E-Locus - Institutional Repository of the University of Crete

Home Search

Results - Details

Search command : Author="Γεωργακόπουλος" And Author="Γεώργιος"

Current Record: 2 of 2

[Add to Basket]

Identifier

000403636

Title

Marginal causal consistency in constraint-based causal learning

Alternative Title

Συνέπεια αιτιακών σχέσεων κατά την εκμάθηση περιθωρίων αιτιακών γράφων με αλγορίθμους βασισμένους σε περιορισμούς

Author

Ρουμπελάκη, Άννα Ν.

Thesis advisor

Τσαμαρδινός, Ιωάννης
Τριανταφύλλου, Σοφία

Reviewer

Τόλλης, Ιωάννης
Γεωργακόπουλος, Γεώργιος

Abstract

Maximal ancestral graphs are probabilistic graphical models that can represent the causal relationships among a set of measured variables in the presence of latent confounders. The Fast Causal Inference algorithm is broadly used to retrieve invariant pairwise features of a class of Markov Equivalent MAGs from observational data sets. We investigate the consistency of causal features obtained by FCI and its variations in different marginals of a data set. Under perfect statistical knowledge, the causal relationships obtained by the algorithm when applied on marginal data sets should not conflict the causal relationships obtained when it is applied on all the measured variables. However, in practice FCI is prone to statistical errors and error propagation. As a result, the output of FCI in different marginals may be conflicting. In order to measure causal consistency among marginal data sets, we employed a novel method to identify all invariant causal relationships that exist among the measured variables. In our empirical evaluation on simulated data sets we show that constraint-based algorithms are mostly consistent when applied on marginal data sets. We also introduce an algorithm for ranking causal relationships. We compared with bootstrapping and studied the effect of combining both scores. We studied the problem of marginal causal consistency in various settings, and even in cases of non-ideal scenarios for the algorithm. Finally, we propose a new method of feature selection. The predictors that are selected by our method remain predictive, even when variables on the system are manipulated. We compare our method with several others and conclude that it has similar performance as the Markov Blanket, and outperforms the rest.

Language

English

Subject

Marginal distributions

Αιτιακή μάθηση

Περιθώριες κατανομές

Issue date

2016-11-18

Collection

School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses

Type of Work--Post-graduate theses

Permanent Link