PhD Defence • Artificial Intelligence • Improved Bayesian Network Structure Learning in the Model Averaging ParadigmExport this event to calendar

Thursday, December 1, 2022 — 2:00 PM to 5:00 PM EST

Please note: This PhD defence will take place online.

Zhenyu (Alister) Liao, PhD candidate
David R. Cheriton School of Computer Science

Supervisor: Professor Peter van Beek

A Bayesian network (BN) is a probabilistic graphical model with applications in knowledge discovery and prediction. Its structure can be learned from data using the well-known score-and-search approach, where a scoring function is used to evaluate the fit of a proposed BN to the data in an unsupervised manner, and the space of directed acyclic graphs is searched for the best-scoring BNs. However, selecting a single model (i.e., the best-scoring BN) is often not the best choice. When one is learning a BN from limited data, selecting a single model may be misleading as there may be many other BNs that have scores that are close to optimal, and the posterior probability of even the best-scoring BN is often close to zero.

A more preferred alternative to committing to a single model is to perform some form of Bayesian or frequentist model averaging. A widely used data analysis methodology is to: (i) learn a set of plausible networks that fit the data, (ii) perform model averaging to obtain confidence measure for each edge, and (iii) select a threshold and report all edges with confidence higher than the threshold. In this manner, a representative network can be constructed from the edges that are deemed significant that can then be examined for probabilistic dependencies and possible cause-effect relations.

This thesis presents several improvements to Bayesian network structure learning that benefit the data analysis methodology. We propose a novel approach to model averaging inspired by performance guarantees in approximation algorithms. Our approach has two primary advantages. First, our approach only considers \emph{credible} models in that they are optimal or near-optimal in score. Second, our approach is more efficient and scales to significantly larger Bayesian networks than existing approaches.

We empirically study a selection of widely used and also recently proposed scoring functions. We address design limitations of previous empirical studies by scaling our experiments to larger BNs, comparing on an extensive set of both ground truth BNs and real-world datasets, considering alternative performance metrics, and comparing scoring functions on two model averaging frameworks: the bootstrap and the credible set. Contrary to previous recommendations based on finding a single structure, we find that for model averaging the BDeu scoring function is the preferred choice in most scenarios for the bootstrap framework and a recent score qNML is the preferred choice for the credible set framework. We identify an important shortcoming in a widely used threshold selection method. We then propose a simple transfer learning approach for maximizing target metrics and selecting a threshold that can be generalized from proxy datasets to the target dataset and show on an extensive set of benchmarks that it can perform significantly better than previous approaches. We demonstrate via ensemble methods that combining results from multiple scores significantly improve both the bootstrap and the credible set approach on various metrics, and that combining all scores from both approaches still yields better results.


To join this PhD defence on MS Teams, please go to https://teams.microsoft.com/l/meetup-join/19%3ameeting_NzhmOWUxN2EtYzBiNS00ZjhkLWJmNzctZTczZThiMDRlZTA4%40thread.v2/0?context=%7b%22Tid%22%3a%22723a5a87-f39a-4a22-9247-3fc240c01396%22%2c%22Oid%22%3a%2255a4dd3f-4336-4fc1-920f-697bade427ea%22%7d.

Location 
Online PhD defence
200 University Avenue West

Waterloo, ON N2L 3G1
Canada
Event tags 

S M T W T F S
27
28
29
30
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
  1. 2023 (2)
    1. January (2)
  2. 2022 (241)
    1. December (16)
    2. November (28)
    3. October (15)
    4. September (12)
    5. August (29)
    6. July (23)
    7. June (17)
    8. May (20)
    9. April (24)
    10. March (22)
    11. February (16)
    12. January (19)
  3. 2021 (210)
    1. December (21)
    2. November (13)
    3. October (12)
    4. September (21)
    5. August (20)
    6. July (17)
    7. June (11)
    8. May (16)
    9. April (27)
    10. March (20)
    11. February (13)
    12. January (19)
  4. 2020 (217)
  5. 2019 (255)
  6. 2018 (217)
  7. 2017 (36)
  8. 2016 (21)
  9. 2015 (36)
  10. 2014 (33)
  11. 2013 (23)
  12. 2012 (4)
  13. 2011 (1)
  14. 2010 (1)
  15. 2009 (1)
  16. 2008 (1)