James Cussens
[Research]
[Topics for prospective PhD students]
[Projects]
[Software]
[Teaching]
[Professional
Activities]
[Administration]
[Personal history]
[Contact information]
[Dept home page]
Research
Recent papers
- James
Cussens.
Bayesian
network learning by compiling to weighted MAX-SAT.
In Proceedings of the 24th Conference on Uncertainty in Artificial
Intelligence (UAI 2008), Helsinki, 2008.
- Vítor
Santos Costa, David Page, and James Cussens.
CLP(BN):
Constraint logic programming for probabilistic knowledge.
In Luc De Raedt, Paolo Frasconi, Kristian Kersting, and Stephen Muggleton,
editors, Probabilistic Inductive Logic Programming, volume 4911
of Lectures Notes in Computer Science, pages 156-188. Springer,
Berlin, 2008.
-
Nicos Angelopoulos and James Cussens.
Bayesian learning of Bayesian networks with informative priors.
(preprint).
Annals of Mathematics and Artificial Intelligence, 54(1):53-98, 2008
Recent talks
- A tutorial on logic-based approaches to SRL. Invited talk at ILP-MLG-SRL 09, July 2009, Leuven.
-
A logical way to find high probability pedigrees, CRiSM Graphical Models and Genetic Applications workshop, University of Warwick, April 2009.
- Bayesian
network learning by compiling to weighted MAX-SAT, May 2008, York.
-
From
correlations to networks, November 2007, York.
-
Applying algebraic statistics to probabilistic-logical
representations, September 2007, Leuven
My main research interests are in machine learning,
particularly focusing on graphical models and inductive logic
programming and combinations thereof - such a combination is
known as statistical relational learning or probabilistic
inductive logic programming. I also continue to have an
interest in the application of machine learning techniques to
natural language.
To undertake research in these areas it is necessary to have a
strong mathematical background. Unfortunately, I have no
specific funding available to support graduate students. Our
department does award funded
research studentships on a competitive basis.
Informative priors for learning graphical models
Graphical models represent relations of conditional
independence between related variables, for example those
between the various symptoms and causes of a
disease. Automatically learning such models from data is
deservedly a hot topic in machine learning. In some
applications it is crucial to include information not
contained in data - prior information. We have been working on
methods to do this using logical representations within a
Bayesian approach.
Probabilistic inductive logic programming
Inductive logic programming (ILP) uses a logical
representation for machine learning. There is currently great
interest in integrating probabilistic models (e.g. graphical
models) into ILP. In principle there should be no barrier to
doing this, since probabilistic inference is just a special
case of logical (deductive) inference, but there remains
considerable debate about how best to achieve this
integration. One useful avenue of research is to approach this
from a Bayesian point of view, since this reduces statistical
inference to probabilistic inference.
-
James
Cussens
Model equivalence of PRISM programs.
In Proceedings of the Dagstuhl seminar: Probabilistic, Logical and
Relational Learning - A Further Synthesis, 2007. (Associated
talk given at CS Dept, KU Leuven, Belgium.)
- James Cussens. Logic-based
formalisms for statistical relational learning. In Lise
Getoor and Ben Taskar, editors, Introduction to
Statistical Relational Learning. pages 269-290, MIT Press, Cambridge,
MA, 2007.
-
James Cussens.
Individuals, relations and structures in probabilistic models.
In Lise Getoor and David Jensen, editors, IJCAI Workshop on Learning
Statistical Models from Relational Data (SRL2003), pages 32-36,
Acapulco, Mexico, August 2003.
- James Cussens.
Attribute-value and relational learning: A statistical viewpoint.
In Luc De Raedt and Stefan Kramer, editors, Proceedings of the
ICML-2000 Workshop on Attribute-Value and Relational Learning: Crossing the
Boundaries, pages 35-39, 2000.
Many popular approaches to machine learning for natural
language take a standard statistical approach, representing
natural language inputs by often very large vectors of
'features'. The many successes of such an approach show the
importance of using principled statistical methods. However,
logical representations of language can be more compact and
analysable than large feature vectors. There remains important
work to be done to see to what extent the benefits of
statistical and logical methods can be combined for natural
language machine learning.
- James Cussens and Sašo Džeroski, editors.
Learning Language in Logic, volume 1925 of LNAI.
Springer, Berlin, September 2000.
LNAI State-of-the-Art Survey.
Software
- gPy is a collection of
Python modules for manipulating discrete hierarchical models
(including Bayesian nets). It is used to support the
teaching of Algorithms for Graphical
Models.
- The
MCMCMS (Markov chain Monte Carlo over Model Structures)
system
uses Stochastic Logic Programs (SLPs) to define priors for
Bayesian inference. The code was written by
Nicos Angelopoulos.
The
user guide
provides a tutorial on its use.
-
Pepl
is an implementation of the Failure-Adjusted Maximisation
(FAM) algorithm.This is an instance of the EM algorithm which
produces maximum likelihood estimates for the parameters of
SLPs. The code was written by Nicos Angelopoulos.
-
Aaron Bate, a final year student in this department, has
produced software for animating the construction of Prolog
proof trees (the software draws a graphical representation
of the proof tree). It uses Sicstus Prolog and Tcl/Tk. You can download the software as a gzipped
tar file. I have included a simple Prolog SLP
interpreter which allows you to sample from a distribution
over proof trees and where each proof tree determines an
acyclic digraph (the structural element of a Bayesian
net). See the file slp_readme in the distribution
for an explanation.
Projects
Teaching
Currently, at York:
Externally I've done:
At York, I previously taught:
If you're interested in doing a PhD in one of my research
areas, then just drop me a line. Here's details about research
students.
Current PhD students
-
Waleed Alsanie - Relation
Extraction from Text
-
Adel Aloraini - Machine
learning of Bayesian networks
-
Joanne Powell - PrediCtoR: Predicting the Recovery of
Ancient DNA and Ancient Proteins (with Matthew
Collins, Archaeology)
Former students
-
Barnaby Fisher -
Inductive Logic Programming and Mercury
-
Heather
Maclaren
-
Inductive Logic Programming for Software Agents:
Algorithms and Implementations
Professional Activities
Programme chair
Member
Invited speaker
Area chair / Senior PC
PC member
-
AAAI-10, ILP 2010, ECAI-2010
-
Terminologie et intelligence artificielle (TIA - 2009),
ILP-09, SRL-09,
IJCAI-09,
ICML 09,
AISTATS
09, CoNLL 09, NAACL-HLT 09,
EACL
Cognitive 2009, NAACL-2009 Workshop on Unsupervised and Minimally Supervised Learning of Lexical Semantics
-
CoNLL 08,
ILP 08,
ECAI 08,
SBIA 08,
ICML 08
-
UAI07, ILP07, ACL-2007 Workshop on Cognitive Aspects of Computational Language Acquisition, TIA'07
-
AAAI-06,
UAI06,
ILP06,
SRL06,
CoNLL06
-
UAI05,
ICML05,
ECML/PKDD05,
ILP05,
LLLL,
CoNLL05,
TIA05
-
ECML04
,
CIFT04,
UAI04
,
ICML04
,
SRL04
,
CoNLL04
,
Psycho-computational models ...
-
UAI03,
ICML03
,
ILP03,
CoNLL03
,
Acquisition, apprentissage et ...
,
SRL2003
,
ECML03
-
UAI02,
CIFT02,
ICML02
,
ECML02
,
ILP02,
CoNLL02
-
ECML01
,
CoNLL01
,
ILP01
,
LLL01
-
CoNLL00
,
ILP00
,
LLL00
-
ILP99
,
LLL99
- ILP98
Reviewer
Miscellaneous
Administration
Personal history
Contact information
|
Address
|
York Centre for Complex Systems Analysis, Department of
Biology (Area 17), University of York, PO Box 373, York YO10
5YW, UK |
|
Direct phone
|
+44 1904 328396 |
|
Fax
|
+44 1904 500159 |
|
firstname.lastname AT cs DOT york DOT ac DOT uk
|