用于相似度计算的 Neo4j Cypher 查询性能缓慢

Question

我是 Neo4j/Graph 数据库的新手，我正在尝试复制 Cypher 食谱中的教程：http://docs.neo4j.org/chunked/stable/cypher-cookbook-similarity-calc.html

随机数据集包含 100 种食物和 1500 人，所有人都通过具有“倍”整数属性的 ATE 关系与食物相关。食物和人被标记并具有属性“名称” - 由自动索引索引

neo4j-sh (?)$ dbinfo -g "Primitive count"
{
  "NumberOfNodeIdsInUse": 1600,
  "NumberOfPropertyIdsInUse": 151600,
  "NumberOfRelationshipIdsInUse": 150000,
  "NumberOfRelationshipTypeIdsInUse": 1
}

neo4j-sh (?)$ index --indexes
Node indexes:
  node_auto_index
Relationship indexes:
  relationship_auto_index

在 neo4j-shell 中从说明书运行修改后的查询永远不会完成（可能是因为节点/关系太多？）：

EXPORT name="Florida Goyette"
MATCH (me:Person { name: {name}})-[r1:ATE]->(food)<-[r2:ATE]-(you:Person)
WITH me,count(DISTINCT r1) AS H1,count(DISTINCT r2) AS H2,you
MATCH (me)-[r1:ATE]->(food)<-[r2:ATE]-(you)
RETURN SUM((1-ABS(r1.times/H1-r2.times/H2))*(r1.times+r2.times)/(H1+H2)) AS similarity
LIMIT 100;

因此，我开始研究如何将之前限制为“第一”100 人并得出以下结论：

EXPORT name="Florida Goyette"
MATCH (me:Person { name: {name} })-[r1:ATE]->(food)
WITH me, food
MATCH (food)<-[r2:ATE]-(you)
WHERE me <> you
WITH me, you
LIMIT 100
MATCH (me)-[r1:ATE]->(food)<-[r2:ATE]-(you)
WITH me, count(DISTINCT r1) AS H1, count(DISTINCT r2) AS H2, you
MATCH (me)-[r1:ATE]->(food)<-[r2:ATE]-(you)
WITH me, you, SUM((1-ABS(r1.times/H1-r2.times/H2))*(r1.times+r2.times)/(H1+H2)) AS similarity
RETURN me.name, you.name, similarity
ORDER BY similarity DESC;

但是这个查询在预热缓存上表现很差

100 rows
16038 ms

对于“实时”使用，是否有机会使此类查询执行得更快？

系统和 Neo4j

Windows 7（64 位）、Intel Core I7-2600K、8GB RAM、SSD 驱动器上的 Neo4j 数据库。

Neo4j 社区版本：2.1.0-M01（也在 2.0.1 stable 上测试过）

neo4j-community.options

 -Xmx2048m
 -Xms2048m

neo4j.properties

neostore.nodestore.db.mapped_memory=200M
neostore.relationshipstore.db.mapped_memory=200M
neostore.propertystore.db.mapped_memory=200M
neostore.propertystore.db.strings.mapped_memory=330M
neostore.propertystore.db.arrays.mapped_memory=330M

node_auto_indexing=true
node_keys_indexable=name
relationship_auto_indexing=true
relationship_keys_indexable=times

Cypher dump of my data（503kb 压缩）

PROFILE 输出

ColumnFilter(symKeys=["similarity", "you", "you.name", "me", "me.name"], returnItemNames=["me.name", "you.name", "similarity"], _rows=100, _db_hits=0)
Sort(descr=["SortItem(similarity,false)"], _rows=100, _db_hits=0)
  Extract(symKeys=["me", "you", "similarity"], exprKeys=["me.name", "you.name"], _rows=100, _db_hits=200)
    ColumnFilter(symKeys=["me", "you", "  INTERNAL_AGGREGATEcb085cf5-8982-4a83-ba3d-9642de570c59"], returnItemNames=["me", "you", "similarity"], _rows=100, _db_hits=0)
      EagerAggregation(keys=["me", "you"], aggregates=["(INTERNAL_AGGREGATEcb085cf5-8982-4a83-ba3d-9642de570c59,Sum(Divide(Multiply(Subtract(Literal(1),AbsFunction(Subtract(Divide(Property(r1,times(1)),H1),Divide(Property(r2,times(1)),H2)))),Add(Property(r1,times(1)),Property(r2,times(1)))),Add(H1,H2))))"], _rows=100, _db_hits=40000)
        SimplePatternMatcher(g="(you)-['r2']-(food),(me)-['r1']-(food)", _rows=10000, _db_hits=0)
          ColumnFilter(symKeys=["me", "you", "  INTERNAL_AGGREGATE677cd11c-ae53-4d7b-8df6-732ffed28bbf", "  INTERNAL_AGGREGATEb5eb877c-de01-4e7a-9596-03cd94cfa47a"], returnItemNames=["me", "H1", "H2", "you"], _rows=100, _db_hits=0)
            EagerAggregation(keys=["me", "you"], aggregates=["(  INTERNAL_AGGREGATE677cd11c-ae53-4d7b-8df6-732ffed28bbf,Distinct(Count(r1),r1))", "(  INTERNAL_AGGREGATEb5eb877c-de01-4e7a-9596-03cd94cfa47a,Distinct(Count(r2),r2))"], _rows=100, _db_hits=0)
              SimplePatternMatcher(g="(you)-['r2']-(food),(me)-['r1']-(food)", _rows=10000, _db_hits=0)
                ColumnFilter(symKeys=["me", "food", "you", "r2"], returnItemNames=["me", "you"], _rows=100, _db_hits=0)
                  Slice(limit="Literal(100)", _rows=100, _db_hits=0)
                    Filter(pred="NOT(me == you)", _rows=100, _db_hits=0)
                      SimplePatternMatcher(g="(you)-['r2']-(food)", _rows=100, _db_hits=0)
                        ColumnFilter(symKeys=["food", "me", "r1"], returnItemNames=["me", "food"], _rows=1, _db_hits=0)
                          Filter(pred="Property(me,name(0)) == {name}", _rows=1,_db_hits=148901)
                            TraversalMatcher(start={"label": "Person", "producer": "NodeByLabel", "identifiers": ["me"]}, trail="(me)-[r1:ATE WHERE true AND true]->(food)", _rows=148901, _db_hits=148901)

Answer 1

您正在多次执行相同的 MATCHing。这样做效果更好吗？

EXPORT name="Florida Goyette"
MATCH (me:Person { name: {name}})-[r1:ATE]->(food)<-[r2:ATE]-(you:Person)
WITH me,r1,r2,count(DISTINCT r1) AS H1,count(DISTINCT r2) AS H2,you
LIMIT 100 
RETURN SUM((1-ABS(r1.times/H1-r2.times/H2))*(r1.times+r2.times)/(H1+H2)) AS similarity;

Answer 2

您使用了错误的索引类型。使用

创建标签索引

CREATE INDEX ON :Person(name)

检查架构索引和约束

neo4j-shell

schema
schema ls -l :User 

或

neo4j 浏览器

:schema 
:schema ls -l :User

可能需要对查询进行优化，但从这里开始。

Answer 3

1. 在 Windows 上，内存映射在 inside 堆中。因此，将堆大小增加到 4G。

2. 您不需要旧的自动索引，而是需要像 jjaderberg 这样的新模式索引。

3. 这会返回多少行？

.

MATCH (me:Person { name: {name}})-[r1:ATE]->(food)<-[r2:ATE]-(you:Person) RETURN count(*)

还有多少：

MATCH (me:Person { name: {name}})-[r1:ATE]->(food)<-[r2:ATE]-(you:Person)
WITH me,count(DISTINCT r1) AS H1,count(DISTINCT r2) AS H2,you
MATCH (me)-[r1:ATE]->(food)<-[r2:ATE]-(you)
RETURN COUNT(*)

你也可以避免匹配两次：

MATCH (me:Person { name: {name}})-[r1:ATE]->(food)<-[r2:ATE]-(you:Person)
WITH me,
     collect([r1,r2]) as rels, 
     count(DISTINCT r1) AS H1,
     count(DISTINCT r2) AS H2,
     you

RETURN me,you, 
       reduce(a=0,r in rels | 
              a + (1-ABS(r[0].times/H1-r[1].times/H2))*
                  (r[0].times+r[1].times)
                  /(H1+H2) as similarity

顺便说一句。如果您使用您的域、用例和一些示例数据创建了一个 GraphGist，那就太棒了！