Scribd
Upload
149 views

Fast Kde

The document discusses a new Java-based fast kernel density estimation (KDE) library and compares its performance to an existing Python library. It tested the libraries on word vector data from the abstracts of six medical journals. The Java library calculated KDEs roughly 100 times faster than the Python alternative. Plots of the KDEs for each journal showed differences in the diversity of their research topics.

Uploaded by

josh
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
149 views

Fast Kde

The document discusses a new Java-based fast kernel density estimation (KDE) library and compares its performance to an existing Python library. It tested the libraries on word vector data from the abstracts of six medical journals. The Java library calculated KDEs roughly 100 times faster than the Python alternative. Plots of the KDEs for each journal showed differences in the diversity of their research topics.

Uploaded by

josh
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 1

Fast Kernel Density Estimation https://github.

com/jsisaacs/fastKDE

FastKDE and its Application Towards Text Analysis


Joshua Isaacson, JT Wolohan

Overview
Kernel density estimation (KDE) is a technique

for approximating the probability distribution

of sample data. Histograms estimate local density,

but points must fall into the same discrete bins to be

counted correctly. With KDE, the distribution is mapped

continuously, which avoids the problems of a rough

estimate and data falling into bins that fail to

represent their true position.

Currently, there isn’t a reliable and fast KDE library

implemented for JVM languages. The project’s goal is

to build such a library and compare it to an existing

Python library. It is expected that the JVM library will be

faster than the Python counterpart because numerical

computing with the Java-based Nd4j library is shown

to be faster than numpy, which is what most statistical

Python packages use.

Data
To test the library, we used the
PubMed dataset from

data.gov, which is made up of


over 26 million citations

for medical research literature.


The entire data was roughly

50 GB, which was then parsed


to contain only article
abstracts and their corresponding journals. We then

trained a Word2vec model with the article abstract texts.

From there, 6 journals were chosen to be analyzed:

Brain Research, Neurology, Journal of Neurochemistry,

Neuroscience, Circulation, and Chest. The last two journals

are of different parts of the body from the first 4, so we

expected their distributions to be slightly different. For each

of the 6 journals, we obtained the corresponding word

vectors, which were the inputs to our JVM-based KDE

library.

For the six journals, the speed to calculate the KDE

Implementation was roughly 100 seconds for Seaborn Kdeplot and

2 seconds for the Java-based FastKDE.


Our library is based off of the
FastKDE algorithm
The above twelve figures are the KDE plots of the six

created by T. O’Brien, journal’s word vectors, with the left side being the

Kashinath, Cavanaugh, Collins,


default Seaborn Kdeplot, and the right being our

and J. O’Brien in 2016. The FastKDE library. As the plots show, a greater area

inputs are 2 one-dimensional


correlates to a greater range of words in the
matrices, as well as optional associated document, in this case the journal.
grid dimensions, weights,
Neurology, Circulation, and Chest all have broad
and adjustment factor. There are distributions, which suggests that the journals have a
seven main components to the algorithm: grid dimension
wider range of research articles with more diverse
optimization, bin setup, grid creation, bandwidth
language, compared to densely packed KDE plots,
determination, kernel calculation, 2-dimensional
like that of Brain Research, Journal of Neurochemistry,
convolution, and scaling by a normalization factor. and Neuroscience, which could explain that their
journals have more concentrated research content.

You might also like