Abstract. We present a new algorithm for Bayesian network structure learning, called Max-Min Hill-Climbing (MMHC). The algorithm combines ideas from local learning, constraint-based, and search-and-score techniques in a principled and effective way. It first reconstructs the skeleton of a Bayesian network and then performs a Bayesian-scoring greedy hill-climbing search to orient the edges. In our extensive empirical evaluation MMHC outperforms on average and in terms of various metrics several prototypical and state-of-the-art algorithms, namely the PC, Sparse Candidate, Three Phase Dependency Analysis, Optimal Reinsertion, Greedy Equivalence Search, and Greedy Search. These are the first empirical results simultaneously comparing most of the major Bayesian network algorithms against each other. MMHC offers certain theoretical advantages, specifically over the Sparse Candidate algorithm, corroborated by our experiments. MMHC and detailed results of our study are publicly available at

BDeu marginal likelihood score is a popular model selection criterion for selecting a Bayesian network structure based on sample data. This non-informative scoring criterion assigns same score for network structures that encode same independence statements. However, before applying the BDeu score, one must determine a single parameter, the equivalent sample size α. Unfortunately no generally accepted rule for determining the α parameter has been suggested. This is disturbing, since in this paper we show through a series of concrete experiments that the solution of the network structure optimization problem is highly sensitive to the chosen α parameter value. Based on these results, we are able to give explanations for how and why this phenomenon happens, and discuss ideas for solving this problem. 1

In Bayesian analysis of multi-way contingency tables, the selection of a prior distribution for either the log-linear parameters or the cell probabilities parameters is a major challenge. In this paper, we define a flexible family of conjugate priors for the wide class of discrete hierarchical log-linear models, which includes the class of graphical models. These priors are defined as the Diaconis–Ylvisaker conjugate priors on the log-linear parameters subject to “baseline constraints ” under multinomial sampling. We also derive the induced prior on the cell probabilities and show that the induced prior is a generalization of the hyper Dirichlet prior. We show that this prior has several desirable properties and illustrate its usefulness by identifying the most probable decomposable, graphical and hierarchical log-linear models for a six-way contingency table. 1. Introduction. We

In the Bayesian approach to structure learning of graphical models, the equivalent sample size (ESS) in the Dirichlet prior over the model parameters was recently shown to have an important effect on the maximum-a-posteriori estimate of the Bayesian network structure. In our first contribution, we theoretically analyze the case of large ESS-values, which complements previous work: among other results, we find that the presence of an edge in a Bayesian network is favoured over its absence even if both the Dirichlet prior and the data imply independence, as long as the conditional empirical distribution is notably different from uniform. In our second contribution, we focus on realistic ESS-values, and provide an analytical approximation to the ‘optimal ’ ESS-value in a predictive sense (its accuracy is also validated experimentally): this approximation provides an understanding as to which properties of the data have the main effect determining the ‘optimal ’ ESS-value. 1

Classical approaches used to learn Bayesian network structure from data have disadvantages in terms of complexity and lower accuracy of their results. However, a recent empirical study has shown that a hybrid algorithm improves sensitively accuracy and speed: it learns a skeleton with an independency test (IT) approach and constrains on the directed acyclic graphs (DAG) considered during the search-and-score phase. Subsequently, we theorize the structural constraint by introducing the concept of super-structure S, which is an undirected graph that restricts the search to networks whose skeleton is a subgraph of S. We develop a super-structure constrained optimal search (COS): its time complexity is upper bounded by O(γm n), where γm < 2 depends on the maximal degree m of S. Empirically, complexity depends on the average degree ˜m and sparse structures allow larger graphs to be calculated. Our algorithm is faster than an optimal search by several orders and even finds more accurate results when given a sound super-structure. Practically, S can be approximated by IT approaches; significance level of the tests controls its sparseness, enabling to control the trade-off between speed and accuracy. For incomplete super-structures, a greedily post-processed version (COS+) still enables to significantly outperform other heuristic searches. Keywords: subset Bayesian networks, structure learning, optimal search, super-structure, connected 1.

We present an approach to discretizing multivariate continuous data while learning the structure of a graphical model. We derive the joint scoring function from the principle of predictive accuracy, which inherently ensures the optimal trade-off between goodness of fit and model complexity (including the number of discretization levels). Using the so-called finest grid implied by the data, our scoring function depends only on the number of data points in the various discretization levels. Not only can it be computed efficiently, but it is also invariant under monotonic transformations of the continuous space. Our experiments show that the discretization method can substantially impact the resulting graph structure. 1

BDeu marginal likelihood score is a popular model selection criterion for selecting a Bayesian network structure based on sample data. This non-informative scoring criterion assigns same score for network structures that encode same independence statements. However, before applying the BDeu score, one must determine a single parameter, the equivalent sample size α. Unfortunately no generally accepted rule for determining the α parameter has been suggested. This is disturbing, since in this paper we show through a series of concrete experiments that the solution of the network structure optimization problem is highly sensitive to the chosen α parameter value. Based on these results, we are able to give explanations for how and why this phenomenon happens, and discuss ideas for solving this problem.

Abstract—Transient faults are a major concern in today’s deep sub-micron semiconductor technology. These faults are rare but they have been known to cause catastrophic system-level failures. Transient errors often occur due to physical effects on deployed systems and hence, diagnosis of transient errors must be performed over manufactured chips or systems assembled from black-box components where arbitrary instrumentation of the system is not possible and hence, the system state is only partially observable. Further, these systems are often composed of components that are third party IP which further adds opaqueness to the system. In this paper, we propose a probabilistic approach to localize transient faults in space and time for such partially observable systems. From a set of correct traces and a failure trace, we seek to locate the faulty component and the cycle of operation at which the fault occurred. Our technique uses correct system traces over monitored components of the system to learn a dynamic Bayesian network (DBN) summarizing the temporal dependencies across the monitored components. This DBN is augmented with different error hypotheses allowed by the fault model. The most probable explanation (MPE) among these hypotheses corresponds to the most likely location of the error. We evaluated the effectiveness of our technique on a set of ISCAS89 benchmarks and a router design used in on-chip networks in a multi-core design. I.

A framework for the analysis of model selection issues is presented. The framework separates model selection into two dimensions: the model-complexity dimension and the model-space dimension. The model-complexity dimension pertains to how the complexity of a single model interacts with its scoring by standard evaluation measures. The model-space dimension pertains to the interpretation of the totality of evaluation scores obtained. Central to the analysis is the concept of evaluation coherence, a property which requires that a measure not produce misleading model evaluations. Of particular interest is whether model evaluation measures are misled by model complexity. Several common evaluation measures — apparent error rate, the BD metric, and MDL scoring — are analyzed, and each is found to lack complexity coherence. These results are used to consider arguments for and against the Occam razor paradigm as it pertains to overfit avoidance in model selection, and also to provide an abstract analysis of what the literature refers to as oversearch. 1.

We propose an information-theoretic approach for predictive modeling with Bayesian networks. Our approach is based on the minimax optimal Normalized Maximum Likelihood (NML) distribution, motivated by the MDL principle. In particular, we present a parameter learning method which, together with a previously introduced NML-based model selection criterion, provides a way to construct highly predictive Bayesian network models from data. The method is parameterfree and robust, unlike the currently popular Bayesian marginal likelihood approach which has been shown to be sensitive to the choice of prior hyperparameters. Empirical tests show that the proposed method compares favorably with the Bayesian approach in predictive tasks. 1