Similarity join using Hadoop

Question

I'm new to hadoop. I'd like to run some approaches with you that I came up with.

Problem:
2 datasets : A and B.
Both datasets represent songs: some top level attributes, titles (1..), performers (1..).
I need to match these datasets either using equality or fuzzy algorithms (such as levenshtein , jaccard, jaro-winkler, etc) based on titles and performer.
The dataset sizes are: A=20-30M , B~=1-6M.

So here there are approaches that I came up with:

1. Load dataset B(smallest) into HDFS. Use mapreduce against dataset A(biggest) , where:
   map phase : for each record in A access HDFS and pull records B for matching;
   reduce phase : writes id pairs

2. load dataset A into distirubted cache (i.e. jboss cache) in optimized form to speed up searching. Use mapreduce against dataset B, where :
   map phase: for each record in B query distributed cache for matching
   reduce : writes id pairs

3. use mapreduce to join both datasets, where
   map phase: gets a record from set A and set B , does matching
   reduce phase: same
   (I'm fuzzy about ths one. 1st: join will be the cartesian product with trillion of records; 2nd: not sure how hadoop can parallize that across cluster)

4. use hive (i'm looking at right now trying to figure out how to plugin custom functions that will do string matching)

I'm loooking for a pointers, which approach would be the best candidate or maybe there are some other approaches that I do not see.

Answer 1

You might want to look at these two papers by Jimmy Lin:

• Brute Force and Indexed Approaches to Pairwise Document Similarity
• Pairwise Document Similarity in Large Collections with MapReduce

The approach you take will be dependent on what kind of similarity metric you use, but a Lucene based approach may work here. You might also want to think of ways to partitioning the data to reduce the number of comparisons you need to make.

Answer 2

You might find this paper and code useful:

Efficient Parallel Set-Similarity Joins Using MapReduce

I've personally implemented it in Cascading with good results. Unfortunately the code is too domain specific to release.

The point of the above work is to reduce the number joins to the candidate pairs that are very likely similar, then the candidate pairs can be compared directly (in a MR join) using any cocktail of algorithms that are relevant. A good side effect is that this join can be performed evenly across the cluster without duplicate comparisons.

Ultimately this is an optimization on performing a cross join between two independent sets or within the same set (the second case implemented slightly differently than the first).

disclosure: I'm the author of Cascading