in reply to Re^2: Fast common substring matching
in thread Fast common substring matching

I came up with an algorithm inspired by bzip's algorithm of generating all substrings and then sorting them. I tried yours on a list of 20 strings of 1000 chars, and it ran in 153 seconds. Mine ran in 0.67 seconds, yielding the same results. 30 strings of 3000 chars runs in 20.3 seconds on mine; scaling up from there starts to get painful, but I would guess the OP's requirement of 300 strings of 3000 chars would run in under an hour, if it had plenty of memory (there will be 900,000 strings averaging 1500 chars in length).

Give it a whirl. 

use warnings;
use strict;
use Time::HiRes;

if (@ARGV == 0) {
    print "Finds longest matching substring between any pair of test s
+trings\n";
    print "in the given file. Pairs of lines are expected with the fir
+st of a\n";
    print "pair being the string name and the second the test string."
+;
    exit (1);
}

my $minmatch = 4;

my $startTime = [Time::HiRes::gettimeofday ()];

my @strings;
while (<>) {
  chomp(my $label = $_);
  chomp(my $string = <>);
  # Compute all substrings
  push @strings, map [substr($string, $_), $label, $_], 0..(length($st
+ring) - $minmatch);
}

print "Loaded. Sorting...\n";

@strings = sort {$a->[0] cmp $b->[0]} @strings;

print "Sorted. Finding matches...\n";

# Now walk through the list. The best match for each string will be th
+e
# previous or next element in the list that is not from the original s
+ubstring,
# so for each entry, just look for the next one. See how many initial 
+letters
# match and track the best matches
my @matchdata = (0); # (length, index1-into-strings, index2-into-strin
+gs)
for my $i1 (0..($#strings - 1)) {
  my $i2 = $i1 + 1;
  ++$i2 while $i2 <= $#strings and $strings[$i2][1] eq $strings[$i1][1
+];
  next if $i2 > $#strings;
  my ($common) = map length, ($strings[$i1][0] ^ $strings[$i2][0]) =~ 
+/^(\0*)/;
  if ($common > $matchdata[0]) {
    @matchdata = ($common, [$i1, $i2]);
  }
  elsif ($common == $matchdata[0]) {
    push @matchdata, [$i1, $i2];
  }
}

print "Best match: $matchdata[0] chars\n";
for my $i (@matchdata[1..$#matchdata]) {
print "$strings[$i->[0]][1] starting at $strings[$i->[0]][2]"
 . " and $strings[$i->[1]][1] starting at $strings[$i->[1]][2]\n";
}

print "Completed in " . Time::HiRes::tv_interval ($startTime) . "\n";

[download]
A test-data generating program follows 
use warnings;
use strict;

my ($howmany, $howlong) = (20, 1000);
# Generate $howmany strings of $howlong characters
for my $s (1..$howmany) {
  print "'String $s'\n";
  my $str = '';
  $str .= (qw(A C G T))[rand 4] for 1..$howlong;
  print "$str\n";
}

[download]
Caution: Contents may have been coded under pressure.

Replies are listed 'Best First'.
Re^4: Fast common substring matching
by bioMan (Beadle) on Nov 29, 2005 at 16:53 UTC

    Roy

    There is one difference between your algorithm and Grandfather's. His code returns the longest substring for each pair of input strings.

    With my original data set your code returns one substring. Grandfather's code returned over three thousand (where $minmatch = 256). On the other hand your code finds multiple occurrences of the longest common substrings, if they all have the same length, which I like.

    Mike

[reply]
      Yes, after I came up with my algorithm, I realized what all the output from GrandFather's code meant. I had thought it was just some sort of cryptic progress meter. :-)

      The (reasonably) obvious way to get the longest substring for each pair of input strings would be to run my algorithm using each pair of strings as input rather than the whole list of strings. That's probably more work than GF's method, though. I thought about trying it, but something shiny caught my attention...

      Update: but now I've done it. It runs on 20 strings of 1000 characters in something under 10 seconds for me. 100 strings of 1000 characters takes about 4 minutes.

      Read more... (4 kB)
      Caution: Contents may have been coded under pressure.
[reply]
[d/l]

        I had thought it was just some sort of cryptic progress meter. :-)

        LOL - I know what you mean.

        I'm still going over your original code to see how you did what you did -- trying to learn some perl :-)

        I'll give the new code a try. I also see that the minimum length in your code doesn't have to be a power of 2. This should allow me to analyze a limit boundary that appears to be present in my data. Grandfather's code allowed me to come up with what I feel is a pretty good estimate for the value of the limit, but this should allow a closer examination of the limit.

        Mike

[reply]

        Update: Important on Windows is starting the shared-manager process immediately if construction for the shared variable comes after loading data. Unix platforms benefit from Copy-on-Write feature which is great.

        ...

use MCE::Hobo;
use MCE::Shared;

# For minimum memory consumption, start the shared-manager process bef
+ore
# loading data locally.

MCE::Shared->start();   # <-- important on Windows

my $minmatch = 4;

my $startTime = [Time::HiRes::gettimeofday ()];

my %strings;
while (<>) {
  chomp(my $label = $_);
  chomp(my $string = <>);
  # Compute all substrings
  @{$strings{$label}} = map [substr($string, $_), $label, $_], 0..(len
+gth($string) - $minmatch);
}

print "Loaded. Generating combos...\n";

my @keys = sort keys %strings;

my $sequence = MCE::Shared->sequence(
   { chunk_size => 1, bounds_only => 1 }, 0, $#keys - 1
);

...

        [download]



        Hello Roy Johnson,

        I am fascinated by the various examples posted here, here, and also the Inline C demonstration.

        Your 2nd demonstration scales wonderfully on multiple cores after loading the strings hash. For testing, I made a file containing 48 sequences. The serial and parallel code complete in 22.6 seconds and 6.1 seconds respectively. My laptop has 4 real cores plus 4 hyper-threads.

        First, the construction for MCE::Hobo. This requires a later 1.699_011 dev release or soon after the final MCE 1.7 release.

        ...

print "Loaded. Generating combos...\n";

my @keys = sort keys %strings;

# Now walk through the list. The best match for each string will be th
+e
# previous or next element in the list that is not from the original s
+ubstring,
# so for each entry, just look for the next one. See how many initial 
+letters
# match and track the best matches

use MCE::Hobo;
use MCE::Shared;

my $sequence = MCE::Shared->sequence(
   { chunk_size => 1, bounds_only => 1 }, 0, $#keys - 1
);

sub walk_list {
  my @best_overall_match = (0);

  # $beg and $end have the same values when chunk_size => 1

  while ( my ( $beg, $end ) = $sequence->next ) {
    for my $ki1 ( $beg .. $end ) {
      for my $ki2 (($ki1 + 1)..$#keys) {

        my @strings = sort {$a->[0] cmp $b->[0]} @{$strings{$keys[$ki1
+]}}, @{$strings{$keys[$ki2]}};

        my @matchdata = (0); # (length, index1-into-strings, index2-in
+to-strings)

        for my $i1 (0..($#strings - 1)) {
          my $i2 = $i1 + 1;
          ++$i2 while $i2 <= $#strings and $strings[$i2][1] eq $string
+s[$i1][1];
          next if $i2 > $#strings;
          my ($common) = map length, ($strings[$i1][0] ^ $strings[$i2]
+[0]) =~ /^(\0*)/;
          next if $common < $minmatch;
          if ($common > $matchdata[0]) {
            @matchdata = ($common, [$i1, $i2]);
          }
          elsif ($common == $matchdata[0]) {
            push @matchdata, [$i1, $i2];
          }
        }

        next if $matchdata[0] < $minmatch;

        if ($matchdata[0] > $best_overall_match[0]) {
          @best_overall_match = ($matchdata[0]);
        }
        if ($matchdata[0] >= $best_overall_match[0]) {
          push @best_overall_match, map {
            ["$strings[$_->[0]][1]:$strings[$_->[0]][2]", "$strings[$_
+->[1]][1]:$strings[$_->[1]][2]"]
          } @matchdata[1..$#matchdata];
        }

      } # $ki2
    } # $ki1
  }

  return @best_overall_match;
};

MCE::Hobo->create( \&walk_list ) for 1 .. 8;

my @best_overall_match = (0);

for my $hobo ( MCE::Hobo->list ) {
  my @ret = $hobo->join;
  if ( $ret[0] > $best_overall_match[0] ) {
    @best_overall_match = @ret;
  }
  elsif ( $ret[0] == $best_overall_match[0] ) {
    shift @ret;
    push  @best_overall_match, @ret;
  }
}

print "Best overall match: $best_overall_match[0] chars\n";

...

        [download]



        MCE::Loop is next and does the same thing.

        ...

print "Loaded. Generating combos...\n";

my @keys = sort keys %strings;

# Now walk through the list. The best match for each string will be th
+e
# previous or next element in the list that is not from the original s
+ubstring,
# so for each entry, just look for the next one. See how many initial 
+letters
# match and track the best matches

use MCE::Loop;

MCE::Loop::init(
   max_workers => 8,
   chunk_size  => 1,
   bounds_only => 1,
);

my @ret = mce_loop_s {
  my ( $mce, $seq, $chunk_id ) = @_;
  my @best_overall_match = (0);

  # $seq->[0] and $seq->[1] have the same values when chunk_size => 1

  for my $ki1 ( $seq->[0] .. $seq->[1] ) {
    for my $ki2 (($ki1 + 1)..$#keys) {

      my @strings = sort {$a->[0] cmp $b->[0]} @{$strings{$keys[$ki1]}
+}, @{$strings{$keys[$ki2]}};

      my @matchdata = (0); # (length, index1-into-strings, index2-into
+-strings)

      for my $i1 (0..($#strings - 1)) {
        my $i2 = $i1 + 1;
        ++$i2 while $i2 <= $#strings and $strings[$i2][1] eq $strings[
+$i1][1];
        next if $i2 > $#strings;
        my ($common) = map length, ($strings[$i1][0] ^ $strings[$i2][0
+]) =~ /^(\0*)/;
        next if $common < $minmatch;
        if ($common > $matchdata[0]) {
          @matchdata = ($common, [$i1, $i2]);
        }
        elsif ($common == $matchdata[0]) {
          push @matchdata, [$i1, $i2];
        }
      }

      next if $matchdata[0] < $minmatch;

      if ($matchdata[0] > $best_overall_match[0]) {
        @best_overall_match = ($matchdata[0]);
      }
      if ($matchdata[0] >= $best_overall_match[0]) {
        push @best_overall_match, map {
          ["$strings[$_->[0]][1]:$strings[$_->[0]][2]", "$strings[$_->
+[1]][1]:$strings[$_->[1]][2]"]
        } @matchdata[1..$#matchdata];
      }

    } # $ki2
  } # $ki1

  MCE->gather(\@best_overall_match);

} 0, $#keys - 1;

MCE::Loop::finish;

my @best_overall_match = (0);

for my $i ( 0 .. $#ret ) {
  if ($ret[$i]->[0] > $best_overall_match[0]) {
    @best_overall_match = @{ $ret[$i] };
  }
  elsif ( $ret[$i]->[0] == $best_overall_match[0] ) {
    shift @{ $ret[$i] };
    push  @best_overall_match, @{ $ret[$i] };
  }
}

print "Best overall match: $best_overall_match[0] chars\n";

...

        [download]



        This has been a lot of fun. I learned some more Perl from it all.

        Regards, Mario

[reply]
[d/l]
[select]
Re^4: Fast common substring matching
by marioroy (Prior) on Feb 17, 2016 at 06:39 UTC

    Greetings Roy Johnson,

    I learn more Perl from reading your code. It took me a while, but somehow overlooked the line sorting $strings before walking through the list.

    Today, came to realization that Perl can do bitwise operations on strings.

    my ($common) = map length, ($strings[$i1][0] ^ $strings[$i2][0]) =~ /^
+(\0*)/;

    [download]

    Thank you for this.

[reply]
[d/l]

      Update: Added options chunk_size and bounds_only to MCE::Shared::Sequence in trunk, similar to MCE options. This allows MCE::Hobo workers to run as fast as MCE workers. Also, corrected the demonstration. Seeing this run faster than serial code made my day.

      I had to try something with the upcoming MCE 1.7 release. Parallelism may be beneficial for big sequences. MCE 1.7 will ship with MCE::Hobo, a threads-like module for processes. Thus, benefiting from Copy-on-Write feature of modern OS'es. In essence, the @strings array is not copied per each worker unless written to by the worker.

      Using Roy Johnson's demonstration, made the following changes to enable parallelism via MCE::Hobo workers. This requires MCE in trunk or a later dev 1.699_011 release.

      ...

print "Sorted. Finding matches...\n";

# Now walk through the list. The best match for each string will be th
+e
# previous or next element in the list that is not from the original s
+ubstring,
# so for each entry, just look for the next one. See how many initial 
+letters
# match and track the best matches
#
# my @matchdata = (0); # (length, index1-into-strings, index2-into-str
+ings)
# for my $i1 (0..($#strings - 1)) {
#   my $i2 = $i1 + 1;
#   ++$i2 while $i2 <= $#strings and $strings[$i2][1] eq $strings[$i1]
+[1];
#   next if $i2 > $#strings;
#   my ($common) = map length, ($strings[$i1][0] ^ $strings[$i2][0]) =
+~ /^(\0*)/;
#   if ($common > $matchdata[0]) {
#     @matchdata = ($common, [$i1, $i2]);
#   }
#   elsif ($common == $matchdata[0]) {
#     push @matchdata, [$i1, $i2];
#   }
# }

use MCE::Hobo;
use MCE::Shared;

my $sequence = MCE::Shared->sequence(
   { chunk_size => 500, bounds_only => 1 }, 0, $#strings - 1
);

sub walk_list {
  my @matchdata = (0); # (length, index1-into-strings, index2-into-str
+ings)

  while ( my ( $beg, $end ) = $sequence->next ) {
    for my $i1 ( $beg .. $end ) {
      my $i2 = $i1 + 1;
      ++$i2 while $i2 <= $#strings and $strings[$i2][1] eq $strings[$i
+1][1];
      next if $i2 > $#strings;
      my ($common) = map length, ($strings[$i1][0] ^ $strings[$i2][0])
+ =~ /^(\0*)/;
      if ($common > $matchdata[0]) {
        @matchdata = ($common, [$i1, $i2]);
      }
      elsif ($common == $matchdata[0]) {
        push @matchdata, [$i1, $i2];
      }
    }
  }

  return @matchdata;
};

MCE::Hobo->create( \&walk_list ) for 1 .. 8;

my @matchdata = (0); # (length, index1-into-strings, index2-into-strin
+gs)

for my $hobo ( MCE::Hobo->list ) {
  my @ret = $hobo->join;
  if ( $ret[0] > $matchdata[0] ) {
    @matchdata = @ret;
  }
  elsif ( $ret[0] == $matchdata[0] ) {
    shift @ret;
    push  @matchdata, @ret;
  }
}

print "Best match: $matchdata[0] chars\n";

...

      [download]

      MCE 1.7 is nearly completed in trunk. The MCE::Shared::Sequence module is helpful. I will try to finish MCE 1.7 by the end of the month.

      Regards, Mario

[reply]
[d/l]

      Update: The update to MCE::Shared::Sequence in trunk allows MCE::Hobo workers to run as fast as MCE workers. Thank you for this. The MCE::Hobo demonstration made me realized the need to beef up MCE::Shared::Sequence with chunk_size and bounds_only options similar to MCE options.

      Using Roy Johnson's demonstration, made the following changes to enable parallelism via MCE::Loop.

      ...

print "Sorted. Finding matches...\n";

use MCE::Loop;

MCE::Loop::init(
   max_workers => 8,
   chunk_size  => 500,
   bounds_only => 1,
);

my @ret = mce_loop_s {
  my ( $mce, $seq, $chunk_id ) = @_;
  my @matchdata = (0); # (length, index1-into-strings, index2-into-str
+ings)

  for my $i1 ( $seq->[0] .. $seq->[1] ) {
    my $i2 = $i1 + 1;
    ++$i2 while $i2 <= $#strings and $strings[$i2][1] eq $strings[$i1]
+[1];
    next if $i2 > $#strings;
    my ($common) = map length, ($strings[$i1][0] ^ $strings[$i2][0]) =
+~ /^(\0*)/;
    if ($common > $matchdata[0]) {
      @matchdata = ($common, [$i1, $i2]);
    }
    elsif ($common == $matchdata[0]) {
      push @matchdata, [$i1, $i2];
    }
  }

  MCE->gather( \@matchdata );

} 0, $#strings - 1;

my @matchdata = (0); # (length, index1-into-strings, index2-into-strin
+gs)

for my $i ( 0 .. $#ret ) {
  if ( $ret[$i]->[0] > $matchdata[0] ) {
    @matchdata = @{ $ret[$i] };
  }
  elsif ( $ret[$i]->[0] == $matchdata[0] ) {
    shift @{ $ret[$i] };
    push  @matchdata, @{ $ret[$i] };
  }
}

print "Best match: $matchdata[0] chars\n";

...

      [download]
[reply]
[d/l]
Re^4: Fast common substring matching
by marioroy (Prior) on Feb 17, 2016 at 20:31 UTC
[reply]

In Section Cool Uses for Perl