Similarity queries are widely used in applications where users need to find records that satisfy a similarity predicate, while exact matching is not sufficient. These queries are especially important for social and Web applications, where errors, abbreviations, and inconsistencies are common. As an example, we may want to find all the movies starring Schwarzenegger, while we don't know the exact spelling of his last name (despite his popularity in both the movie industry and politics :-)). As another example, we want to find all the Facebook users who have similar friends. To meet this type of needs, AsterixDB supports similarity queries using efficient indexes and algorithms.
AsterixDB supports edit distance (on strings) and Jaccard (on sets). For instance, in our TinySocial example, the friend-ids of a Facebook user forms a set of friends, and we can define a similarity between the sets of friends of two users. We can also convert a string to a set of grams of a length "n" (called "n-grams") and define the Jaccard similarity between the two gram sets of the two strings. Formally, the "n-grams" of a string are its substrings of length "n". For instance, the 3-grams of the string schwarzenegger are sch, chw, hwa, ..., ger.
AsterixDB provides tokenization functions to convert strings to sets, and the similarity functions.
The following query asks for all the Facebook users whose name is similar to Suzanna Tilson, i.e., their edit distance is at most 2.
use dataverse TinySocial;
for $user in dataset('FacebookUsers')
let $ed := edit-distance($user.name, "Suzanna Tilson")
where $ed <= 2
return $user
The following query asks for all the Facebook users whose set of friend ids is similar to [1,5,9], i.e., their Jaccard similarity is at least 0.6.
use dataverse TinySocial;
for $user in dataset('FacebookUsers')
let $sim := similarity-jaccard($user.friend-ids, [1,5,9])
where $sim >= 0.6f
return $user
AsterixDB allows a user to use a similarity operator ~= to express a condition by defining the similarity function and threshold using "set" statements earlier. For instance, the above query can be equivalently written as:
use dataverse TinySocial;
set simfunction "jaccard";
set simthreshold "0.6f";
for $user in dataset('FacebookUsers')
where $user.friend-ids ~= [1,5,9]
return $user
In this query, we first declare Jaccard as the similarity function using simfunction and then specify the threshold 0.6f using simthreshold.
AsterixDB supports fuzzy joins between two sets. The following query finds, for each Facebook user, all Twitter users with names similar to their name based on the edit distance.
use dataverse TinySocial;
set simfunction "edit-distance";
set simthreshold "3";
for $fbu in dataset FacebookUsers
return {
"id": $fbu.id,
"name": $fbu.name,
"similar-users": for $t in dataset TweetMessages
let $tu := $t.user
where $tu.name ~= $fbu.name
return {
"twitter-screenname": $tu.screen-name,
"twitter-name": $tu.name
}
};
AsterixDB uses two types of indexes to support similarity queries, namely "ngram index" and "keyword index".
An "ngram index" is constructed on a set of strings. We generate n-grams for each string, and build an inverted list for each n-gram that includes the ids of the strings with this gram. A similarity query can be answered efficiently by accessing the inverted lists of the grams in the query and counting the number of occurrences of the string ids on these inverted lists. The similar idea can be used to answer queries with Jaccard similarity. A detailed description of these techniques is available at this paper.
For instance, the following DDL statements create an ngram index on the FacebookUsers.name attribute using an inverted index of 3-grams.
use dataverse TinySocial;
create index fbUserIdx on FacebookUsers(name) type ngram(3);
The number "3" in "ngram(3)" is the length "n" in the grams. This index can be used to optimize similarity queries on this attribute using edit-distance, edit-distance-check, jaccard, or jaccard-check queries on this attribute where the similarity is defined on sets of 3-grams. This index can also be used to optimize queries with the "contains()" predicate (i.e., substring matching) since it can be also be solved by counting on the inverted lists of the grams in the query string.
A "keyword index" is constructed on a set of strings or sets (e.g., OrderedList, UnorderedList). Instead of generating grams as in an ngram index, we generate tokens (e.g., words) and for each token, construct an inverted list that includes the ids of the records with this token. The following two examples show how to create keyword index on two different types:
use dataverse TinySocial;
create index fbMessageIdx on FacebookMessages(message) type keyword;
for $o in dataset('FacebookMessages')
let $jacc := similarity-jaccard-check(word-tokens($o.message), word-tokens("love like verizon"), 0.2f)
where $jacc[0]
return $o
use dataverse TinySocial;
create index fbUserIdx_fids on FacebookUsers(friend-ids) type keyword;
for $c in dataset('FacebookUsers')
let $jacc := similarity-jaccard-check($c.friend-ids, {{3,10}}, 0.5f)
where $jacc[0]
return $c
As shown above, keyword index can be used to optimize queries with token-based similarity predicates, including similarity-jaccard and similarity-jaccard-check.