How to use correctly ZipfDistribution from Apache commons math library in Java?

Question

I want to create a source of data (in Java) based on words (from a dictionary) that follow a Zipf distribution. So I come to ZipfDistribution and NormalDistribution of the Apache commons library. Unfortunately, information about how to use these classes are rarely. I tried to do some tests but I am not sure if I am using it in the right manner. I am following only what is written in the documentation of each constructor. But the results don't seem to be "well-distributed".

import org.apache.commons.math3.distribution.NormalDistribution;
import org.apache.commons.math3.distribution.ZipfDistribution;

import java.io.BufferedReader;
import java.io.IOException;
import java.io.InputStream;
import java.io.InputStreamReader;
import java.net.URL;

public class ZipfDistributionDataSource extends RichSourceFunction<String> {
    private static final String DISTINCT_WORDS_URL = "https://raw.githubusercontent.com/dwyl/english-words/master/words_alpha.txt";

    public static void main(String[] args) throws Exception {
        ZipfDistributionDataSource zipfDistributionDataSource = new ZipfDistributionDataSource();
        StringBuffer stringBuffer = new StringBuffer(zipfDistributionDataSource.readDataFromResource());
        String[] words = stringBuffer.toString().split("\n");
        System.out.println("size: " + words.length);

        System.out.println("Normal Distribution");
        NormalDistribution normalDistribution = new NormalDistribution(words.length / 2, 1);
        for (int i = 0; i < 10; i++) {
            int sample = (int) normalDistribution.sample();
            System.out.print("sample[" + sample + "]: ");
            System.out.println(words[sample]);
        }

        System.out.println();
        System.out.println("Zipf Distribution");
        ZipfDistribution zipfDistribution = new ZipfDistribution(words.length - 1, 1);
        for (int i = 0; i < 10; i++) {
            int sample = zipfDistribution.sample();
            System.out.print("sample[" + sample + "]: ");
            System.out.println(words[sample]);
        }
    }

    private String readDataFromResource() throws Exception {
        URL url = new URL(DISTINCT_WORDS_URL);
        InputStream in = url.openStream();
        BufferedReader bufferedReader = new BufferedReader(new InputStreamReader(in));
        StringBuilder builder = new StringBuilder();
        String line;
        try {
            while ((line = bufferedReader.readLine()) != null) {
                builder.append(line + "\n");
            }
            bufferedReader.close();

        } catch (IOException ioe) {
            ioe.printStackTrace();
        } catch (Exception e) {
            e.printStackTrace();
        }
        return builder.toString();
    }
}

output

size: 370103
Normal Distribution
sample[185049]: metathesize
sample[185052]: metathetically
sample[185051]: metathetical
sample[185050]: metathetic
sample[185049]: metathesize
sample[185050]: metathetic
sample[185052]: metathetically
sample[185050]: metathetic
sample[185052]: metathetically
sample[185050]: metathetic

Zipf Distribution
sample[11891]: anaphasic
sample[314]: abegge
sample[92]: abandoner
sample[3]: aah
sample[36131]: blepharosynechia
sample[218]: abbozzo
sample[8]: aalii
sample[5382]: affing
sample[6394]: agoraphobia
sample[4360]: adossed

score 1 · Accepted Answer · answered Oct 28 '19 at 13:27

1

You are using it just fine from a code perspective :) The problem is in assuming the source material is ordered by Zipf when it is clearly alphabetical. The whole point of using ZipfDistribution is that words[0] must be the most common word (hint: it's 'the') and roughly twice the freq of words[1]) etc.

https://en.wikipedia.org/wiki/Word_lists_by_frequency https://en.wikipedia.org/wiki/Most_common_words_in_English

answered Oct 28 '19 at 13:27

drekbour

2,895
18
28

I am afraid I didn't get your point. I see that the problem is that my list of words is alphabetically sorted. Then, I didn't understand afterwards. I mean, if I have a number which represents my zipf distribution I can just select the words that correspond to that number, can't I? – Felipe Oct 28 '19 at 13:39
I am not sure if I am right. But if I use the Moby Dick book it already follows very close the Zipf distribution. – Felipe Oct 28 '19 at 13:48
1

Note your example ZipfDistribution `sample` values all lie close to 0 and NormalDistribution `sample` values all lie close to N/2. That is because those are the highest probability areas of their distribution; A random sample is most likely to be from there. The classes have no knowledge of the data except its size N and the prerequisite that it _is_ sorted in the expected distribution. Their purpose is merely to encapsulate the probability computations (say "what is the probability a random word is in the top 10", "How probable is the word 'teapot'" ). – drekbour Oct 28 '19 at 13:54

How to use correctly ZipfDistribution from Apache commons math library in Java?

1 Answers1