Defining Domain Specific Blocking And Similarity Functions

To tune Zingg for more accurate matching with higher recall

You can add your own blocking functions which will be evaluated by Zingg to build the blocking tree.

The blocking tree works on the matched records provided by the user as part of the training. At every node, it selects the hash function and the field on which it should be applied so that there is the least elimination of the matching pairs. Say we have data like this :

Pair 1	firstname	lastname
Record A	john	doe
Record B	johnh	d oe

Pair 2	firstname	lastname
Rrecord A	mary	ann
Record B	marry

Let us assume we have hash function first1char and we want to check if it is a good function to apply to firstname :

Pair	Record	Output
1	Record A	j
1	Record B	j
2	Record A	m
2	Record B	m

There is no elimination in the pairs above, hence it is a good function.

Now let us try last1char on firstname :

Pair	Record	Output
1	Record A	n
1	Record B	h
2	Record A	y
2	Record B	y

Pair 1 is getting eliminated above, hence last1char is not a good function.

So, first1char(firstname) will be chosen. This brings near similar records together - in a way, clusters them to break the cartesian join.

These business-specific blocking functions go into Hash Functions and must be added to HashFunctionRegistry and hash functions config.

Also, for similarity, you can define your own measures. Each dataType has predefined features, for example, String fuzzy type is configured for Affine and Jaro.

You can define your own comparisons and use them.