Catálogo de publicaciones - libros

Given ordered classes, one is not only concerned to maximize the classification accuracy, but also to minimize the distances between the actual and the predicted classes. This paper offers an organized study on the various methodologies that have tried to handle this problem and presents an experimental study of these methodologies with the proposed local ordinal technique, which locally converts the original ordinal class problem into a set of binary class problems that encode the ordering of the original classes. The paper concludes that the proposed technique can be a more robust solution to the problem because it minimizes the distances between the actual and the predicted classes as well as improves the classification accuracy.

Many students who enrol in the undergraduate program on informatics at the Hellenic Open University (HOU) fail the introductory course exams and drop out. We analyze their academic performance, derive short rules that explain success or failure in the exams and use the accuracy of these rules to reflect on specific tutoring practices that could enhance success.

Decision trees are commonly employed as data classifiers in various research fields, but also in real-world application domains. In the fuzzy neural framework, decision trees can offer valuable assistance in determining a proper initial system structure, which means not only feature selection, but also rule extraction and organization. This paper proposes a synergistic model that combines the advantages of a subsethood-product neural fuzzy inference system and a CART algorithm, in order to create a novel architecture and generate fuzzy rules of the form “IF - THEN IF”, where the first “IF” concerns the primary attributes and the second “IF” the secondary attributes of the given dataset as defined by our method. The resulted structure eliminates certain drawbacks of both techniques and produces a compact, comprehensible and efficient rulebase. Experiments in benchmark classification tasks prove that this method does not only reduce computational cost, but it also maintains performance at high levels, offering fast and accurate processing during realtime operations.

In this paper, we discuss the visualization of multidimensional data. A well-known procedure for mapping data from a high-dimensional space onto a lower-dimensional one is Sammon’s mapping. The algorithm is oriented to minimize the projection error. We investigate an unsupervised backpropagation algorithm to train a multilayer feed-forward neural network (SAMANN) to perform the Sammon’s nonlinear projection. Sammon mapping has a disadvantage. It lacks generalization, which means that new points cannot be added to the obtained map without recalculating it. The SAMANN network offers the generalization ability of projecting new data, which is not present in the original Sammon’s projection algorithm. Retraining of the network when the new data points appear has been analyzed in this paper.

The paper presents a hybrid adaptation method that combines a knowledge-based approach with reinforcement learning and a simulated annealing technique, and is applied in the navigation of an educational mobile robot. The experimental results of simulations showed a good behaviour of the robot when doing an adaptive navigation in a dynamic environment by using the proposed hybrid method.

This paper describes an information system, which classifies web pages in specific categories according to a proposed relevance feedback mechanism. The proposed relevance feedback mechanism is called Balanced Relevance Weighting Mechanism — BRWM and uses the proportion of the already relevant categorized information amount for feature classification. Experimental measurements over an e-commerce framework, which describes the fundamental phases of web commercial transactions verified the robustness of using the mechanism on real data. Except from revealing the accomplished sequences in a web commerce transaction, the system can be used as an assistant and consultation tool for classification purposes. In addition, BRWM was compared with a similar relevance feedback mechanism from the literature over the established corpus of Reuters-21578 text categorization test collection, presenting promising results.

Linear regression and regression tree models are among the most known regression models used in the machine learning community and recently many researchers have examined their sufficiency in ensembles. Although many methods of ensemble design have been proposed, there is as yet no obvious picture of which method is best. One notable successful adoption of ensemble learning is the distributed scenario. In this work, we propose an efficient distributed method that uses different subsets of the same training set with the parallel usage of an averaging methodology that combines linear regression and regression tree models. We performed a comparison of the presented ensemble with other ensembles that use either the linear regression or the regression trees as base learner and the performance of the proposed method was better in most cases.

Over the last few years, argumentation has been gaining increasing importance in several Al-related areas, mainly as a vehicle for facilitating rationally justifiable decision making when handling incomplete and potentially inconsistent information. In this setting, user support systems can rely on argumentation techniques to automatize reasoning and decision making in several situations such as the handling of complex policies or managing change in dynamic environments. This paper presents a generic argument-based approach to characterize user support systems, in which knowledge representation and inference are captured in terms of Defeasible Logic Programming, a general-purpose defeasible argumentation formalism based on logic programming. We discuss a particular application which has emerged as an instance of this approach oriented towards providing user decision support for web search.

This paper discusses platform independent conceptual modeling of a knowledge intensive application, focusing on the use of knowledge-based systems (KBS) in the context of model-driven engineering. An extension to the Unified Modeling Language (UML) for knowledge modeling is presented based on the profiling extension mechanism of UML. The UML profile discussed in this paper has been successfully captured in a Meta-Object-Facility (MOF) based UML tool — the executable Modeling Framework (XMF). The example is that of modeling a knowledge-based system for the Ulcer Clinical Practical Guidelines (CPG) Recommendations. It demonstrates the use of the profile, with the prototype system implemented in the Java Expert System Shell (JESS).

Building coalitions between autonomous domains and managing the negotiation process between multiple security policies in a multi-domain environment is a challenging task. The negotiation process requires efficient modeling methods for the determination of secure access states and demands support from automated tools aiming to support administrators and to minimize human intervention; thus making the whole process more efficient and less error-prone. In this paper we define a framework that enables the representation of policy merging between autonomous domains, as a constraint satisfaction problem, while remaining neutral in regard to the policy language. Role and permission hierarchies are modeled using the constraint programming formalism. Policy mappings are utilized in order to enable cross-organizational role assignment. Further optimization on policy mappings is achieved by casting the problem to a partially ordered multi-criteria shortest path problem.