Doctoral theses
Current Record: 58 of 125
|
Identifier |
000395038 |
Title |
Integrative causal analysis of heterogeneous data sets |
Alternative Title |
Ολοκληρωμένη αιτιακή ανάλυση ετερογενών συνόλων δεδομένων |
Author
|
Τριανταφύλλου, Σοφία
|
Thesis advisor
|
Τσαμαρδινός, Ιωάννης
|
Reviewer
|
Μπένος, Παναγιώτης
Πλεξουσάκης, Δημήτριος
Χριστοφίδης, Βασίλειος
Cooper, Gregory
Glymour, Clark
Maathuis, Marloes
|
Abstract |
Scientific practice typically involves repeatedly studying a system, each time trying to unravel a dif¬ferent perspective. In each study, the scientist may take measurements under different experimental conditions (interventions, manipulations, perturbations) and measure different sets of quantities (variables). The result is a collection of heterogeneous data sets coming from different data distribu¬tions. These data sets are analyzed in isolation and results are manually synthesized by the scientific community into scientific knowledge.
This thesis argues that heterogeneous data sets measuring the same system under study must all stem from, and therefore reflect, the same underlying causal mechanism, and that they can be co-analysed based on this premise. We define the problem of identifying one or all causal models that best fit all available data sets. We call this approach Integrative Causal Analysis.
The standard assumptions of causal modelling connect the statistical properties entailed in the available data sets to the underlying causal mechanism. Particularly, multivariate statistical relations of the measured variables constrain the search space of possible underlying causal models. Thus, the problem can be recast as a constraint satisfaction problem.
We propose an efficient conversion that translates statistical constraints into a SAT instance that can be solved with state-of-the-art SAT solvers. To improve scalability of our method we employ a series of approximate or exact steps that restrict the complexity of the conversion. Additionally, we introduce a scalable method for resolving conflicts arising from statistical errors. Finally, we identify a minimal example where INCA can produce a non-trivial prediction. We then test this prediction extensively in public data sets from a wide range of scientific domains, in an attempt to test whether causally-inspired predictions are verified.
We test our methods in a variety of different data sets and conditions. Results indicate that (a) our methods are robust and behave reasonably against different input parameters (b) our methods outperform state-of-the-art alternatives and (c) while causal assumptions cannot be easily verified, they lead to statistical predictions that are largely validated in real-world data sets.
|
Language |
English |
Subject |
Bayesian networks |
|
Μπαϋεσιανά δίκτυα |
Issue date |
2015-05-12 |
Collection
|
School/Department--School of Sciences and Engineering--Department of Computer Science--Doctoral theses
|
|
Type of Work--Doctoral theses
|
Permanent Link |
https://elocus.lib.uoc.gr//dlib/c/1/a/metadata-dlib-1435310689-323364-20375.tkl
|
Views |
640 |