Formal explainability is an emerging field that aims to provide mathematically guaranteed explanations for the predictions made by machine learning models. Recent work in this area focuses on computing “probabilistic explanations” for the predictions made by classifiers based on specific data instances. The goal of this paper is to extend the concept of probabilistic explanations to the regression setting, treating the target regressor as a black box function. The class of probabilistic explanations consists of linear functions that meet a sparsity constraint, alongside a hyperplane constraint defined for the data instance being explained. While minimizing the precision error of such explanations is generally $\text{NP}^{\text{PP}}$-hard, we demonstrate that it can be approximated by substituting the precision measure with a fidelity measure. Optimal explanations based on this fidelity objective can be effectively approached using Mixed Integer Programming (MIP). Moreover, we show that for certain distributions used to define the precision measure, explanations with approximation guarantees can be computed in polynomial time using a variant of Iterative Hard Thresholding (IHT). Experiments conducted on various datasets indicate that both the MIP and IHT approaches outperform the state-of-the-art LIME and MAPLE explainers.

Explaining in accurate and intelligible terms the predictions made by classifiers is a key challenge of eXplainable Artificial Intelligence (XAI). To this end, an abductive explanation for the predicted label of some data instance is a subset-minimal collection of features such that the restriction of the instance to these features is sufficient to determine the prediction. However, due to cognitive limitations, abductive explanations are often too large to be interpretable. In those cases, we need to reduce the size of abductive explanations, while still determining the predicted label with high probability. In this paper, we show that finding such probabilistic explanations is NP-hard, even for decision trees. In order to circumvent this issue, we investigate the approximability of probabilistic explanations through the lens of supermodularity. We examine both greedy descent and greedy ascent approaches for supermodular minimization, whose approximation guarantees depend on the curvature of the “unnormalized” error function that evaluates the precision of the explanation. Based on various experiments for explaining decision tree predictions, we show that our greedy algorithms provide an efficient alternative to the state-of-the-art constraint optimization method.

Random forests have long been considered as powerful model ensembles in machine learning. By training multiple decision trees, whose diversity is fostered through data and feature subsampling, the resulting random forest can lead to more stable and reliable predictions than a single decision tree. This however comes at the cost of decreased interpretability: while decision trees are often easily interpretable, the predictions made by random forests are much more difficult to understand, as they involve a majority vote over multiple decision trees. In this paper, we examine different types of reasons that explain "why" an input instance is classified as positive or negative by a Boolean random forest. Notably, as an alternative to prime-implicant explanations taking the form of subset-minimal implicants of the random forest, we introduce majoritary reasons which are subset-minimal implicants of a strict majority of decision trees. For these abductive explanations, the tractability of the generation problem (finding one reason) and the optimization problem (finding one minimum-sized reason) are investigated. Unlike prime-implicant explanations, majoritary reasons may contain redundant features. However, in practice, prime-implicant explanations - for which the identification problem is DP-complete - are slightly larger than majoritary reasons that can be generated using a simple linear-time greedy algorithm. They are also significantly larger than minimum-sized majoritary reasons which can be approached using an anytime Partial MaxSAT algorithm.

Abductive explanations take a central place in eXplainable Artificial Intelligence (XAI) by clarifying with few features the way data instances are classified. However, instances may have exponentially many minimum-size abductive explanations, and this source of complexity holds even for ``intelligible'' classifiers, such as decision trees. When the number of such abductive explanations is huge, computing one of them, only, is often not informative enough. Especially, better explanations than the one that is derived may exist. As a way to circumvent this issue, we propose to leverage a model of the explainee, making precise her / his preferences about explanations, and to compute only preferred explanations. In this paper, several models are pointed out and discussed. For each model, we present and evaluate an algorithm for computing preferred majoritary reasons, where majoritary reasons are specific abductive explanations suited to random forests. We show that in practice the preferred majoritary reasons for an instance can be far less numerous than its majoritary reasons.

In constraint satisfaction problems, the variable ordering heuristic takes a central place by selecting the variables to branch on during backtrack search. As many hand-crafted branching heuristics have been proposed in the literature, a key issue is to identify, from a pool of candidate heuristics, which one is the best for solving a given constraint satisfaction task. Based on the observation that modern constraint solvers are using restart sequences, the best heuristic identification problem can be cast in the context of multi-armed bandits as a non-stochastic best arm identification problem. Namely, during each run of some given restart sequence, the bandit algorithm selects a branching heuristic and receives a reward for this heuristic before proceeding to the next run. The goal is to identify the best heuristic using few runs, and without any stochastic assumption about the constraint solver. In this study, we propose an adaptive variant of Successive Halving that exploits Luby's universal restart sequence. We analyze the convergence of this bandit algorithm in the non-stochastic setting, and we demonstrate its empirical effectiveness on various constraint satisfaction benchmarks.

In this paper, we investigate the computational intelligibility of Boolean classifiers, characterized by their ability to answer XAI queries in polynomial time. The classifiers under consideration are decision trees, DNF formulae, decision lists, decision rules, tree ensembles, and Boolean neural nets. Using 9 XAI queries, including both explanation queries and verification queries, we show the existence of large intelligibility gap between the families of classifiers. On the one hand, all the 9 XAI queries are tractable for decision trees. On the other hand, none of them is tractable for DNF formulae, decision lists, random forests, boosted decision trees, Boolean multilayer perceptrons, and binarized neural networks.

In constraint-based applications, the user is often required to be an expert as, for a given problem instance, many parameters of the used solver must be manually tuned to improve its efficiency. Clearly, this background knowledge burdens the spread of constraint programming technology to non-expert users. In order to alleviate this issue, the idea of "autonomous" constraint solving is to adjust the solver parameters and to efficiently handle any problem instance without manual tuning. Notably, the choice of the variable ordering heuristic can lead to drastically different performances. A key question arises then: how can we find the best variable ordering heuristic for a problem instance, given a set of available heuristics provided by the solver? To answer this question, we propose an algorithmic framework that combines multi-armed bandits and restarts. Each candidate heuristic is viewed as an arm, and the framework learns to estimate the best heuristic using a multi-armed bandit algorithm. The common mechanism of restarts is used to provide feedback for reinforcing the bandit algorithm. Based on a thorough experimental evaluation, we demonstrate that this framework is able to find the best heuristic for most problem instances; notably, it outperforms the state-of-the-art in terms of time and solved instances.

One of the key purposes of eXplainable AI (XAI) is to develop techniques for understanding predictions made by Machine Learning (ML) models and for assessing how much reliable they are. Several encoding schemas have recently been pointed out, showing how ML classifiers of various types can be mapped to Boolean circuits exhibiting the same input-output behaviours. Thanks to such mappings, XAI queries about classifiers can be delegated to the corresponding circuits. In this paper, we define new explanation and/or verification queries about classifiers. We show how they can be addressed by combining queries and transformations about the associated Boolean circuits. Taking advantage of previous results from the knowledge compilation map, this allows us to identify a number of XAI queries that are tractable provided that the circuit has been first turned into a compiled representation.

In this paper, we investigate the problem of sparse online linear classification in changing environments. We first analyze the tracking performance of standard online linear classifiers, which use gradient descent for minimizing the regularized hinge loss. The derived shifting bounds highlight the importance of choosing appropriate step sizes in the presence of concept drifts. Notably, we show that a better adaptability to concept drifts can be achieved using constant step sizes rather than the state-of-the-art decreasing step sizes. Based on these observations, we then propose a novel sparse approximated linear classifier, called sparse approximated linear classification (SALC), which uses a constant step size. In essence, SALC simply rounds small weights to zero for achieving sparsity and controls the truncation error in a principled way for achieving a low tracking regret. The degree of sparsity obtained by SALC is continuous and can be controlled by a parameter which captures the tradeoff between the sparsity of the model and the regret performance of the algorithm. Experiments on nine stationary data sets show that SALC is superior to the state-of-the-art sparse online learning algorithms, especially when the solution is required to be sparse; on seven groups of nonstationary data sets with various total shifting amounts, SALC also presents a good ability to track drifts. When wrapped with a drift detector, SALC achieves a remarkable tracking performance regardless of the total shifting amount.

In online optimization, the goal is to iteratively choose solutions from a decision space, so as to minimize the average cost over time. As long as this decision space is described by combinatorial constraints, the problem is generally intractable. In this paper, we consider the paradigm of compiling the set of combinatorial constraints into a deterministic and Decomposable Negation Normal Form (dDNNF) circuit, for which the tasks of linear optimization and solution sampling take linear time. Based on this framework, we provide efficient characterizations of existing combinatorial prediction strategies, with a particular attention to mirror descent techniques. These strategies are compared on several real-world benchmarks for which the set of Boolean constraints is preliminarily compiled into a dDNNF circuit.

Symmetry detection is a promising approach for reducing the search tree of games. In General Game Playing (GGP), where any game is compactly represented by a set of rules in the Game Description Language (GDL), the state-of-the-art methods for symmetry detection rely on a rule graph associated with the GDL description of the game. Though such rule-based symmetry detection methods can be applied to various tree search algorithms, they cover only a limited number of symmetries which are apparent in the GDL description. In this paper, we develop an alternative approach to symmetry detection in stochastic games that exploits constraint programming techniques. The minimax optimization problem in a GDL game is cast as a stochastic constraint satisfaction problem (SCSP), which can be viewed as a sequence of one-stage SCSPs. Minimax symmetries are inferred according to themicrostructure complement of these one-stage constraint networks. Based on a theoretical analysis of this approach, we experimentally show on various games that the recent stochastic constraint solver MAC-UCB, coupled with constraint-based symmetry detection, significantly outperforms the standard Monte Carlo Tree Search algorithms, coupled with rule-based symmetry detection. This constraint-driven approach is also validated by the excellent results obtained by our player during the last GGP competition.

We present and evaluate a new CNF encoding scheme for reducing probabilistic inference from a graphical model to weighted model counting. This new encoding scheme elaborates on the CNF encoding scheme ENC4 introduced by Chavira and Darwiche, and improves it by taking advantage of log encodings of the elementary variable/value assignments and of the implicit encoding of the most frequent probability value per conditional probability table. From the theory side, we show that our encoding scheme is faithful, and that for each input network, the CNF formula it leads to contains less variables and less clauses than the CNF formula obtained using ENC4. From the practical side, we show that the C2D compiler empowered by our encoding scheme performs in many cases significantly better than when ENC4 is used, or when the state-of-the-art ACE compiler is considered instead.

In this paper, symmetries are exploited for achieving significant space savings in a knowledge compilation perspective. More precisely, the languages FBDD and DDG of decision diagrams are extended to the languages Sym-FBDDX,Yand Sym-DDGX,Yof symmetry-driven decision diagrams, where X is a set of “symmetry-free” variables and Y is a set of “top” variables. Both the time efficiency and the space efficiency of Sym-FBDDX,Yand Sym-DDGX,Yare analyzed, in order to put those languages in the knowledge compilation map for propositional representations. It turns out that each of Sym-FBDDX,Yand Sym-DDGX,Ysatisfies CT (the model counting query). We prove that no propositional language over a set X∪Y of variables, satisfying both CO (the consistency query) and CD (the conditioning transformation), is at least as succinct as any of Sym-FBDDX,Yand Sym-DDGX,Yunless the polynomial hierarchy collapses. The price to be paid is that only a restricted form of conditioning and a restricted form of forgetting are offered by Sym-FBDDX,Yand Sym-DDGX,Y. Nevertheless, this proves sufficient for a number of applications, including configuration and planning. We describe a compiler targeting Sym-FBDDX,Yand Sym-DDGX,Yand give some experimental results on planning domains, highlighting the practical significance of these languages.

Learning discrete linear functions is a notoriously difficult challenge. In this paper, the learning task is cast as combinatorial optimization problem: given a set of positive and negative feature vectors in the Euclidean space, the goal is to find a discrete linear function that minimizes the cumulative hinge loss of this training set. Since this problem is NP-hard, we propose two simple rounding algorithms that discretize the fractional solution of the problem. Generalization bounds are derived for two important classes of binary-weighted linear functions, by establishing the Rademacher complexity of these classes and proving approximation bounds for rounding methods. These methods are compared on both synthetic and real-world data.

Online multiplicative weight-update learning algorithms, such as Winnow, have proven to behave remarkably for learning simple disjunctions with few relevant attributes. The aim of this paper is to extend the Winnow algorithm to more expressive concepts characterized by DNF formulas with few relevant rules. For such problems, the convergence of Winnow is still fast, since the number of mistakes increases only linearly with the number of attributes. Yet, the learner is confronted with an important computational barrier: during any prediction, it must evaluate the weighted sum of an exponential number of rules. To circumvent this issue, we convert the prediction problem into a Weighted Model Counting problem. The resulting algorithm, SharpNow, is an exact simulation of Winnow equipped with backtracking, caching, and decomposition techniques. Experiments on static and drifting problems demonstrate the performance of the algorithm in terms of accuracy and speed.

When different knowledge-based systems must cooperate to perform decision tasks that are beyond their individual capabilities, we are faced with the problem of combining knowledge in a multi-source environment. In particular, we are confronted with two main difficulties: the prospect of inconsistency, which arises when different knowledge bases are merged together, and the high computational complexity of reasoning with very large pools of combined information. In this paper, we define a formal framework which handles both aspects of consistency and tractability, and which is useful to specify knowledge retrievers. This framework tolerates inconsistency and enables a knowledge retriever to infer non-degenerative conclusions when conflicting viewpoints are combined. Furthermore, approximate reasoning is incorporated in order to perform efficient query answering using combined knowledge. Finally, a stepwise procedure is included for improving approximate answers and allowing their convergence to the right answer.