Publication: Bayesçi Ağlarda Zamansal Değişkenlerin Kullanımı
Abstract
Bayes Ağları (BA)' nın zaman kavramı ile genişletilmiş bir versiyonu olan Dinamik Bayesçi Ağlar (DBA), sıralı ve/veya zamansal süreçlerin modellenmesinde ve herhangi bir rastgele değişkenler kümesi için nedensellik ilişkilerinin hem görsel, hem de olasılıksal olarak ifade edilmesinde kullanılmaktadır. DBA' da mevcut veri seti kullanılarak belirli algoritmalar ile veri setindeki ilişkileri yansıtan ağ yapısı oluşturulmaktadır. Bu durum 'yapısal öğrenme' olarak adlandırılmaktır. DBA' da yapı öğrenme süreci için üç farklı yöntem bulunmaktadır: Skor tabanlı yöntemler, kısıt tabanlı yöntemler ve karma yöntemlerdir. Bu çalışmada yapı öğrenme için karma yöntemler kullanılmıştır. Karma yöntemlerde, önce kısıt tabanlı yöntemler daha sonra skor tabanlı yöntemler uygulanmaktadır. Yapı öğrenme süreci üzerinde kullanılan skor kriterlerinin önemli bir etkiye sahip olduğu söylenebilmektedir. Çalışmada amaç; sürekli durumlu DBA' da karma yapı öğrenme süreci üzerinde alternatif skor kriterlerinin etkilerinin incelenmesidir. Bunun için DBA karma yapı öğrenme sürecinde farklı skor ve kısıt tabanlı algoritmalar ve skor kriterleri ile oluşturulan kombinasyonlar ile farklı boyuttaki simüle ve gerçek (UCI) çok değişkenli zaman serisi verileri için DBA yapı öğrenme işlemi gerçekleştirilmiştir. Öğrenilen ağ yapılarının performansları için ortalama skor ölçümleri elde edilerek, sonuçlar skor kriterleri açısından incelenmiştir. Hem simüle hem de UCI verisi için elde edilen sonuçlar genel olarak incelendiğinde, değişken sayısı 5 olan simüle veriler için BGE skoru ile, değişken sayısı 7, 10 olan simüle veriler ve UCI verisi için BIC-G skoru ile daha güçlü sonuçlar elde edilmiştir. Sonuçlar sadece kullanılan alternatif AIC ve BIC skor kriteleri açısından incelendiğinde, farklı BIC skorlarının AIC skorlarına göre daha güçlü sonuçlar verdiği söylenebilmektedir.
A Bayesian Network (BN) is a probabilistic graphical model that represents a set of variables and their conditional dependencies via a Directed Acyclic Graph (DAG). Dynamic Bayesian Networks (BDNs) extend standard BN with the concept of time and they are probabilistic graphical models dedicated to modeling multivariate time series. DBNs can deal with discrete, continuous and both discerete and continuos variable states. The task of structure learning for DBNs refers to learn the structure of the DAG from dataset and there are three approaches for the structure learning: score-based approach, constraint-based approach and hybrid-based approach. In this study, hybrid-based approachs are used for DBN structure learning. They aggregate both independence-based and score-based structure learning algorithms. Firstly, constraint-based algorithms are used to determine the initial network structure and the score-based algorithms are used which include search strategies and scoring functions to find the highest score network structure. In addition, the effect of the number of time slices determined in the DBNs on structure learning has been examined. The aim of this study is to examine the effects of alternative score criteria on hybrid structure learning process in continuous-state DBN. For this, DBN structure learning process was performed for different size simulated and real (UCI) multivariate time series datasets with combinations of different score-based, constraint-based algorithms and score criteria. Average score measurements were obtained for the performances of the learned network structures and the results were examined in terms of score criteria. For both simulated and UCI datasets, general results showed that BGE score for simulated data with 5 variables and BIC-G score for data sets with 7 and 10 variables showed stronger performance. When the results for the alternate AIC and BIC score criteria used were examined, it was concluded that different BIC scores gave stronger results than AIC scores.
A Bayesian Network (BN) is a probabilistic graphical model that represents a set of variables and their conditional dependencies via a Directed Acyclic Graph (DAG). Dynamic Bayesian Networks (BDNs) extend standard BN with the concept of time and they are probabilistic graphical models dedicated to modeling multivariate time series. DBNs can deal with discrete, continuous and both discerete and continuos variable states. The task of structure learning for DBNs refers to learn the structure of the DAG from dataset and there are three approaches for the structure learning: score-based approach, constraint-based approach and hybrid-based approach. In this study, hybrid-based approachs are used for DBN structure learning. They aggregate both independence-based and score-based structure learning algorithms. Firstly, constraint-based algorithms are used to determine the initial network structure and the score-based algorithms are used which include search strategies and scoring functions to find the highest score network structure. In addition, the effect of the number of time slices determined in the DBNs on structure learning has been examined. The aim of this study is to examine the effects of alternative score criteria on hybrid structure learning process in continuous-state DBN. For this, DBN structure learning process was performed for different size simulated and real (UCI) multivariate time series datasets with combinations of different score-based, constraint-based algorithms and score criteria. Average score measurements were obtained for the performances of the learned network structures and the results were examined in terms of score criteria. For both simulated and UCI datasets, general results showed that BGE score for simulated data with 5 variables and BIC-G score for data sets with 7 and 10 variables showed stronger performance. When the results for the alternate AIC and BIC score criteria used were examined, it was concluded that different BIC scores gave stronger results than AIC scores.
Description
Citation
WoS Q
Scopus Q
Source
Volume
Issue
Start Page
End Page
105
