-
Notifications
You must be signed in to change notification settings - Fork 0
/
discussion.tex
20 lines (19 loc) · 890 Bytes
/
discussion.tex
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
\section{Discussion}
From the experiment results,
we can see that
\begin{itemize}
\item Solving the problem using LP requires lots of time.
and it is infeasible for large scale problems.
In our experiments, we were forced to truncate the dataset
to a small scale (20000 rows) due to this limitation.
\item The preparation time for the sketching is negligible in
experiments.
This is different from HW1.
\item The performances of the two sketching methods are similar.
Increasing the number of rows sampled helps imporve performance.
We can get a reletive error of $0.58\%$
by sampling $2.5\%$ of all rows.
\item Time consumption of sketching methods is extremely small.
This makes it suitable for large scale dataset.
\end{itemize}
Therefore sketching methods is very useful for large scale problems.