L. M. de Campos

Partial abductive inference in Bayesian belief networks using a genetic algorithm

Abductive inference in Bayesian belief networks is the process of generating the K most probable configurations given an observed evidence. When we are only interested in a subset of the network’s variables, this problem is called partial abductive inference. Both problems are NP-hard, and so exact computation is not always possible. This paper describes an approximate method based on genetic algorithms to perform partial abductive inference. We have tested the algorithm using the alarm network and from the experimental results we can conclude that the algorithm presented here is a good tool to perform this kind of probabilistic reasoning. ” 1999 Elsevier Science B.V. All rights reserved.

Building Bayesian network-based information retrieval systems

Bayesian networks are suitable models to deal with the information retrieval problem because they are appropriate tools to manage the intrinsic uncertainty with which this area is pervaded. In this paper we introduce several modifications to the previous works on this field adding new features and showing how a good retrieval effectiveness can be achieved by improving the quality of the Bayesian networks used in the model and tuning some of their parameters.

Distances between fuzzy measures through associated probabilities: some applications

In a recent paper a method to study fuzzy measures by means of certain sets of associated probabilities has been developed. In this paper, distances in the set of fuzzy measures are defined trough distances between the associated probabilities. Finally, some applications of these distances to measure the uncertainty and specificity of fuzzy measures as well as to approximate fuzzy measures, are considered.

Group Recommending: A methodological Approach based on Bayesian Networks

The problem of building recommender systems has attracted considerable attention in recent years, but most recommender systems are designed for recommending items for individuals. The aim of this paper is to automatically recommend and rank a list of new items to a group of users. The proposed model can be considered as a collaborative Bayesian network-based group recommender system, where the groups rates are computed from past voting patterns of other users with similar tastes. The use of Bayesian networks allows us to obtain an intuitive representation of the mechanisms that govern the relationships between the group members.

Independence Relationships in Possibility Theory and Their Application to Learning Belief Networks

This paper is devoted to the study of the concept of independence and conditional independence for possibility distributions. Different alternatives are considered, and their properties are analized with respect to a well-known set of axioms which try to capture the intuitive idea of independence. Moreover we carry out a formal study of the conditions that would lead to appropriate definitions of independence for possibilities.

A decision-based approach for recommending in hierarchical domains

Recommendation Systems are tools designed to help users to find items within a given domain, according to their own preferences expressed by means of a user profile. A general model for recommendation systems based on probabilistic graphical models is proposed in this paper. It is designed to deal with hierarchical domains, where the items can be grouped in a hierarchy, each item being only contained in another, more general item. The model makes decisions about which items in the hierarchy are more useful for the user, and carries out the necessary computations in a very efficient way.

Using second-hand information in collaborative recommender systems

Building recommender systems (RSs) has attracted considerable attention in the recent years. The main problem with these systems lies in those items for which we have little information and which cause incorrect predictions. One accredited solution involves using the items’ content information to improve these recommendations, but this cannot be applied in situations where the content information is unavailable. In this paper we present a novel idea to deal with this problem, using only the available users’ ratings. The objective is to use all possible information in the dataset to improve recommendations made with little information. For this purpose we will use what we call second-hand information: in the recommendation process, when a similar user has not rated the target item, we will guess his/her preferences using the information available. This idea is independent from the RS used and, in order to test it, we will employ two different collaborative RS. The results obtained confirm the soundness of our proposal.

Propagation of linguistic labels in causal networks

The problem of the propagation of linguistic labels in polytrees is considered. The approach is purely symbolic and does not consider semantic translations of the terms. It starts from a general axiomatic framework to propagate information in graphs, which is later particularized to the case of linguistic labels. It is observed that it is very difficult to define operations with linguistic labels verifying all the required properties, mainly because of the lack of granularity of finite sets of terms. To cope with these problems, modifications of the propagation algorithms are proposed.<<ETX>>

Fuzzy Measures with Different Levels of Granularity

In this paper we will represent the uncertainty about a particular event, A, by means of an interval [l(A), u(A)]. The extremes of such interval are linguistic values from a particular scale or level of granularity. The problem is how to transform an uncertainty value or interval with arbitrary extremes into an interval expressed on a specific scale. Two conditions are requested to do this transformation: i) Improper information, not contained in the original, must not be obtained, ii) The loss of information must not be too great. Several approaches are considered, and their properties and behaviour are studied. Most of them are based on different methods of ranking fuzzy numbers. Finally, several examples of particular transformations are considered.

Hierarchical naive bayes models for representing user profiles

In this paper, we show how a user profile can be enhanced when a more detailed description of the products is included. Two main assumptions have been considered: the first implies that the set of features used to describe an item can be organized into a well-defined set of components or categories, and the second is that the users rating for a given item is obtained by combining user opinions of the relevance of each component.