Simhash Near-Duplicate Detection

This library enables the efficient identification of near-duplicate documents using simhash using a C++ extension.

Usage

simhash differs from most hashes in that its goal is to have two similar documents produce similar hashes, where most hashes have the goal of producing very different hashes even in the face of small changes to the input.

The input to simhash is a list of hashes representative of a document. The output is an unsigned 64-bit integer. The input list of hashes can be produced in several ways, but one common mechanism is to:

1. tokenize the document
2. consider overlapping shingles of these tokens (simhash.shingle)
3. hash these overlapping shingles
4. input these hashes into simhash.compute

This has the effect of considering phrases in a document, rather than just a bag of the words in it.

Once we've produced a simhash, we would like to compare it to other documents. For two documents to be considered near-duplicates, they must have few bits that differ. We can compare two documents:

import simhash

a = simhash.compute(...)
b = simhash.compute(...)
simhash.num_differing_bits(a, b)

One of the key advantages of simhash is that it does not require O(n^2) time to find all near-duplicate pairs from a set of hashes. Given a whole set of simhashes, we can find all pairs efficiently:

import simhash

# The `simhash`-es from our documents
hashes = []

# Number of blocks to use (more in the next section)
blocks = 4
# Number of bits that may differ in matching pairs
distance = 3
matches = simhash.find_all(hashes, blocks, distance)

All the matches returned are guaranteed to be all pairs where the hashes differ by distance bits or fewer. The blocks parameter is less intuitive, but is best described in this article or in the paper. The best parameter to choose depends on the distribution of the input simhashes, but it must always be at least one greater than the provided distance.

Internally, find_all takes blocks C distance passes to complete. The idea is that as that value increases (for instance by increasing blocks), each pass completes faster. In terms of memory, find_all takes O(hashes + matches) memory.

Building

This is installable via pip:

pip install git+https://github.com/seomoz/simhash-py.git

It can also be built from git:

git submodule update --init --recursive
python setup.py install

or

pip install simhash-py

under osx, you should

export MACOSX_DEPLOYMENT_TARGET = 10.x (10.9,10.10...)

first

Benchmark

This is a rough benchmark, but should help to give you an idea of the order of magnitude for the performance available. Running on a single core on a vagrant instance on a 2015 MacBook Pro:

$ ./bench.py --random 1000000 --blocks 5 --bits 3
Generating 1000000 hashes
Starting Find all
     Ran Find all in 1.595416s

Architecture

Each document gets associated with a 64-bit hash calculated using a rolling hash function and simhash. This hash can be thought of as a fingerprint for the content. Two documents are considered near-duplicates if their hashes differ by at most k bits, a parameter chosen by the user.

In this context, there is a large corpus of known fingerprints, and we would like to determine all the fingerprints that differ by our query by k or fewer bits. To accomplish this, we divide up the 64 bits into at m blocks, where m is greater than k. If hashes A and B differ by at most k bits, then at least m - k groups are the same.

Choosing all the unique combinations of m - k blocks, we perform a permutation on each of the hashes for the documents so that those blocks are first in the hash. Perhaps a picture would illustrate it better:

63------53|52------42|41-----32|31------21|20------10|09------0|
|    A    |     B    |    C    |     D    |     E    |    F    |

If m = 6, k = 3, we'll choose permutations:
- A B C D E F
- A B D C E F
- A B E C D F
...
- C D F A B E
- C E F A B D
- D E F A B C

This generates a number of tables that can be put into sorted order, and then a small range of candidates can be found in each of those tables for a query, and then each candidate in that range can be compared to our query.

The corpus is represented by the union of these tables, could conceivably be hosted on a separate machine. And each of these tables is also amenable to sharding, where each shard would comprise a contiguous range of numbers. For example, you might divide a table into 256 shards, where each shard is associated with each of the possible first bytes.

The best partitioning remains to be seen, likely from experimentation, but the basis of this is the table. The table tracks hashes inserted into it subject to a permutation associated with the table. This permutation is described as a vector of bitmasks of contiguous bit ranges, whose populations sum to 64.

Example

Let's suppose that our corpus has a fingerprint:

0100101110111011001000101111101110111100001010011101100110110101

and we have a query:

0100101110111011011000101111101110011100001010011100100110110101

and they differ by only three bits which happen to fall in blocks B, D and E:

63------53|52------42|41-----32|31------21|20------10|09------0|
|    A    |     B    |    C    |     D    |     E    |    F    |
|         |          |         |          |          |         |
0000000000000000010000000000000000100000000000000001000000000000

Since any fingerprint matching the query differs by at most 3 bits, at most 3 blocks can differ, and at least 3 must match. Whatever table has the 3 blocks that do not differ as the leading blocks will match the query when doing a scan. In this case, the table that's permuted A C F B D E will match. It's important to note that it's possible for a query to match from more than one table. For example, if two of the non-matching bits are in the same block, or the query differs by fewer than 3 bits.

32-Bit Systems

The only requirement of simhash-py is that it has uint64_t.