Added some notes on sampling to recommendation recipe CTR
Project: http://git-wip-us.apache.org/repos/asf/tinkerpop/repo Commit: http://git-wip-us.apache.org/repos/asf/tinkerpop/commit/3fe223bd Tree: http://git-wip-us.apache.org/repos/asf/tinkerpop/tree/3fe223bd Diff: http://git-wip-us.apache.org/repos/asf/tinkerpop/diff/3fe223bd Branch: refs/heads/TINKERPOP-1642 Commit: 3fe223bdcbf0695529aa9f5fd58b3bff573845b6 Parents: df285d3 Author: Stephen Mallette <[email protected]> Authored: Thu Mar 9 07:34:23 2017 -0500 Committer: Stephen Mallette <[email protected]> Committed: Thu Mar 9 07:34:23 2017 -0500 ---------------------------------------------------------------------- docs/src/recipes/recommendation.asciidoc | 45 ++++++++++++++++++++++++++- 1 file changed, 44 insertions(+), 1 deletion(-) ---------------------------------------------------------------------- http://git-wip-us.apache.org/repos/asf/tinkerpop/blob/3fe223bd/docs/src/recipes/recommendation.asciidoc ---------------------------------------------------------------------- diff --git a/docs/src/recipes/recommendation.asciidoc b/docs/src/recipes/recommendation.asciidoc index 8d5f1ec..0aaa7e4 100644 --- a/docs/src/recipes/recommendation.asciidoc +++ b/docs/src/recipes/recommendation.asciidoc @@ -245,4 +245,47 @@ g.V().has("person","name","alice").as("alice"). by(values, decr). by(select(keys).values("name")). unfold().select(keys).values("name") ----- \ No newline at end of file +---- + +In considering the practical applications of this recipe, it is worth revisiting the earlier "basic" version of the +reccomendation algorithm: + +[gremlin-groovy,existing] +---- +g.V().has('person','name','alice').as('her'). + out('bought').aggregate('self'). + in('bought').where(neq('her')). + out('bought').where(without('self')). + groupCount(). + order(local). + by(values, decr) +---- + +The above traversal performs a full ranking of items based on all the connected data. That could be a time consuming +operation depending on the number of paths being traversed. As it turns out, recommendations don't need to have perfect +knowledge of all data to provide a "pretty good" approximation of a recommendation. It can therefore make sense to +place additional limits on the traversal to have it better return more quickly at the expense of examining less data. + + +Gremlin provides a number of steps that can help with these limits like: +link:http://tinkerpop.apache.org/docs/x.y.z/reference/#coin-step[coin()], +link:http://tinkerpop.apache.org/docs/x.y.z/reference/#sample-step[sample()], and +link:http://tinkerpop.apache.org/docs/current/reference/#timelimit-step[timeLimit()]. For example, to have the +traversal sample the data for no longer than one second, the previous "basic" recommendation could be changed to: + +[gremlin-groovy,existing] +---- +g.V().has('person','name','alice').as('her'). + out('bought').aggregate('self'). + in('bought').where(neq('her')). + out('bought').where(without('self')).timeLimit(1000). + groupCount(). + order(local). + by(values, decr) +---- + +In using sampling methods, it is important to consider that the natural ordering of edges in the graph may not produce +an ideal sample for the recommendation. For example, if the edges end up being returned oldest first, then the +recommendation will be based on the oldest data, which would not be ideal. As with any traversal, it is important to +understand the nature of the graph being traversed and the behavior of the underlying graph database to properly +achieve the desired outcome. \ No newline at end of file
