Alexander Karlsson

An empirical comparison of Bayesian and credal networks for dependable high-level information fusion

Bayesian networks are often proposed as a method for high-level information fusion. However, a Bayesian network relies on strong assumptions about the underlying probabilities. In many cases it is not realistic to require such precise probability assessments. We show that there exists a significant set of problems where credal networks outperform Bayesian networks, thus enabling more dependable decision making for this type of problems. A credal network is a graphical probabilistic method that utilizes sets of probability distributions, e.g., interval probabilities, for representation of belief. Such a representation allows one to properly express epistemic uncertainty, i.e., uncertainty that can be reduced if more information becomes available. Since reducing uncertainty has been proposed as one of the main goals of information fusion, the ability to represent epistemic uncertainty becomes an important aspect in all fusion applications.

On evidential combination rules for ensemble classifiers

Ensemble classifiers are known to generally perform better than each individual classifier of which they consist. One approach to classifier fusion is to apply Shaferpsilas theory of evidence. While most approaches have adopted Dempsterpsilas rule of combination, a multitude of combination rules have been proposed. A number of combination rules as well as two voting rules are compared when used in conjunction with a specific kind of ensemble classifier, known as random forests, w.r.t. accuracy, area under ROC curve and Brier score on 27 datasets. The empirical evaluation shows that the choice of combination rule can have a significant impact on the performance for a single dataset, but in general the evidential combination rules do not perform better than the voting rules for this particular ensemble design. Furthermore, among the evidential rules, the associative ones appear to have better performance than the non-associative.

Evaluating precise and imprecise State-Based Anomaly detectors for maritime surveillance

We extend the State-Based Anomaly Detection approach by introducing precise and imprecise anomaly detectors using the Bayesian and credal combination operators, where evidences over time are combined into a joint evidence. We use imprecision in order to represent the sensitivity of the classification regarding an object being normal or anomalous. We evaluate the detectors on a real-world maritime dataset containing recorded AIS data and show that the anomaly detectors outperform previously proposed detectors based on Gaussian mixture models and kernel density estimators. We also show that our introduced anomaly detectors perform slightly better than the State-Based Anomaly Detection approach with a sliding window.

Modeling uncertainty in bibliometrics and information retrieval: an information fusion approach

Abstract We describe ongoing research where the aim is to apply recent results from the research field of information fusion to bibliometric analysis and information retrieval. We highlight the importance of ‘uncertainty’ within information fusion and argue that this concept is crucial also for bibliometrics and information retrieval. More specifically, we elaborate on three research strategies related to uncertainty: uncertainty management methods, explanation of uncertainty and visualization of uncertainty. We exemplify our strategies to the classical problem of author name disambiguation where we show how uncertainty can be modeled explained and visualized using information fusion. We show how an information seeker can benefit from tracing increases/decreases of uncertainty in the reasoning process. We also present how such changes can be explained for the information seeker through visualization techniques, which are employed to highlight the complexity involved in the process of modeling and managing uncertainty in bibliometric analysis. Finally we argue that a further integration of information fusion approaches in the research area of bibliometrics and information retrieval may results in new and fruitful venues of research.

An Empirical Comparison of Bayesian and Credal Set Theory for Discrete State Estimation

We are interested in whether or not there exist any advantages of utilizing credal set theory for the discrete state estimation problem. We present an experiment where we compare in total six different methods, three based on Bayesian theory and three on credal set theory. The results show that Bayesian updating performed on centroids of operand credal sets significantly outperforms the other methods. We analyze the result based on degree of imprecision, position of extreme points, and second-order distributions.

Decision Making with Hierarchical Credal Sets

We elaborate on hierarchical credal sets, which are sets of probability mass functions paired with second-order distributions. A new criterion to make decisions based on these models is proposed. This is achieved by sampling from the set of mass functions and considering the Kullback-Leibler divergence from the weighted center of mass of the set. We evaluate this criterion in a simple classification scenario: the results show performance improvements when compared to a credal classifier where the second-order distribution is not taken into account.

Uncertainty Levels of Second-Order Probability

Since second-order probability distributions assign probabilities to probabilities there is uncertainty on two levels. Although different types of uncertainty have been distinguished before and corresponding measures suggested, the distinction made here between first- and second-order levels of uncertainty has not been considered before. In this paper previously existing measures are considered from the perspective of first- and second-order uncertainty and new measures are introduced. We conclude that the concepts of uncertainty and informativeness needs to be qualified if used in a second-order probability context and suggest that from a certain point of view information can not be minimized, just shifted from one level to another

A study on class-specifically discounted belief for ensemble classifiers

Ensemble classifiers are known to generally perform better than their constituent classifiers. Whereas a lot of work has been focusing on the generation of classifiers for ensembles, much less attention has been given to the fusion of individual classifier outputs. One approach to fuse the outputs is to apply Shaferpsilas theory of evidence, which provides a flexible framework for expressing and fusing beliefs. However, representing and fusing beliefs is non-trivial since it can be performed in a multitude of ways within the evidential framework. In a previous article, we compared different evidential combination rules for ensemble fusion. The study involved a single belief representation which involved discounting (i.e., weighting) the classifier outputs with classifier reliability. The classifier reliability was interpreted as the classifierpsilas estimated accuracy, i.e., the percentage of correctly classified examples. However, classifiers may have different performance for different classes and in this work we assign the reliability of a classifier output depending on the class-specific reliability of the classifier. Using 27 UCI datasets, we compare the two different ways of expressing beliefs and some evidential combination rules. The result of the study indicates that there is indeed an advantage of utilizing class-specific reliability compared to accuracy in an evidential framework for combining classifiers in the ensemble design considered.

Imprecise Probability as an Approach to Improved Dependability in High-Level Information Fusion

The main goal of information fusion can be seen as improving human or automatic decision-making by exploiting diversities in information from multiple sources. High-level information fusion aims specifically at decision support regarding situations, often expressed as “achieving situation awareness”. A crucial issue for decision making based on such support is trust that can be defined as “accepted dependence”, where dependence or dependability is an overall term for many other concepts, e.g., reliability. This position paper reports on ongoing and planned research concerning imprecise probability as an approach to improved dependability in high-level information fusion. We elaborate on high-level information fusion from a generic perspective and a partial mapping from a taxonomy of dependability to high-level information fusion is presented. Three application domains: defense, manufacturing, and precision agriculture, where experiments are planned to be implemented are depicted. We conclude that high-level information fusion as an application-oriented research area, where precise probability (Bayesian theory) is commonly adopted, provides an excellent evaluation ground for imprecise probability.

Mode tracking using multiple data streams

Most existing work in information fusion focuses on combining information with well-defined meaning towards a concrete, pre-specified goal. In contradistinction, we instead aim for autonomous disco ...