Bayesian networks (BN) are versatile probabilistic graphical models that enable efficient knowledge representation and inference. These models have proven effective across diverse domains, including healthcare, bioinformatics, economics, law, and image processing. The structure and parameters of a BN can be obtained by domain experts or directly learned from available data. However, as privacy concerns escalate, it becomes increasingly critical for publicly released models to safeguard sensitive information in training data. Typically, released models do not prioritize privacy by design, and the issue equally affects BNs. In particular, tracing attacks from adversaries can combine the released BN with auxiliary data to determine whether specific individuals belong to the data from which the BN was learned. The current approach to addressing this privacy issue involves introducing noise into the learned parameters. While this method offers robust protection against tracing attacks, it also significantly impacts the model’s utility, in terms of both the significance and accuracy of the resulting inferences. Hence, high privacy may be attained, but at the cost of releasing a possibly ineffective model. This paper introduces credal networks (CN) as a novel and practical solution for balancing the model’s privacy and utility. Specifically, after adapting the notion of tracing attacks, we demonstrate that a CN enables the masking of the learned BN, thereby reducing the probability of successful tracing attacks. As CNs are obfuscated but not noisy versions of BNs, they can achieve meaningful inferences while safeguarding the privacy of the released model. Moreover, we identify key learning information that must be concealed to prevent attackers from recovering the BN underlying the released CN. Finally, we conduct a set of numerical experiments to analyze how privacy gains can be modulated by tuning the CN hyperparameters. Our results confirm that CNs provide a principled, practical, and effective approach towards the development of privacy-aware probabilistic graphical models.

This paper presents a novel approach to address contextual bandit problems with partially observable, delayed feedback by introducing an approximate Thompson sampling technique. This is a common setting, with applications ranging from online marketing to vaccine trials. Leveraging Bootstrapped Thompson sampling (BTS), we obtain an approximate posterior distribution over delay distributions and conversion probabilities, thereby extending an Expectation-Maximisation (EM) model to the Bayesian domain. Unlike prior methodologies, our approach does not overlook uncertainty on delays. Within the EM framework, we employ the Kaplan-Meier estimator to place no restriction on delay distributions. Through extensive benchmarking against state-of-the-art techniques, our approach demonstrates superior performance across the majority of tested environments, with comparable performance in the remaining cases. Furthermore, our method offers practical implementation using off-the-shelf libraries, facilitating broader adoption. Our technique lays a foundation for extending to other bandit settings, such as non-contextual bandits or action-dependent delay distributions, promising wider applicability and versatility in real-world applications.

Causality is gaining more and more attention in the machine learning community and consequently also in recommender systems research. The limitations of learning offline from observed data are widely recognized, however, applying debiasing strategies like Inverse Propensity Weighting does not always solve the problem of making wrong estimates. This concept paper contributes a summary of debiasing strategies in recommender systems and the design of several toy examples demonstrating the limits of these commonly applied approaches. Therefore, we propose to map the causality frameworks of potential outcomes and structural causal models onto the recommender systems domain in order to foster future research and development. For instance, applying causal discovery strategies on offline data to learn the causal graph in order to compute counterfactuals or improve debiasing strategies.

Learning the structure of continuous-time Bayesian networks directly from data has traditionally been performed using score-based structure learning algorithms. Only recently has a constraint-based method been proposed, proving to be more suitable under specific settings, as in modelling systems with variables having more than two states. As a result, studying diverse structure learning algorithms is essential to learn the most appropriate models according to data characteristics and task-related priorities, such as learning speed or accuracy. This article proposes alternative algorithms for learning multidimensional continuous-time Bayesian network classifiers, introducing, for the first time, constraint-based and hybrid algorithms for these models. Nevertheless, these contributions also apply to the simpler one-dimensional classification problem for which only score-based solutions exist in the literature. More specifically, the aforementioned constraint-based structure learning algorithm is first adapted to the supervised classification setting. Then, a novel algorithm of this kind, specifically tailored for the multidimensional classification problem, is presented to improve the learning times for the induction of multidimensional classifiers. Finally, a hybrid algorithm is introduced, attempting to combine the strengths of the score- and constraint-based approaches. Experiments with synthetic and real-world data are performed not only to validate the capabilities of the proposed algorithms but also to conduct a comparative study of the available competitors.

A key challenge in computer vision and deep learning is the definition of robust strategies for the detection of adversarial examples. In this work, we propose the adoption of ensemble approaches to leverage the effectiveness of multiple detectors in exploiting distinct properties of the input data. To this end, the ENsemble Adversarial Detector (ENAD) framework integrates scoring functions from state-of-the-art detectors based on Mahalanobis distance, Local Intrinsic Dimensionality, and One-Class Support Vector Machines, which process the hidden features of deep neural networks. ENAD is designed to ensure high standardization and reproducibility to the computational workflow. Extensive tests on benchmark datasets, models and adversarial attacks show that ENAD outperforms all competing methods in the large majority of settings. The improvement over the state-of-the-art and the intrinsic generality of the framework, which allows one to easily extend ENAD to include any set of detectors and integration strategies, set the foundations for the new area of ensemble adversarial detection.

Reproducibility is a main principle in science and fundamental to ensure scientific progress. However, many recent works point out that there are widespread deficiencies for this aspect in the AI field, making the reproducibility of results impractical or even impossible. We therefore studied the state of reproducibility support on the topic of Reinforcement Learning & Recommender Systems to analyse the situation in this context. We collected a total of 60 papers and analysed them by defining a set of variables to inspect the most important aspects that enable reproducibility, such as dataset, pre-processing code, hardware specifications, software dependencies, algorithm implementation, algorithm hyperparameters, and experiment code. Furthermore, we used the ACM Badges definitions assigning them to the selected papers. We discovered that, like in many other AI domains, the Reinforcement Learning & Recommender Systems field is grappling with a reproducibility crisis, as none of the selected papers were reproducible when strictly applying the ACM Badges definitions according to our analysis.

This article presents the development of an expert system to support the diagnosis of post-harvest diseases of stored apples. We propose a picture-based and conversational interaction with users, where sampled images depicting symptoms of apples with known diseases are presented to users to elicit their feedback on perceived similarities in order to determine the most likely diagnosis of a diseased target apple. This article makes, besides the description of the industrial application scenario, multiple contributions circled around three rounds of user studies: (i) an usability and effectiveness assessment of the approach, where three user interface configurations are put to a test and the effectiveness of different types of user feedback mechanisms is assessed; (ii) contextual multi-armed bandit approaches for dynamic selection of displayed images with symptoms of diseased apples, that clearly outperform random and greedy sampling baseline strategies; (iii) a comparison of two different strategies for determining the context representation of a contextual multi-armed bandit approach, namely based on PCA of image features and a gamified large-scale user study. We therefore provide design insights for the development of such diagnosis applications on diseases that manifest themselves through visual symptoms in general and, hence, the findings can be also valid for domains other than post-harvest fruit diseases.

This paper describes a decision support system to fulfill the diagnosis of post-harvest diseases in apple fruit. The diagnostic system builds on user feedback elicitation via multiple conversational rounds. The interaction with the user is conducted by selecting and displaying images depicting symptoms of diseased apples. The system is able to adapt the image selection mechanism by exploiting previous user feedback through a contextual multi-armed bandit approach. We performed a large scale user experiment, where different strategies for image selection have been compared in order to identify which reloading strategy makes users more effective in the diagnosis task. Concretely, we compared contextual multi-armed bandit methods with two baseline strategies and identified that the exploration-exploitation principle significantly paid off in comparison to a greedy and to a random stratified selection strategy. Although the application context is very domain specific, we believe that the general methodology of conversational item selection for eliciting user preferences applies to other scenarios in decision support and recommendation systems.

Non-stationary continuous time Bayesian networks are introduced. They allow the parents set of each node in a continuous time Bayesian network to change over time. Structural learning of nonstationary continuous time Bayesian networks is developed under different knowledge settings. A macroeconomic dataset is used to assess the effectiveness of learning non-stationary continuous time Bayesian networks from real-world data.

This demo presents a conversational navigation approach for a diagnostic application of postharvest diseases of apple with the goal to educate users on the diagnosed diseases as well as to recommend consequences for the storage facility and what action to take for the next growing period. It thus builds on earlier works on picture-based navigation for conversational recommender systems and provides evidence for its usability based on a first small-scale comparative usability study.

When looking for recently published scientific papers, a researcher usually focuses on the topics related to her/his scientific interests. The task of a recommender system is to provide a list of unseen papers that match these topics. The core idea of this paper is to leverage the latent topics of interest in the publications of the researchers, and to take advantage of the social structure of the researchers (relations among researchers in the same field) as reliable sources of knowledge to improve the recommendation effectiveness. In particular, we introduce a hybrid approach to the task of scientific papers recommendation, which combines content analysis based on probabilistic topic modeling and ideas from collaborative filtering based on a relevance-based language model. We conducted an experimental study on DBLP, which demonstrates that our approach is promising.

The deep learning wave is propagating through many research areas and communities. In the last years it quickly propagated to Recommendation Systems, a research area which aims to recommend items to users. Indeed, many deep learning models and architectures have been proposed for Recommendation Systems to improve collaborative filtering and content based algorithms. In this paper we propose a hybrid recommendation system combining user ratings and natural language text processing to solve the 0/1 recommendation problem. In particular, we describe a deep learning architecture combining two information sources, namely natural language text and user rating. Natural language text is used to learn a user-specific content-based classifier, while user ratings are used to develop user-adaptive collaborative filtering recommendations. We perform numerical experiments on MovieLens 1M and reach first preliminary, but promising results, showing the proposed architecture has the potential to combine content-based and collaborative filtering recommendation mechanisms using a deep learning supervisor.

Most evaluations of novel algorithmic contributions assess their accuracy in predicting what was withheld in an offline evaluation scenario. However, several doubts have been raised that standard offline evaluation practices are not appropriate to select the best algorithm for field deployment. The goal of this work is therefore to compare the offline and the online evaluation methodology with the same study participants, i.e. a within users experimental design. This paper presents empirical evidence that the ranking of algorithms based on offline accuracy measurements clearly contradicts the results from the online study with the same set of users. Thus the external validity of the most commonly applied evaluation methodology is not guaranteed.

Non-stationary continuous time Bayesian networks are introduced. They allow the parents set of each node to change over continuous time. Three settings are developed for learning non-stationary continuous time Bayesian networks from data: known transition times, known number of epochs and unknown number of epochs. A score function for each setting is derived and the corresponding learning algorithm is developed. A set of numerical experiments on synthetic data is used to compare the effectiveness of non-stationary continuous time Bayesian networks to that of non-stationary dynamic Bayesian networks. Furthermore, the performance achieved by non-stationary continuous time Bayesian networks is compared to that achieved by state-of-the-art algorithms on four real-world datasets, namely drosophila, saccharomyces cerevisiae, songbird and macroeconomics.

BACKGROUND: Dynamic aspects of gene regulatory networks are typically investigated by measuring system variables at multiple time points. Current state-of-the-art computational approaches for reconstructing gene networks directly build on such data, making a strong assumption that the system evolves in a synchronous fashion at fixed points in time. However, nowadays omics data are being generated with increasing time course granularity. Thus, modellers now have the possibility to represent the system as evolving in continuous time and to improve the models' expressiveness. RESULTS: Continuous time Bayesian networks are proposed as a new approach for gene network reconstruction from time course expression data. Their performance was compared to two state-of-the-art methods: dynamic Bayesian networks and Granger causality analysis. On simulated data, the methods comparison was carried out for networks of increasing size, for measurements taken at different time granularity densities and for measurements unevenly spaced over time. Continuous time Bayesian networks outperformed the other methods in terms of the accuracy of regulatory interactions learnt from data for all network sizes. Furthermore, their performance degraded smoothly as the size of the network increased. Continuous time Bayesian networks were significantly better than dynamic Bayesian networks for all time granularities tested and better than Granger causality for dense time series. Both continuous time Bayesian networks and Granger causality performed robustly for unevenly spaced time series, with no significant loss of performance compared to the evenly spaced case, while the same did not hold true for dynamic Bayesian networks. The comparison included the IRMA experimental datasets which confirmed the effectiveness of the proposed method. Continuous time Bayesian networks were then applied to elucidate the regulatory mechanisms controlling murine T helper 17 (Th17) cell differentiation and were found to be effective in discovering well-known regulatory mechanisms, as well as new plausible biological insights. CONCLUSIONS: Continuous time Bayesian networks were effective on networks of both small and large size and were particularly feasible when the measurements were not evenly distributed over time. Reconstruction of the murine Th17 cell differentiation network using continuous time Bayesian networks revealed several autocrine loops, suggesting that Th17 cells may be auto regulating their own differentiation process.

In this paper we consider the case of the ICT adoption and use in an agriculture cluster in Lombardy, a northern region of Italy. At the state of the art, relationships among key factors of adoption and use of ICT in agriculture area received little attention by the academic literature. Thus, in this paper we aim to identify a research model in order to provide evidence of four different research questions concerning the determining factors for ICT adoption. The proposed case study reports and discusses the results obtained by analysing data from a survey of about 600 agricultural farms. Finally, Belief Bayesian Networks (BBNs) are used to analyse the complex influence relationships detected between research variables.

BACKGROUND: Molecular dynamics (MD) simulations are powerful tools to investigate the conformational dynamics of proteins that is often a critical element of their function. Identification of functionally relevant conformations is generally done clustering the large ensemble of structures that are generated. Recently, Self-Organising Maps (SOMs) were reported performing more accurately and providing more consistent results than traditional clustering algorithms in various data mining problems. We present a novel strategy to analyse and compare conformational ensembles of protein domains using a two-level approach that combines SOMs and hierarchical clustering. RESULTS: The conformational dynamics of the α-spectrin SH3 protein domain and six single mutants were analysed by MD simulations. The Cα's Cartesian coordinates of conformations sampled in the essential space were used as input data vectors for SOM training, then complete linkage clustering was performed on the SOM prototype vectors. A specific protocol to optimize a SOM for structural ensembles was proposed: the optimal SOM was selected by means of a Taguchi experimental design plan applied to different data sets, and the optimal sampling rate of the MD trajectory was selected. The proposed two-level approach was applied to single trajectories of the SH3 domain independently as well as to groups of them at the same time. The results demonstrated the potential of this approach in the analysis of large ensembles of molecular structures: the possibility of producing a topological mapping of the conformational space in a simple 2D visualisation, as well as of effectively highlighting differences in the conformational dynamics directly related to biological functions. CONCLUSIONS: The use of a two-level approach combining SOMs and hierarchical clustering for conformational analysis of structural ensembles of proteins was proposed. It can easily be extended to other study cases and to conformational ensembles from other sources.

In this paper the problem of performing external validation of the semantic coherence of topic models is considered. The Fowlkes-Mallows index, a known clustering validation metric, is generalized for the case of overlapping partitions and multi-labeled collections, thus making it suitable for validating topic modeling algorithms. In addition, we propose new probabilistic metrics inspired by the concepts of recall and precision. The proposed metrics also have clear probabilistic interpretations and can be applied to validate and compare other soft and overlapping clustering algorithms. The approach is exemplified by using the Reuters-21578 multi-labeled collection to validate LDA models, then using Monte Carlo simulations to show the convergence to the predicted results. Additional statistical evidence is provided to better understand the relation of the metrics presented.

An algorithm for the automatic labeling of topics accordingly to a hierarchy is presented. Its main ingredients are a set of similarity measures and a set of topic labeling rules. The labeling rules are specifically designed to find the most agreed labels between the given topic and the hierarchy. The hierarchy is obtained from the Google Directory service, extracted via an ad-hoc developed software procedure and expanded through the use of the OpenOffice English Thesaurus. The performance of the proposed algorithm is investigated by using a document corpus consisting of 33,801 documents and a dictionary consisting of 111,795 words. The results are encouraging, while particularly interesting and significant labeling cases emerged.

A software system for topic extraction and automatic document classification is presented. Given a set of documents, the system automatically extracts the mentioned topics and assists the user to select their optimal number. The user-validated topics are exploited to build a model for multi-label document classification. While topic extraction is performed by using an optimized implementation of the Latent Dirichlet Allocation model, multi-label document classification is performed by using a specialized version of the Multi-Net Naive Bayes model. The performance of the system is investigated by using 10,056 documents retrieved from the WEB through a set of queries formed by exploiting the Italian Google Directory. This dataset is used for topic extraction while an independent dataset, consisting of 1,012 elements labeled by humans, is used to evaluate the performance of the Multi-Net Naive Bayes model. The results are satisfactory, with precision being consistently better than recall for the labels associated with the four most frequent topics.