Skip to main content

Package for consistent sampling with or without replacement.

Project description

consistent_sampler

Routine sampler for providing 'consistent sampling' --- sampling that is consistent across subsets (as explained below).

Here we call the elements to be sampled "ids", although they may be arbitrary python objects (strings, tuples, whatever). We assume that ids are distinct.

Consistent sampling works by associating a random "ticket number" with each id; the desired sample is found by taking the subset of the desired sample size containing those elements with the smallest associated random numbers.

The random ticket numbers are computed using a given "seed"; this seed may be an arbitrary python object, typically a large integer or long string.

The sampling is consistent since it consistently favors elements with small ticket numbers; if two sets S and T have substantial overlap, then their samples of a given size will also have substantial overlap (for the same random seed).

This routine takes as input a finite collection of distinct object ids, a random seed, and some other parameters. The sampling may be "with replacement" or "without replacement". One of the additional parameters to the routine is "take" -- the size of the desired sample.

It provides as output a "sampling order" --- an ordered list of object ids that determine the sample. Each object id as associated with a random value (its "ticket number") that depends on the id and the seed; ids are output in order of increasing ticket number. For efficiency and portability, the ticket number is represented as a decimal fraction 0.dddd...dddd between 0 and 1.

For sampling without replacement, the output can not be longer than the input, as no id may appear in the sample more than once.

For sampling with replacement, the output may be infinite in length, as an id may appear in the sample an arbitrarily large (even infinite) number of times. When an id is sampled and then replaced in the set of ids being sampled, it is given a new random ticket number drawn uniformly from the set of numbers in (0, 1) larger than its previous ticket number.

The output of sampler is always a python generator, capable of producing an infinitely long stream of ids.

Example 1.

As a small example of sampling without replacement:

g = sampler(['A-1', 'A-2', 'A-3', 'B-1', 'B-2', 'B-3'], 
            with_replacement=False, take=4, seed=314159, output='id')

yields a generator g whose output can be printed:

print(list(id for id in g))

which produces:

['B-2', 'B-3', 'A-3', 'A-2']

Example 2.

Here is an example where sampling is with replacement from a set of 6 ids, and the output gives a triple (tuple) for each selected id, giving 1. the associated random ticket number, 2. the selected id, and 3. the "generation" (number of times the id has been selected so far).

>>> for t in sampler(['a1', 'b2', 'c3', 'd4', 'e5', 'f6'],
                     with_replacement=True, seed=19283746, take=10):
...     print(t)
('0.303241347', 'e5', 1)
('0.432145156', 'b2', 1)
('0.487135586', 'c3', 1)
('0.581779914', 'b2', 2)
('0.680782907', 'b2', 3)
('0.700258702', 'c3', 2)
('0.816686725', 'b2', 4)
('0.841870265', 'a1', 1)
('0.857737141', 'a1', 2)
('0.866227993', 'f6', 1)

Discussion

This routine is designed for use in election audits, where the ids being sampled are ballot ids, but this routine is suitable for general use.

For a similar election audit sampling method, see Stark's election audit tools: https://www.stat.berkeley.edu/~stark/Vote/auditTools.htm

Consistent sampling is not a new idea, see for example https://arxiv.org/abs/1612.01041 and the references to consistent sampling therein. The routine here may (or may not) be novel in that it extends consistent sampling to sampling with replacement.

For our applications, one big advantage of consistent sampling is the following. If each county collects cast ballots separately, then they can order their own ballots for sampling and interpretation independently of what other counties are doing. An overall sample order can be constructed from the individual county sample order, by merging the list of triples each produces into an overall sorted order.

Usage

Further documentation and examples are: 1. in the code consistent_sampler.py 2. obtainable by running the program demo_consistent_sampler.py 3. described in the file USAGE_EXAMPLES.md 4. in the file docs/consistent_sampling_with_replacement.pdf (This file will soon appear on arXiv as well.)

This code has been packaged and uploaded to PyPI. From python3 you can say

from consistent_sampler import sampler

and then say help(sampler) for more documentation, or run sampler as in the above examples.

Efficiency

The bulk of the work is in computing the SHA256 hash function, which takes about one microsecond per call on a typical laptop. Thus, the running time of sampler is proportional to the length if id_list (to set up the priority queue) plus, if sampling is done with replacement, the value of drop+take, where the constant of proportionality is about 1 microsecond.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

consistent_sampler-1.0.10.tar.gz (13.8 kB view details)

Uploaded Source

Built Distribution

consistent_sampler-1.0.10-py3-none-any.whl (14.1 kB view details)

Uploaded Python 3

File details

Details for the file consistent_sampler-1.0.10.tar.gz.

File metadata

  • Download URL: consistent_sampler-1.0.10.tar.gz
  • Upload date:
  • Size: 13.8 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/1.11.0 pkginfo/1.4.2 requests/2.19.1 setuptools/40.2.0 requests-toolbelt/0.8.0 tqdm/4.25.0 CPython/3.7.0

File hashes

Hashes for consistent_sampler-1.0.10.tar.gz
Algorithm Hash digest
SHA256 89dd222271fdf0bb941f5eccee75be39b2ac0c0205af7b7023a12bab99d38fba
MD5 86a7943449f72746b5f3c8dfaf5b55d6
BLAKE2b-256 1d8e3571bb805337f61d7976567d19208006fcea6867416ccc4e95584e640de5

See more details on using hashes here.

File details

Details for the file consistent_sampler-1.0.10-py3-none-any.whl.

File metadata

  • Download URL: consistent_sampler-1.0.10-py3-none-any.whl
  • Upload date:
  • Size: 14.1 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/1.11.0 pkginfo/1.4.2 requests/2.19.1 setuptools/40.2.0 requests-toolbelt/0.8.0 tqdm/4.25.0 CPython/3.7.0

File hashes

Hashes for consistent_sampler-1.0.10-py3-none-any.whl
Algorithm Hash digest
SHA256 80cd24a9bfb8089b929915805744a9c3604c97314de492599bd3be43252fa0a4
MD5 00ab2b0f535acce6c97536368290baee
BLAKE2b-256 c4abe074ffa44ff637542a72518936296b6a82e3a34b0e98750affe95ac07b5d

See more details on using hashes here.

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page