I implemented SimHash in Python. It took me a while to figure out all the bit fiddling, but I think it is pretty useful (although, it might require some experimentation to choose a good similarity...
I implemented SimHash in Python. It took me a while to figure out all the bit fiddling, but I think it is pretty useful (although, it might require some experimentation to choose a good similarity threshold that minimizes false positives and false negatives on many data sets). This algorithm could be applied to data types other than text/strings (such as image or audio data), but I leave that as an exercise for the reader. Note that there are other efficient LSHs such as Nilsimsa or Minhash.
Also note that using Python’s built-in hash as an underlying hashing function is not ideal because the hash values are salted with a seed that varies across different invocations of Python (at least since Python 3.3). I opted to use cityhash instead, because it is a fast string hashing algorithm.
Excerpting from the last section of the README, which is where the meat is:
Locality Sensitive Hashing
The script uses locality sensitive hashing (LSH) to compare texts. LSHs are a kind of hashing function that, unlike normal hashing functions, have the property that the probability of a hash collision is proportional to the similarity of objects hashed. In this case, the LSH implemented is simhash (described here).
Running dedup.py with the -v/--verbose option prints the binary values of the computed LSHs and shows the similarity scores:
# 32 bit LSH
(dedup) $ find test -name "*.txt" | ./dedup.py - -v
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████| 7/7 [00:00<00:00, 5169.95files/s]
0%| | 0/32 [00:00<?, ?rotations/s]test/1.txt : the cat sat in the hat
test/2.txt : the cat sat on the mat
00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001001110100011100101111100010
00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001001010100111101101110111010
0.1875
test/3.txt : we all scream for ice scream
test/4.txt : we all scream for ice cream
00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000010001000010001111101100001011010
00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000010001000010001111101000101011110
0.09375
test/3.txt : we all scream for ice scream
test/5.txt : we all like ice scream
00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000010001000010001111101100001011010
00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000010011000010011111101100011001110
0.15625
test/4.txt : we all scream for ice cream
test/5.txt : we all like ice scream
00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000011010001000010001111101000101011
00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000011010011000010011111101100011001
0.1875
test/7.txt : abcdefg
hijklmnop
test/6.txt : abcdefg
00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001101100000001101001111010110110
00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001101000001000101001111010110001
0.1875
100%|██████████████████████████████████████████████████████████████████████████████████████████████| 32/32 [00:00<00:00, 17233.91rotations/s]
found 1 duplicate(s) of file: test/1.txt
test/1.txt test/2.txt
found 2 duplicate(s) of file: test/3.txt
test/3.txt test/4.txt
test/3.txt test/5.txt
found 1 duplicate(s) of file: test/4.txt
test/4.txt test/5.txt
found 1 duplicate(s) of file: test/7.txt
test/7.txt test/6.txt
found 5 duplicate(s) in total
# 64 bit LSH
(dedup) $ find test -name "*.txt" | ./dedup.py - -v -b 64
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████| 7/7 [00:00<00:00, 3116.79files/s]
0%| | 0/64 [00:00<?, ?rotations/s]test/7.txt : abcdefg
hijklmnop
test/6.txt : abcdefg
00000000000000000000000000000000000000000000000000000000000000000000011110111011000101110111001001000010001001010011000101110100
00000000000000000000000000000000000000000000000000000000000000000010011110011011111100110111000001000010000001010011001111011100
0.1875
test/1.txt : the cat sat in the hat
test/2.txt : the cat sat on the mat
00000000000000000000000000000000000000000000000000000000000000000011000110100111001010111010011101000000100011001101111110100010
00000000000000000000000000000000000000000000000000000000000000000011000111100111100010111011011101000000100011001101111001100010
0.109375
test/3.txt : we all scream for ice scream
test/4.txt : we all scream for ice cream
00000000000000000000000000000000000000000000000000000000000000000101000000000111011010000000001000010000101010110001001001101001
00000000000000000000000000000000000000000000000000000000000000000101000000000111011011001000001000010100101011110001001001101001
0.0625
test/4.txt : we all scream for ice cream
test/5.txt : we all like ice scream
00000000000000000000000000000000000000000000000000000000000000000101000000000111011011001000001000010100101011110001001001101001
00000000000000000000000000000000000000000000000000000000000000000101000000001110011010000000001100010000100001011011001010001001
0.21875
test/3.txt : we all scream for ice scream
test/5.txt : we all like ice scream
00000000000000000000000000000000000000000000000000000000000000001100010010011010010101000000000111011010000000001000010000101010
00000000000000000000000000000000000000000000000000000000000000000110110010100010010101000000001110011010000000001100010000100001
0.1875
100%|██████████████████████████████████████████████████████████████████████████████████████████████| 64/64 [00:00<00:00, 15947.92rotations/s]
found 1 duplicate(s) of file: test/7.txt
test/7.txt test/6.txt
found 1 duplicate(s) of file: test/1.txt
test/1.txt test/2.txt
found 2 duplicate(s) of file: test/3.txt
test/3.txt test/5.txt
test/3.txt test/4.txt
found 1 duplicate(s) of file: test/4.txt
test/4.txt test/5.txt
found 5 duplicate(s) in total
# 128 bit LSH
(dedup) $ find test -name "*.txt" | ./dedup.py - -v -b 128
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████| 7/7 [00:00<00:00, 829.01files/s]
0%| | 0/128 [00:00<?, ?rotations/s]test/1.txt : the cat sat in the hat
test/2.txt : the cat sat on the mat
00110000010010111010000110000100000111111100100001001000000000100110110000111010010001001001101101011010011000001110001010111001
00111010010010111010110110000000000111111000100100111000000000100010110000010010010001001101101101010010011100001010001010111001
0.1328125
test/3.txt : we all scream for ice scream
test/4.txt : we all scream for ice cream
11110111010000001011001000010000101100100000101000010011110010000011011100100111100110111100110011011000101101000110100011000110
11110111110000011011001000010000101100100000101000010011110010000111001110101111101110111100110011011000101101000110100011000110
0.0546875
test/7.txt : abcdefg
hijklmnop
test/6.txt : abcdefg
01110101000101011011101010010001100110011100101001010010000100100010110010011011010001101000100100000001000011000011000011101000
10011101110101011000101010010001110110010100111001111110000111101010100010011011010011101100100100110001101001011001000011110001
0.2421875
100%|█████████████████████████████████████████████████████████████████████████████████████████████| 128/128 [00:00<00:00, 6732.18rotations/s]
found 1 duplicate(s) of file: test/1.txt
test/1.txt test/2.txt
found 1 duplicate(s) of file: test/3.txt
test/3.txt test/4.txt
found 1 duplicate(s) of file: test/7.txt
test/7.txt test/6.txt
found 3 duplicate(s) in total
After computing an LSH for each document, the list of documents are sorted by their LSH and pairs of adjacent LSHs in the list are compared. If an adjacent pair's LSHs are similar enough (based on a threshold), then the pair is considered a duplicate. Considering the sorted list of LSHs once is not sufficient however, because if a pair of documents have high bits that differ, but most bits are still the same, they won't be considered, because they will be unlikely to be adjacent in the sorted list. To mitigate this, each LSH is rotated and the lists of LSHs are sorted again, and again adjacent pairs are checked.
This strategy requires a pass over each of N documents to compute the LSHs, but once the LSHs are computed, it requires only N-1 * M comparisons where M is size of the LSH in bits. When N is much larger than M, this is a significant improvement in time complexity over the naive approach of comparing N documents pairwise, which would require N choose 2 or N! / 2! * (N - 2)! comparisons.
If we have, say, over 100k documents to compare, performing N choose 2 comparisons becomes very expensive:
# ~241 hours is a long time
(dedup) $ find ~/Desktop/words -type f | ./dedup_pairwise.py - > ~/Desktop/duplicates_pairwise.tsv
100%|███████████████████████████████████████████████████████████████████████████████████████████| 128959/128959 [00:20<00:00, 6251.87files/s]
0%| | 1272202/8315147361 [01:58<241:35:38, 9559.05pairs/s
But, if we use the strategy of rotating the LSHs, we can find duplicates in a reasonable amount of time:
I implemented SimHash in Python. It took me a while to figure out all the bit fiddling, but I think it is pretty useful (although, it might require some experimentation to choose a good similarity threshold that minimizes false positives and false negatives on many data sets). This algorithm could be applied to data types other than text/strings (such as image or audio data), but I leave that as an exercise for the reader. Note that there are other efficient LSHs such as Nilsimsa or Minhash.
Also note that using Python’s built-in
hashas an underlying hashing function is not ideal because the hash values are salted with a seed that varies across different invocations of Python (at least since Python 3.3). I opted to usecityhashinstead, because it is a fast string hashing algorithm.Excerpting from the last section of the README, which is where the meat is:
Locality Sensitive Hashing
The script uses locality sensitive hashing (LSH) to compare texts. LSHs are a kind of hashing function that, unlike normal hashing functions, have the property that the probability of a hash collision is proportional to the similarity of objects hashed. In this case, the LSH implemented is simhash (described here).
Running
dedup.pywith the-v/--verboseoption prints the binary values of the computed LSHs and shows the similarity scores:After computing an LSH for each document, the list of documents are sorted by their LSH and pairs of adjacent LSHs in the list are compared. If an adjacent pair's LSHs are similar enough (based on a threshold), then the pair is considered a duplicate. Considering the sorted list of LSHs once is not sufficient however, because if a pair of documents have high bits that differ, but most bits are still the same, they won't be considered, because they will be unlikely to be adjacent in the sorted list. To mitigate this, each LSH is rotated and the lists of LSHs are sorted again, and again adjacent pairs are checked.
This strategy requires a pass over each of N documents to compute the LSHs, but once the LSHs are computed, it requires only N-1 * M comparisons where M is size of the LSH in bits. When N is much larger than M, this is a significant improvement in time complexity over the naive approach of comparing N documents pairwise, which would require N choose 2 or N! / 2! * (N - 2)! comparisons.
If we have, say, over 100k documents to compare, performing N choose 2 comparisons becomes very expensive:
But, if we use the strategy of rotating the LSHs, we can find duplicates in a reasonable amount of time: