose HyDRA, a novel preferences data generation model driven by three main factors: user-item interaction level, item popularity, and user engagement level.The key innovations of the proposed process include the ability to generate user communities characterized by similar item adoptions, reflecting real-world social influences and trends.Additionally, HyDRA considers item popularity and user engagement as mixtures of different probability distributions, allowing for a more realistic simulation of diverse scenarios.This approach enhances the model's capacity to simulate a wide range of real-world cases, capturing the complexity and variability found in actual user behavior.We demonstrate the effectiveness of HyDRA through extensive experiments on well-known benchmark datasets.The results highlight its capability to replicate real-world data patterns, offering valuable insights for developing and testing recommendation systems in a controlled and realistic manner.The code used to perform the experiments is publicly available: https://github.com/flexibledatageneration/HYDRA.

User navigation on social media platforms is often driven by recommendation algorithms . A growing body of literature questions whether these recommendation systems may exacerbate detrimental phenomena, perpetrate intrinsic biases, and alter user preferences in the long-term. Driven by this premise, the present study formalizes the concept of “ algorithmic drift ”, further introducing a novel framework and two metrics to quantify it. Our methodology involves a simulation process that models user behavior through random walks , reflecting user navigation under the influence and guidance of recommendation systems. This approach highlights that each user may respond differently to such stimuli, varying in both resistance to recommendation influence and inertia in selecting new steps in the random walk. The proposed metrics measure the drift in user behavior and item consumption over time in the random walks. We conduct a comprehensive evaluation over both synthetic and real-world datasets to validate the framework’s ability to measure drift across different parameter settings. All code and data used in our experimentation are publicly accessible online. 1

Abstract The capability to devise robust outlier and anomaly detection tools is an important research topic in machine learning and data mining. Recent techniques have been focusing on reinforcing detection with sophisticated data generation tools that successfully refine the learning process by generating variants of the data that expand the recognition capabilities of the outlier detector. In this paper, we propose $$\textrm{ARN}$$ ARN , a semi-supervised anomaly detection and generation method based on adversarial counterfactual reconstruction. $$\textrm{ARN}$$ ARN exploits a regularized autoencoder to optimize the reconstruction of variants of normal examples with minimal differences that are recognized as outliers. The combination of regularization and counterfactual reconstruction helps to stabilize the learning process, which results in both realistic outlier generation and substantially extended detection capability. In fact, the counterfactual generation enables a smart exploration of the search space by successfully relating small changes in all the actual samples from the true distribution to high anomaly scores. Experiments on several benchmark datasets show that our model improves the current state of the art by valuable margins because of its ability to model the true boundaries of the data manifold.

Popularity bias is the tendency of recommender systems to further suggest popular items while disregarding niche ones, hence giving no chance for items with low popularity to emerge. Although the literature is rich in debiasing techniques, it still lacks quality measures that effectively enable their analyses and comparisons. In this article, we first introduce a formal, data-driven, and parameter-free strategy for classifying items into low, medium, and high popularity categories. Then we introduce Balanced Quality Score (BQS) , a quality measure that rewards the debiasing techniques that successfully push a recommender system to suggest niche items, without losing points in its predictive capability in terms of global accuracy. We conduct tests of BQS on three distinct baseline collaborative filtering frameworks: one based on history-embedding and two on user/item-embedding modeling. These evaluations are performed on multiple benchmark datasets and against various state-of-the-art competitors, demonstrating the effectiveness of BQS.

Abstract In real-world applications, audio surveillance is often performed by large models that can detect many types of anomalies. However, typical approaches are based on centralized solutions characterized by significant issues related to privacy and data transport costs. In addition, the large size of these models prevented a shift to contexts with limited resources, such as edge devices computing. In this work we propose conv-SPAD , a method for convolutional SPectral audio-based Anomaly Detection that takes advantage of common tools for spectral analysis and a simple autoencoder to learn the underlying condition of normality of real scenarios. Using audio data collected from real scenarios and artificially corrupted with anomalous sound events, we test the ability of the proposed model to learn normal conditions and detect anomalous events . It shows performances in line with larger models, often outperforming them. Moreover, the model’s small size makes it usable in contexts with limited resources, such as edge devices hardware.

We propose ARN, a semisupervised anomaly detection and generation method based on adversarial reconstruction. ARN exploits a regularized autoencoder to optimize the reconstruction of variants of normal examples with minimal differences, that are recognized as outliers. The combination of regularization and adversarial reconstruction helps to stabilize the learning process, which results in both realistic outlier generation and substantial detection capability. Experiments on several benchmark datasets show that our model improves the current state-of-the-art by valuable margins because of its ability to model the true boundaries of the data manifold.

In this paper, we propose a survival factorization framework that models information cascades by tying together social influence patterns, topical structure, and temporal dynamics. This is achieved through the introduction of a latent space which encodes: (a) the relevance of an information cascade on a topic; (b) the topical authoritativeness and the susceptibility of each individual involved in the information cascade, and (c) temporal topical patterns. By exploiting the cumulative properties of the survival function and of the likelihood of the model on a given adoption log, which records the observed activation times of users and side-information for each cascade, we show that the inference phase is linear in the number of users and in the number of adoptions. The evaluation on both synthetic and real-world data shows the effectiveness of the model in detecting the interplay between topics and social influence patterns, which ultimately provides high accuracy in predicting users activation times.

Variational autoencoders were proven successful in domains such as computer vision and speech processing. Their adoption for modeling user preferences is still unexplored, although recently it is starting to gain attention in the current literature. In this work, we propose a model which extends variational autoencoders by exploiting the rich information present in the past preference history. We introduce a recurrent version of the VAE, where instead of passing a subset of the whole history regardless of temporal dependencies, we rather pass the consumption sequence subset through a recurrent neural network. At each time-step of the RNN, the sequence is fed through a series of fully-connected layers, the output of which models the probability distribution of the most likely future preferences. We show that handling temporal information is crucial for improving the accuracy of the VAE: In fact, our model beats the current state-of-the-art by valuable margins because of its ability to capture temporal dependencies among the user-consumption sequence using the recurrent encoder still keeping the fundamentals of variational autoencoders intact.

The supervised classification of XML documents by structure involves learning predictive models in which certain structural regularities discriminate the individual document classes. Hitherto, research has focused on the adoption of prespecified substructures. This is detrimental for classification effectiveness, since the a priori chosen substructures may not accord with the structural properties of the XML documents. Therein, an unexplored question is how to choose the type of structural regularity that best adapts to the structures of the available XML documents. We tackle this problem through X-Class, an approach that handles all types of tree-like substructures and allows for choosing the most discriminatory one. Algorithms are designed to learn compact rule-based classifiers in which the chosen substructures discriminate the classes of XML documents. X-Class is studied across various domains and types of substructures. Its classification performance is compared against several rule-based and SVM-based competitors. Empirical evidence reveals that the classifiers induced by X-Class are compact, scalable, and at least as effective as the established competitors. In particular, certain substructures allow the induction of very compact classifiers that generally outperform the rule-based competitors in terms of effectiveness over all chosen corpora of XML data. Furthermore, such classifiers are substantially as effective as the SVM-based competitor, with the additional advantage of a high-degree of interpretability.

Recent works in Recommender Systems (RS) have investigated the relationships between the prediction accuracy, i.e. the ability of a RS to minimize a cost function (for instance the RMSE measure) in estimating users' preferences, and the accuracy of the recommendation list provided to users. State-of-the-art recommendation algorithms, which focus on the minimization of RMSE, have shown to achieve weak results from the recommendation accuracy perspective, and vice versa. In this work we present a novel Bayesian probabilistic hierarchical approach for users' preference data, which is designed to overcome the limitation of current methodologies and thus to meet both prediction and recommendation accuracy. According to the generative semantics of this technique, each user is modeled as a random mixture over latent factors, which identify users community interests. Each individual user community is then modeled as a mixture of topics, which capture the preferences of the members on a set of items. We provide two different formalization of the basic hierarchical model: BH-Forced focuses on rating prediction, while BH-Free models both the popularity of items and the distribution over item ratings. The combined modeling of item popularity and rating provides a powerful framework for the generation of highly accurate recommendations. An extensive evaluation over two popular benchmark datasets reveals the effectiveness and the quality of the proposed algorithms, showing that BH-Free realizes the most satisfactory compromise between prediction and recommendation accuracy with respect to several state-of-the-art competitors.

The Web is an evolving system, which tries to adapt to the needs of users. The transition to Web2.0, and, currently, to Web3.0, are the expression of this trend: the goal is to focus on the leading role of the end user in Web browsing, which should be supported by adequate tools. In this paper, we propose Borè, an architectural paradigm for developing content-based web applications based on cooperative interaction, whose foundations are based on the principles of the model Web3.0. The proposed architecture is extremely innovative in three respects. The first one is the possibility of defining, organizing, storing, querying and displaying the information as customizable objects and relations: a notexpert user can create the Web that he/she may prefer. A second aspect is the realization of social networks (Social Cooperations), which spontaneously arise, through user resource sharing. Finally, there is the possibility of analyzing users' browsing activities, through learning tools that enable the user to enrich his/her Web browsing experience with new knowledge.

This paper presents a hierarchical probabilistic approach to collaborative filtering which allows the discovery and analysis of both global patterns (i.e., tendency of some products of being ‘universally appreciated’) and local patterns (tendency of users within a community to express a common preference on the same group of items). We reformulate the collaborative filtering approach as a clustering problem in a high-dimensional setting, and propose a probabilistic approach to model the data. The core of our approach is a co-clustering strategy, arranged in a hierarchical fashion: first, user communities are discovered, and then the information provided by each user community is used to discover topics, grouping items into categories. The resulting probabilistic framework can be used for detecting interesting relationships between users and items within user communities. The experimental evaluation shows that the proposed model achieves a competitive prediction accuracy with respect to the state-of-art collaborative filtering approaches.

In this work we propose DAEDALUS, a formal framework and system, specifically focussed on progressive combination of mining and querying operators. The core component of DAEDALUS is the MO-DMQL query language that extends SQL in two respects, namely a pattern definition operator and the capability to uniform manipulating both raw data and unveiled patterns. DAEDALUS system is specifically focussed on movement data and has been implemented as a query execution layer on top of the Hermes Moving Object Database. The expressiveness and usefulness of the MODMQL language as well as the computational capabilities of DAEDALUS are qualitatively evaluated by means of a case study.