My Implementation of the Apriori Algorithm

this is a self imposed machine problem i wrote over a frantic afternoon for my lesson on frequent itemsets and the apriori algorithm .

i wanted to write a program that would find the top five frequent item sets among a set of baskets.

consider the item set below:

d = [ ('milk', 'cheese', 'tuna'),
      ('cheese', 'mushroom'),
      ('cheese', 'eggs'),
      ('milk', 'cheese', 'mushroom'),
      ('milk', 'eggs'),
      ('cheese', 'eggs'),
      ('milk', 'cheese', 'mushroom', 'tuna'), 
      ('milk', 'cheese', 'eggs') ]

the goal i want for my program is to be able to induce that

(the presence of milk somehow implies the presence of cheese as well)

or

(milk, cheese and mushroom) somehow go together.

just right now, i think i could also find a correlation between item sets and dates.

anyway, as i have learned the key is to use map reduce and i was able to do this quite easily with a nosql database like mongo db .

a link to the code i wrote is available on github .

so far, i am able to generate the candidate item set for 2-combinations of the universal sets. this is another way of saying that the program i wrote is to count the number of times a two combination appears in all the baskets. for example, it is able to count that (‘milk’, ‘cheese’) occurs four times. below is a screenshot of the script doing just that:

the next step after this is to generate the frequent item sets from the the candidate item set. this just means, i have to filter out non frequent 2-combinations. this can easily be done in two steps and is based on a support count that is defined before hand.

let’s say a support count of 2 is defined, then the first thing to do is to check if the count of the 1 item sets that make up the two item sets have support count . remember that the key to the apriori algorithm is that the subsets of a frequent item set must also be frequent.

this can easily be derived by running the script i wrote:

now that i am able to generate the candidate item sets for any n-combination of the universal set, some next steps for my program would be to

1. generate the frequent item set given a candidate item set.
2. report the top 5 item sets with the highest interest and highest confidence.

performance considerations also come to mind as my implementation of the algorithm has not been tested for a large data set.