Learning Writing A Paper in Causal Discovery
Motivation
In different disciplines of the machine learning research, there are different ways, i.e., traditions and habits, to write a paper. If we cannot follow the same these rules to write papers, our papers will be treated as “not professional” and even rejected. Therefore, being familiar with the traditions is necessary for every newbie. Learning from the classical paper is one of the best way to mimic a writing style. In this blog, I chose paper Causal Discovery from Nonstationary/Heterogeneous Data: Skeleton Estimation and Orientation Determination as my first paper to learn the writing style in causal discovery. If you have any advice for improving writing, please leave your comments. I would be grateful for your generous suggestions.
Expectation
Learn something about
- how to organise paragraphs with a flow in a paper,
- how to introduce a method for being understood better,
Key points/ What I learned
Abstract
1. A phenomenon existed in the real world
2. The phenomenon presents problems, challenges opportunities
3. What we do, what is our contribution
4. Experimental results show
Introduction
1. causal discovery is important
2. Focus on the problem, XXX we want to solve: one feature of the data, distribution shift
— introduce what is XXX
— For examples of XXX
— How XXX influence the existed methods
3. Our work”
— we assume
— we aim to
— address the following questions:
4. Organized as follow:
Problem Definition and Related work
1. Introduce all the basic concepts and notations we need: observed variable -> underlying causal structure -> factorize joint probability distribution of V -> causal module -> changing causal module -> we assume that
2. How the problem influences the current methods -> Example -> Consequently,
3. Related work:
4. In contrast, what we do,
We can also show that distribution shifts actually contain useful information for the purpose of determining causal directions and develop practical algorithms accordingly
CD-NOD Phase 1
Assumptions:
not only list the assumptions, but motivate the assumption and give the consequences.
Detecting changing modules and recovering causal skeleton:
Aim => basic idea =>this is achieved by Algorithm … and supported by Theorem …=> Explain the procedures: in step 1, ….. ; in step 2, ….
CD-NOD Phase 2
Experiment Results
Conclusion and Discussions
Phrases:
empirical science,
causal knowledge,
carry out randomized experiments,
inferring causal relations from purely observational data,
economics,
neuroscience,
data exhibit features
distribution shift
invariance of causal mechanisms
data sets
The above two give the procedure of CD-NOD
a distribution underlying the observed data
the conditional independence graph
for the observed data
spurious connections
a set of observed variables
underlying causal structure
develop a nonparametric and computationally efficient causal discovery procedure
to discover the causal skeleton and orientations from all data points simultaneously
Presentation:
with the rapid accumulation of huge volumes of data of various types, drawn much attention,
occur across data sets
define and motivate the problem in more details
develop a method for determining … by …
be checked by …
report experimental results tested on both synthetic and real-world
To these situations, …
… fail to apply, as …
underlying causal structure over V is represented by a DAG G
… Moreover, …
be written …, and denoted by …
If … happens, … will be … Therefore, … may not be work
As an illustration, suppose …; the influence. Moreover, …
To tackle the issues, one… Improved versions include … Such method may suffer …
Some methods aim to …, but …
several methods aim to …. Compared to them, … more difficult.
Moreover, … assume …, limiting their applicability to complex problems
As a consequence,
it is crucial to …
