Using Value-Based Potentials for Making Approximate Inference on Probabilistic Graphical Models

Abstract

The computerization of many everyday tasks generates vast amounts of data, and this has lead to the development of machine-learning methods which are capable of extracting useful information from the data so that the data can be used in future decision-making processes. For a long time now, a number of fields, such as medicine (and all healthcare-related areas) and education, have been particularly interested in obtaining relevant information from this stored data. This interest has resulted in the need to deal with increasingly complex problems which involve many different variables with a high degree of interdependency. This produces models (and in our case probabilistic graphical models) that are difficult to handle and that require very efficient techniques to store and use the information that quantifies the relationships between the problem variables. It has therefore been necessary to develop efficient structures, such as probability trees or value-based potentials, to represent the information. Even so, there are problems that must be treated using approximation since this is the only way that results can be obtained, despite the corresponding loss of information. The aim of this article is to show how the approximation can be performed with value-based potentials. Our experimental work is based on checking the behavior of this approximation technique on several Bayesian networks related to medical problems, and our experiments show that in some cases there are notable savings in memory space with limited information loss.

Advances in Artificial Intelligence and Statistical Techniques with Applications to Health and Education

“…Topics addressed in this Special Issue include data mining, machine learning, learning analytics, prediction methods, pattern recognition, decision analysis, probabilistic reasoning, fuzzy systems, student or patient modelling, adaptive systems, collaborative systems, recommendation systems, experimental design, and empirical study cases. Specifically, there are twenty rigorously reviewed papers included in this Special Issue, with eleven specializing in the field of medicine and health [1][2][3][4][5][6][7][8][9][10][11] and nine specializing in the field of education [12][13][14][15][16][17][18][19][20]. In total, fifteen papers (43% of the received) were rejected for publication.…”

“…In the area of health, most papers aim at improving the diagnosis of a disease, such as influenza [3], Alzheimer's disease [7], and cardiac arrhythmias [11], or of a specific symptom, such as changes in the retinal vasculature [1], mental fatigue [4], gait variability [5] or vitamin D deficiency [8]. There is one for the personalization of rehabilitation exercises for stroke patients [6], and there are three theoretical papers [2,9,10] aimed at improving the representation of models, which can be applied to different and actual medical problems.…”

“…Regarding the techniques used, those related to artificial intelligence prevail. Specifically, the majority make use of different machine learning algorithms [1][2][3][4]6,8,11], although with distinct objectives. For example, some works propose new specific classification models.…”

“…Another approach was proposed by [2] with respect to improving the efficiency of classification algorithms when there are memory limitations. Then, an improvement in value-based potentials to better represent approximate information and positive results was obtained after verifying it on several Bayesian networks representing medical problems.…”

“…It should be noted that the data used for the above studies have been obtained from public repositories in papers [1,2,11]. The other papers obtain their data from different types of experimental studies, most of them conducted with individuals [3][4][5][6][7][8].…”