Near-Duplicate Detection

This program identifies near-duplicates in a corpus using techniques described by Professor William Arms of Cornell University in his lectures to the students of INFO 4300, Information Retrieval, Fall 2012.

This program was written by Parker Moore (pjm336), Fall 2012.

It is hosten on GitHub at https://github.com/parkr/near-dup-detection

Install

pip install git://github.com/parkr/near-dup-detection.git#egg=NearDuplicatesDetection

Usage

python ndd.py

Explanation of Methodology

All of the logic for the program is built into the Detector class (detector.py). This class contains the methods and instance variables needed to detect near-duplicates, such as the get_jaccard(file1, file2) method, the calculate_sketches() method and the fundamental create_3grams() method.

This program implements the standard procedure for detecting near-duplicates:

1. Generate n-grams (3-grams in this case) for each document. Assign these n-grams a unique ID based on a 64-bit hash.
2. Create 25 sketches for document based on 50 randomly selected numbers and some stuff we generated earlier:
   • p is the closest prime number to the # of n-grams
   • a_s random, in the range [1, p-1]
   • b_s random, in the range [0, p-1]
   • x is the n-gram ID (the hash generated in step 1)
   • using the equation: f_s(x) = (a_s*x + b_s) % p
   • note: this equation is calculated 25 times per document (one time per random pair a_s and b_s), but only the minimum result of f_s(x) for each of the 25 pairs is saved. Thus, at the end of the calculation, each document has 25 f_min's, one for each pair of random numbers.
3. Go through each document and compare it to all the other documents using the Jaccard coefficient estimation equation : J(d1, d2) = m/n, where:
   • m = number of sketch values (must be at the same index in respective lists) which are the same between the two documents
   • n = number of samples (# of sketches)
4. Having defined an arbitrary Jaccard coefficient threshold of 0.5, the program prints out the names of the documents whose Jaccard coefficient is greater than the threshold previously defined, as well as the corresponding Jaccard coefficient.

As an addendum to the project, the three "nearest neighbors" to the first ten documents is calculated at the end using the same method (and the data from before).

License

Standard GPLv2 license applies. Copyright (2012) Parker Moore.