The need for feasible inference in Probabilistic Graphical Models (PGMs) has lead to tractable models like Sum-Product Networks (SPNs). Their highly expressive power and their ability to provide exact and tractable inference make them very attractive for several real world applications, from computer vision to NLP. Recently, great attention around SPNs has focused on structure learning, leading to different algorithms being able to learn both the network and its parameters from data. Here, we enhance one of the best structure learner, LearnSPN, aiming to improve both the structural quality of the learned networks and their achieved likelihoods. Our algorithmic variations are able to learn simpler, deeper and more robust networks. These results have been obtained by exploiting some insights in the building process done by LearnSPN, by hybridizing the network adopting tree-structured models as leaves, and by blending bagging estimations into mixture creation. We prove our claims by empirically evaluating the learned SPNs on several benchmark datasets against other competitive SPN and PGM structure learners

No abstract

Networks had an increasing impact on modern life since network cybersecurity has become an important research field. Several machine learning techniques have been developed to build network intrusion detection systems for correctly detecting unforeseen cyber-attacks at the network-level. For example, deep artificial neural network architectures have recently achieved state-of-the-art results. In this paper a novel deep neural network architecture is defined, in order to learn flexible and effective intrusion detection models, by combining an unsupervised stage for multi-channel feature learning with a supervised one exploiting feature dependencies on cross channels. The aim is to investigate whether class-specific features of the network flows could be learned and added to the original ones in order to increase the model accuracy. In particular, in the unsupervised stage, two autoencoders are separately learned on normal and attack flows, respectively. As the top layer in the decoder of these autoencoders reconstructs samples in the same space as the input one, they could be used to define two new feature vectors allowing the representation of each network flow as a multi-channel sample. In the supervised stage, a multi-channel parametric convolution is adopted, in order to learn the effect of each channel on the others. In particular, as the samples belong to two different distributions (normal and attack flows), the samples labelled as normal should be more similar to the representation reconstructed with the normal autoencoder than that of the attack one, and viceversa. This expected dependency will be exploited to better disentangle the differences between normal and attack flows. The proposed neural network architecture leads to better predictive accuracy when compared to competitive intrusion detection architectures on three benchmark datasets.

The coalition structure generation problem represents an active research area in multi-agent systems. A coalition structure is defined as a partition of the agents involved in a system into disjoint coalitions. The problem of finding the optimal coalition structure is NP-complete. In order to find the optimal solution in a combinatorial optimization problem it is theoretically possible to enumerate the solutions and evaluate each. But this approach is infeasible since the number of solutions often grows exponentially with the size of the problem. In this paper we present a greedy adaptive search procedure (GRASP) to efficiently search the space of coalition structures in order to find an optimal one

In this paper we propose to apply the Information Bottleneck (IB) approach to the sub-class of Statistical Relational Learning (SRL) languages that are reducible to Bayesian networks. When the resulting networks involve hidden variables, learning these languages requires the use of techniques for learning from incomplete data such as the Expectation Maximization (EM) algorithm. Recently, the IB approach was shown to be able to avoid some of the local maxima in which EM can get trapped when learning with hidden variables. Here we present the algorithm Relational Information Bottleneck (RIB) that learns the parameters of SRL languages reducible to Bayesian Networks. In particular, we present the specialization of RIB to a language belonging to the family of languages based on the distribution semantics, Logic Programs with Annotated Disjunction (LPADs). This language is prototypical for such a family and its equivalent Bayesian networks contain hidden variables. RIB is evaluated on the IMDB, Cora and artificial datasets and compared with LeProbLog, EM, Alchemy and PRISM. The experimental results show that RIB has good performances especially when some logical atoms are unobserved. Moreover, it is particularly suitable when learning from interpretations that share the same Herbrand base.

This paper describes GRAPE, an expert component for a scientific Conference Management System (CMS), to automatically assign reviewers to papers, one of the most difficult processes of conference management. In the current practice, this is typically done by a manual and time-consuming procedure, with a risk of bad quality results due to the many aspects and parameters to be taken into account, and on their interrelationships and (often contrasting) requirements. The proposed rule-based system was evaluated on real conference datasets obtaining good results when compared to the handmade ones, both in terms of quality of the assignments, and of reduction in execution time

Sum-Product Networks (SPNs) are deep tractable probabilistic models by which several kinds of inference queries can be answered exactly and in a tractable time. They have been largely used as black box density estimators, assessed by comparing their likelihood scores on different tasks. In this paper we explore and exploit the inner representations learned by SPNs. By taking a closer look at the inner workings of SPNs, we aim to better understand what and how meaningful the representations they learn are, as in a classic Representation Learning framework. We firstly propose an interpretation of SPNs as Multi-Layer Perceptrons, we then devise several criteria to extract representations from SPNs and finally we empirically evaluate them in several (semi-)supervised tasks showing they are competitive against classical feature extractors like RBMs, DBNs and deep probabilistic autoencoders, like MADEs and VAEs.

No abstract