Denis Deratani Mauá

Evaluating credal classifiers by utility-discounted predictive accuracy

Predictions made by imprecise-probability models are often indeterminate (that is, set-valued). Measuring the quality of an indeterminate prediction by a single number is important to fairly compare different models, but a principled approach to this problem is currently missing. In this paper we derive, from a set of assumptions, a metric to evaluate the predictions of credal classifiers. These are supervised learning models that issue set-valued predictions. The metric turns out to be made of an objective component, and another that is related to the decision-makers degree of risk aversion to the variability of predictions. We discuss when the measure can be rendered independent of such a degree, and provide insights as to how the comparison of classifiers based on the new measure changes with the number of predictions to be made. Finally, we make extensive empirical tests of credal, as well as precise, classifiers by using the new metric. This shows the practical usefulness of the metric, while yielding a first insightful and extensive comparison of credal classifiers.

Solving limited memory influence diagrams

We present a new algorithm for exactly solving decision making problems represented as influence diagrams. We do not require the usual assumptions of no forgetting and regularity; this allows us to solve problems with simultaneous decisions and limited information. The algorithm is empirically shown to outperform a state-of-the-art algorithm on randomly generated problems of up to 150 variables and 1064 solutions. We show that these problems are NP-hard even if the underlying graph structure of the problem has low treewidth and the variables take on a bounded number of states, and that they admit no provably good approximation if variables can take on an arbitrary number of states.

Trading off Speed and Accuracy in Multilabel Classification

In previous work, we devised an approach for multilabel classification based on an ensemble of Bayesian networks. It was characterized by an efficient structural learning and by high accuracy. Its shortcoming was the high computational complexity of the MAP inference, necessary to identify the most probable joint configuration of all classes. In this work, we switch from the ensemble approach to the single model approach. This allows important computational savings. The reduction of inference times is exponential in the difference between the treewidth of the single model and the number of classes. We adopt moreover a more sophisticated approach for the structural learning of the class subgraph. The proposed single models outperforms alternative approaches for multilabel classification such as binary relevance and ensemble of classifier chains.

On the complexity of solving polytree-shaped limited memory influence diagrams with binary variables

Influence diagrams are intuitive and concise representations of structured decision problems. When the problem is non-Markovian, an optimal strategy can be exponentially large in the size of the diagram. We can avoid the inherent intractability by constraining the size of admissible strategies, giving rise to limited memory influence diagrams. A valuable question is then how small do strategies need to be to enable efficient optimal planning. Arguably, the smallest strategies one can conceive simply prescribe an action for each time step, without considering past decisions or observations. Previous work has shown that finding such optimal strategies even for polytree-shaped diagrams with ternary variables and a single value node is NP-hard, but the case of binary variables was left open. In this paper we address such a case, by first noting that optimal strategies can be obtained in polynomial time for polytree-shaped diagrams with binary variables and a single value node. We then show that the same problem is NP-hard if the diagram has multiple value nodes. These two results close the fixed-parameter complexity analysis of optimal strategy selection in influence diagrams parametrized by the shape of the diagram, the number of value nodes and the maximum variable cardinality.

Algorithms for Hidden Markov Models with Imprecisely Specified Parameters

Hidden Markov models (HMMs) are widely used models for sequential data. As with other probabilistic models, they require the specification of local conditional probability distributions, which can be too difficult and error-prone, especially when data are scarce or costly to acquire. The imprecise HMM (iHMM) generalizes HMMs by allowing the quantification to be done by sets of, instead of single, probability distributions. iHMMs have the ability to suspend judgment when there is not enough statistical evidence, and can serve as a sensitivity analysis tool for standard non-stationary HMMs. In this paper, we formalize iHMMs and develop efficient inference algorithms to address standard HMM usage such as the computation of likelihoods and most probable explanations. Experiments with real data show that iHMMs produce more reliable inferences without compromising efficiency.

DL-Lite Bayesian Networks: A Tractable Probabilistic Graphical Model

The construction of probabilistic models that can represent large systems requires the ability to describe repetitive and hierarchical structures. To do so, one can resort to constructs from description logics. In this paper we present a class of relational Bayesian networks based on the popular description logic DL-Lite. Our main result is that, for this modeling language, marginal inference and most probable explanation require polynomial effort. We show this by reductions to edge covering problems, and derive a result of independent interest; namely, that counting edge covers in a particular class of graphs requires polynomial effort.

The Complexity of Plate Probabilistic Models

Plate-based probabilistic models combine a few relational constructs with Bayesian networks, so as to allow one to specify large and repetitive probabilistic networks in a compact and intuitive manner. In this paper we investigate the combined, data and domain complexity of plate models, showing that they range from polynomial to \(\#\mathsf {P}\)-complete to \(\#\mathsf {EXP}\)-complete.

Equivalences between maximum a posteriori inference in Bayesian networks and maximum expected utility computation in influence diagrams

Two important tasks in probabilistic reasoning are the computation of the maximum posterior probability of a given subset of the variables in a Bayesian network (MAP), and the computation of the maximum expected utility of a strategy in an influence diagram (MEU). Both problems are NPPP-hard to solve, and NP-hard to approximate when the treewidth of the underlying graph is bounded. Despite the similarities, researches on both problems have largely been conducted independently, with algorithmic solutions and insights designed for one problem not (trivially) transferable to the other one. In this work, we show constructively that these two problems are equivalent in the sense that any algorithm designed for one problem can be used to solve the other with small overhead. Moreover, the reductions preserve the boundedness of treewidth. Building on the known complexity of MAP on networks whose parameters are imprecisely specified, we show how to use the reductions to characterize the complexity of MEU when the parameters are set-valued. These equivalences extend the toolbox of either problem, and shall foster new insights into their solution. We show a polynomial-time reduction from MAP to MEU that preserves the boundedness of treewidth.We show a polynomial-time reduction from MEU to MAP that also preserves the boundedness of treewidth.We use the reduction from MAP to MEU to characterize the complexity of MEU in credal influence diagrams.Experiments show that reducing MEU to MAP can be an effective alternative for solving MEU problems.

Speeding Up k-Neighborhood Local Search in Limited Memory Influence Diagrams

Limited memory influence diagrams are graph-based models that describe decision problems with limited information, as in the case of teams and agents with imperfect recall. Solving a (limited memory) influence diagram is an NP-hard problem, often approached through local search. In this paper we investigate algorithms for k-neighborhood local search. We show that finding a k-neighbor that improves on the current solution is W[1]-hard and hence unlikely to be polynomial-time tractable. We then develop fast schema to perform approximate k-local search; experiments show that our methods improve on current local search algorithms both with respect to time and to accuracy.

The Finite Model Theory of Bayesian Networks: Descriptive Complexity.

We adapt the theory of descriptive complexity to encompass Bayesian networks, so as to quantify the expressivity of Bayesian network specifications based on predicates and quantifiers. We show that Bayesian network specifications that employ firstorder quantification capture the complexity class PP; by allowing quantification over predicates, the resulting Bayesian network specifications capture each class in the hierarchy PP ...NP , a result that does not seem to have equivalent in the literature.1