Sampling in cypher Query

Question

I'm struggling with Cypher to perform a sampling over my user activity.

What does my graph look like

I have a few millions users recorder in my Graph with an indexed property UserId.

I have few hundreds Products with the indexed property ItemId.

My users can [INTERACTS] with my products.

What I'm trying to do

I would like to have an average idea over the path user-[INTERACTS]-product-[INTERACTS]-user-[INTERACTS]-product . In plain english I would like to know which products the look-alike user are interacting with. Eg if I interacts with products A and B, users interacting with these products generally interacts with these other products.

It's what Amazon do.

My problem

I can't simply match the above pattern, it takes way too long to execute. So I thought that I can only sample my user and that I could do the following :

Take only the last 20 products my user interacts with
Take only the last 20 users using each of the above products
Take only the last 20 products these users interacted with, and counting each product occurrence

But I don't know if this is even possible in one single Cypher query.

The closest I came was the following query. But it still is too long, and it does not do what I want. Basically it gives me the latest products used by the "latest" user which has the "latest" products in common with the root user, which seems logic to me, but fails to sample my users' activity.

START u=node:node_auto_index('UserId:9554')  
MATCH 
    u-[i1:INTERACTS]-p1  
WITH 
    u,p1,i1 
LIMIT 20  
MATCH 
    p1-[i2:INTERACTS]-u1 
WHERE 
    NOT(u1=u) 
WITH 
    i1,i2,u1,p1 
LIMIT 400  
MATCH 
    u1-[i3:INTERACTS]-p 
WHERE 
    NOT(p1=p) AND p.ProjectId = {ProjectId} AND p.IsActive? = 1 
RETURN 
    i1.Label, i2.Label,i3.Label, p.ItemId,count(p) as count
LIMIT 8000

Where am I now

After some more unsuccessful testing I tried to code it using the Java API, and it is way more simpler and straight forward. But out of curiosity and because for now my system use Cypher I would like to know how to do this

For the sake of testing I think my question could be reduce to : Given a pattern, what is the last 2 node of each depth.

I created http://console.neo4j.org/?id=inf2hn in order to test it. I think the final result I'm looking for should look like :

Product 2 | User 3 | Product 5
Product 2 | User 3 | Product 6
Product 2 | User 4 | Product 5
Product 2 | User 4 | Product 6
Product 3 | User 3 | Product 5
Product 3 | User 3 | Product 6
Product 3 | User 4 | Product 5
Product 3 | User 4 | Product 6

Thanks for your help

Luanne Luanne · Accepted Answer · 2013-06-12T13:41:20

I think what you mentioned in the comments reflects the fact that it may not just be a simple query. I've written recommendations using Neo4j and to me it looks like what you need it to account for all the various types of interactions with a product and how important they are. Based on that you form a ranked list of product interaction for a user i.e. products that the user is most "interested" in. To find products to recommend to another user, you can calculate how similar the user is to others, and then use those as the starting point for recommending products (if my tastes are similar to yours, I might be interested in the same products that you are).

Some blogs that might help:

http://blog.everymansoftware.com/2012/02/similarity-based-recommendation-engines.html

http://thought-bytes.blogspot.in/2012/02/similarity-based-recommendations-with.html

Disclaimer: I am the owner of the thought bytes blog.

Sampling in cypher Query

2 Answers