Optimizing near-duplicate value search

I'm trying to find near duplicate values in a set of fields in order to allow an administrator to clean them up.

There are two criteria that I am matching on

1. One string is wholly contained within the other, and is at least 1/4 of its length
2. The strings have an edit distance less than 5% of the total length of the two strings

The Pseudo-PHP code:

foreach($values as $value){
$matches = array();
foreach($values as $match){
  if(
    (
      $value['length'] < $match['length']
      &&
      $value['length'] * 4 > $match['length']
      &&
      stripos($match['value'], $value['value']) !== false
    )
    ||
    (
      $match['length'] < $value['length']
      &&
      $match['length'] * 4 > $value['length']
      &&
      stripos($value['value'], $match['value']) !== false
    )
    ||
  开发者_Python百科  (
      abs($value['length'] - $match['length']) * 20 < ($value['length'] + $match['length'])
      &&
      0 < ($match['changes'] = levenshtein($value['value'], $match['value']))
      &&
      $match['changes'] * 20 <= ($value['length'] + $match['length'])
      )
    ){
      $matches[] = &$match;
    }
}
// output matches for current outer loop value
}

I've tried to reduce calls to the comparatively expensive stripos and levenshtein functions where possible, which has reduced the execution time quite a bit. However, as an O(n^2) operation this just doesn't scale to the larger sets of values and it seems that a significant amount of the processing time is spent simply iterating through the arrays.

Some properties of a few sets of values being operated on

Total   | Strings      | # of matches per string |          |
Strings | With Matches | Average | Median |  Max | Time (s) |
--------+--------------+---------+--------+------+----------+
    844 |          413 |     1.8 |      1 |   58 |    140   |
    593 |          156 |     1.2 |      1 |    5 |     62   | 
    272 |          168 |     3.2 |      2 |   26 |     10   |
    157 |           47 |     1.5 |      1 |    4 |      3.2 |
    106 |           48 |     1.8 |      1 |    8 |      1.3 |
     62 |           47 |     2.9 |      2 |   16 |      0.4 |

Are there any other things I can do to reduce the time to check criteria, and more importantly are there any ways for me to reduce the number of criteria checks required (for example, by pre-processing the input values), since there is such low selectivity?

---

Edit: Implemented solution

// $values is ordered from shortest to longest string length
$values_count = count($values); // saves a ton of time, especially on linux
for($vid = 0; $vid < $values_count; $vid++){
for($mid = $vid+1; $mid < $values_count; $mid++){ // only check against longer strings
  if(
    (
      $value['length'] * 4 > $match['length']
      &&
      stripos($match['value'], $value['value']) !== false
    )
    ||
    (
      ($match['length'] - $value['length']) * 20 < ($value['length'] + $match['length'])
      &&
      0 < ($changes = levenshtein($value['value'], $match['value']))
      &&
      $changes * 20 <= ($value['length'] + $match['length'])
      )
    ){
      // store match in both directions
      $matches[$vid][$mid] = true;
      $matches[$mid][$vid] = true;
    }

}
}
// Sort outer array of matches alphabetically with uksort()
foreach($matches as $vid => $mids){
  // sort inner array of matches by usage count with uksort()
  // output matches
}

---

You could first order the strings by length ( O(N) ) and then only check smaller strings to be substrings or larger strings, plus only check with levenshtein in string pairs for which the difference is not too large.

You already perform these checks, but now you do it for all N x N pairs, while preselecting first by length will help you reduce the pairs to check first. Avoid the N x N loop, even if it contains only tests that will fail.

For substring matching you could further improve by creating an index for all smaller items, and update this accordingly as you parse larger items. The index should can form a tree structure branching on letters, where each word (string) forms a path from root to leaf. This way you can find if any of the words in the index compare to some string to match. For each character in your match string try to proceed any pointers in the tree index, and create a new pointer at the index. If a pointer can not be proceeded to a following character in the index, you remove it. If any pointer reaches a leaf note, you've found a substring match. Implementing this is, I think, not difficult, but not trivial either.

---

You can get an instant 100% improvement by tightening your inner loop. Aren't you getting duplicate matches in your results?

For a preprocess step I'd go through and calculate character frequencies (assuming your set of characters is small like a-z0-9, which, given that you're using stripos, I think is likely). Then rather than comparing sequences (expensive) compare frequencies (cheap). This will give you false positives which you can either live with, or plug into the test you've currently got to weed out.