JMLR Special Topic on
Causality
Large-scale Experiment Design and Inference of Causal
Mechanisms
Second call-for-papers, deadline: September 15, 2014
Please submit your papers to JMLR.org by selecting the special topic LSED and send email with your SUBMISSION NUMBER to the guest
editors at causality <at> chalearn <dot> org
The
problem of attributing causes to effects is pervasive in
science, medicine, economy and almost every aspects of our
everyday life involving human reasoning and decision making.
What affects your health? the economy? climate changes? The
gold standard to establish causal relationships is to
perform randomized controlled experiments. However,
experiments are costly while non-experimental
"observational" data collected routinely around the world
are readily available. Unraveling potential cause-effect
relationships from such observational data could save a lot
of time and effort by allowing us to prioritize confirmatory
experiments. This could be complemented by new strategies of
incremental experimental design combining observational and
experimental data. This special topic will include methods
of experimental design, which involve machine learning in
the process of data collection.
Specific
areas relevant to the special topic include, but are not
limited to:
a. Methods to discover causal structure from data and
to perform causal inference (e.g., estimate causal effects,
predict effects of actions, produce most probable causal
explanations, perform inference with counter-factuals, etc.).
Methods based on the use of multiple types of data (e.g.,
observational, experimental, case control) and methods based
on combining knowledge (e.g., in the form of constraints or
prior beliefs) and data, are encouraged. Such methods may be
based on Bayesian Networks and other Probabilistic Graphical
Models, Markov Decision Processes, Structural Equation Models,
Propensity Scoring, Information Theory, Granger Causality, or
other appropriate frameworks. New methods of experimental
design capitalizing on massive amounts of available
observational data and minimal interventions. Methods of
pseudo-experiments, quasi-experiments, and natural
experiments.
b. Theory:
- Identifiability of causal relationships from observational
data or a combination of observational and experimental data.
- Definitions of causality bridging the gap between data
generative definitions, interventional definitions, and
counterfactuals.
- Formal criteria (e.g., statistical tests of significance of
causal relationships, confidence intervals, model scoring
measures.) for causal model selection.
- Properties (e.g., soundness/consistency, stability, sample
efficiency, computational efficiency) of existing and novel
causal discovery methods.
- Formal connections relevant to experiment design and causal
discovery among diverse fields such as Statistics, Artificial
Intelligence, Decision Theory, Econometrics, Markov Decision
Processes, Control Theory, Operations Research, Planning, etc.
c. Assumptions for causal discovery. Theoretical and
empirical study of:
- Study of violations of typical assumptions for causal
discovery (e.g., Causal Faithfulness Condition, Causal Markov
Condition, Causal Sufficiency, causal graph sparseness,
linearity, specific parametric forms of data distributions,
etc.).
- Prevalence and severity of violations of assumptions and
study of worst-case and average-case effects of such
violations.
- Novel or modified assumptions and their properties.
d. Evaluation methods, including the study of
appropriate performance measures, research designs, benchmarks
etc. to empirically study the performance and pros and cons of
experimental design and causal discovery methods.
e. Real-world applications and benchmarking of experimental
design and causal discovery algorithms, including
rigorous studies of highly innovative software environments.
Guest Editors:
Isabelle Guyon, ChaLearn, Berkeley, California, USA.
Alexander Statnikov, New York University, New York, USA.
For further instructions about the submission procedure please
read the JMLR policies or send an email to the special
topic guest editors to causality <at> chalearn
<dot> org.
The call is open to the general public. The participants on
the NIPS
2013 workshop on causality and the cause-effect
pairs challenge are encouraged
to submit a paper.
Recommendations to competitors invited to write a
JMLR paper:
The papers will be judged according to to following criteria:
(1) Performance in the challenge,
(2) Novelty/Originality,
(3) Sanity (correct proofs, good experiments),
(4) Insight, and
(5) Clarity of presentation.
Papers merely describing the steps taken to produce a
challenge entry will not be judged favorably. Please include
in your submissions:
- The choices and advantages of the methods employed should be
supported by a literature overview and qualitative and
quantitative comparisons with other methods on the data of the
challenge and possibly other data.
- The various building blocks of the presented methods should
be analyzed separately and key novel elements contributing to
boosting performance significantly should be singled out.
- The authors are also encouraged to motivate new approaches
in a principled way and draw insights that go beyond the
framework of the challenge.
JMLR is a very selective publication and your paper will
undergo a regular journal review. Your chances of acceptance
will be increased if you:
- clearly motivate your approach from a practical and
theoretical standpoint
- present a consistent set of experiments (using the
development data) showing a significant advantage over other
methods
- cite your final evaluation results in the challenge
- make sure that your paper is well organized, well written,
with good references, figures, and tables
We recommend not to exceed 20 pages.