[SimHash] the Hash-based Similarity Detection Algorithm

The current information explosion has resulted in an increasing number of applications that need to deal with large volumes of data. While many of the data contains useless redundancy data, especially in mass media, web crawler/analytic fields, wasted many precious resources (power, bandwidth, CPU and storage, etc.). This has resulted in an increased interest in algorithms that process the input data in restricted ways.

But traditional hash algorithms have two problems, first it assumes that the data fits in main memory, it is unreasonable when dealing with massive data such as multimedia data, web crawler/analytic repositories and so on. And second, traditional hash can only indentify the identical data. this brings to light the importance of simhash.

 

Simhash 5 steps: Tokenize, Hash, Weigh Values, Merge, Dimensionality Reduction

• tokenize

  • tokenize your data, assign weights to each token, weights and tokenize function are depend on your business

• hash (md5, SHA1)

  • calculate token's hash value and convert it to binary (101011 )

• weigh values

  • for each hash value, do hash*w, in this way: (101011 ) -> (w,-w,w,-w,w,w)

• merge

  • add up tokens' values, to merge to 1 hash, for example, merge (4 -4 -4 4 -4 4) and (5 -5 5 -5 5 5) , results to (4+5 -4+-5 -4+5 4+-5 -4+5 4+5)，which is (9 -9 1 -1 1)

• Dimensionality Reduction

  • Finally, signs of elements of V corresponds to the bits of the final fingerprint, for example (9 -9 1 -1 1) -> (1 0 1 0 1), we get 10101 as the fingerprint.

 

How to use SimHash fingerprints?

Hamming distance can be used to find the similarity between two given data, calculate the Hamming distance between 2 fingerprints.

Based on my experience, for 64 bit SimHash values, with elaborate weight values,  distance of similar data often differ appreciably in magnitude from those unsimilar data.

 

how to calculate Hamming distance: 

　　XOR, 只有兩個位不一樣時結果是1 ，不然爲0，兩個二進制value「異或」後獲得1的個數 爲海明距離 。

 

  

SimHash algorithm, introduced by Charikar and is patented by Google.

simhash 0.1.0 : Python Package Index