The tutorial describes existing approaches to model graph databases and different techniques implemented in RDF and Database engines including their main drawbacks when a large volume of interconnected data needs to be traversed.
Global Lehigh Strategic Initiatives (without descriptions)
Semantic Data Management in Graph Databases
1. Semantic Data Management in
Graph Databases
-Tutorial at ESWC 2013-
Maria-Esther
Vidal
Edna
Ruckhaus
Maribel
Acosta
Cosmin
Basca
USB
1
2. 2
! " #
$ %% #
&
Network of Friends in a High School Annotation Graph representing relations
between clinical trails
Air-traffic between US cities A Fragment of Facebook Relationships between comments
about a topic in Twitter
Graphs …
3. 3
! " #
$ %% #
&
A significant increase of graph data in the form of social & biological information.
Tasks to be Solved …
Patternsof
connectionsbetween people to
understand functioning of society.
Topological properties
of graphs can be used to identify
patterns that reveal phenomena,
anomalies and potentially lead to
a discovery.
4. Tasks to be Solved … (2)
Relatedness
between two nodes.
4
The importance of the information relies on the relations more or equal
than on the entities.
[Anderson et al. 2012]
Best
matchingbetwee
n two set of sub-
graphs.
Keyword
queries.
Proximitypa
tterns.
Frequents
ub-graphs.
Graph
ranking.
5. Basic Graph Operations
Basic Graph operations that need to be efficiently
performed:
Graph traversal.
Measuring path length.
Finding the most interesting central node.
5
6. NoSQL-based Approaches
These approaches can handle unstructured,
unpredictable or messy data.
6
Wide-column
stores
BigTable
model of
Google
Cassandra
Document
Stores
Semi-structured
Data
MongoDB
Key-value
stores
Berkeley DB
Graph
Databases
Graph-based
oriented data
9. Abstract Data Type Graph
G=(V, E,Σ,L) is a graph:
V is a finite set of nodes or vertices,
e.g. V={Term, forOffice, Organization,…}
E is a set of edges representing binary
relationship between elements in V,
e.g. E={(forOffice,Term)
(forOffice,Organization),(Office,Organization)…}
Σis a set of labels,
e.g., Σ={domain, range, sc, type, …}
L is a function: V x V Σ,
9
e.g., L={((forOffice,Term),domain), ((forOffice,Organization),range)… }
10. Abstract Data Type Multi-Graph
G=(V, E,Σ,L) is a multi-graph:
V is a finite set of nodes or vertices,
e.g. V={Term, forOffice, Organization,…}
E is a set of edges representing binary
relationship between elements in V,
e.g. E={(forOffice,Term)
(forOffice,Organization),(Office,Organization)…}
Σis a set of labels,
e.g., Σ={domain, range, sc, type, …}
L is a function: V x V PowerSet(Σ),
10
e.g., L={((forOffice,Term),{domain}), ((forOffice,Organization),{range}),
((_id0,AZ),{forOffice, forOrganization})… }
11. Basic Operations
Given a graph G, the following are operations over G:
AddNode(G,x): adds node x to the graph G.
DeleteNode(G,x): deletes the node x from graph G.
Adjacent(G,x,y):tests if there is an edge from x to y.
Neighbors(G,x):nodes y s.t. there is a node from x to y.
AdjacentEdges(G,x,y):set of labels of edges from x to y.
Add(G,x,y,l):adds an edge between x and y with label l.
Delete(G,x,y,l):deletes an edge between x and y with label l.
Reach(G,x,y):tests if there a path from x to y.
Path(G,x,y): a (shortest) path from x to y.
2-hop(G,x):set of nodes y s.t. there is a path of length 2 from x to y, or from
y to x.
n-hop(G,x):set of nodes y s.t. there is a path of length n from x to y, or from
y to x.
11
12. Implementation of Graphs
12
Adjacency
List
For each node a list
of neighbors.
If the graph is
directed,
adjacency list of i
contains only the
outgoing nodes of
i.
Cheaper for
obtaining the
neighbors of a
node.
Not suitable for
checking if there
is an edge
between two
nodes.
Incidence
List
Vertices and edges
are stored as records
or objects.
Each vertex stores
incident edges.
Each edge stores
incident nodes.
Incidence
Matrix
Bidimensional
representation of
graph.
Rows represent
Vertices.
Columns
represent edges
An entry of 1
represents that
the Source
Vertex is incident
to the Edge.
Adjacency
Matrix
Bidimensional
representation of
graph.
Rows represent
Source Vertices.
Columns represent
Destination Vertices.
Each entry with 1
represents that
there is an edge
from the source
node to the
destination node.
Compressed
Adjacency
Matrix
Differential
encoding
between two
consecutive
nodes
[Sakr and Pardede2012]
14. Implementation of Graphs
14
Adjacency
List
For each node a list
of neighbors.
If the graph is
directed, adjacency
list of i contains
only the outgoing
nodes of i.
Cheaper for
obtaining the
neighbors of a
node.
Not suitable for
checking if there
is an edge
between two
nodes.
Incidence
List
Vertices and edges
are stored as records
of objects.
Each vertex stores
incident edges.
Each edge stores
incident nodes.
Adjacency
Matrix
Bidimensional
representation of
graph.
Rows represent
Source Vertices.
Columns represent
Destination Vertices.
Each entry with 1
represents that
there is an edge
from the source
node to the
destination node.
Incidence
Matrix
Bi-dimensional
representation
of graph.
Rows
represent
Vertices.
Columns
represent
edges
An entry of 1 represents
that the Source Vertex is
incident to the Edge.
Compressed
Adjacency
Matrix
Differential
encoding
between two
consecutive
nodes
[Sakr and Pardede2012]
16. Implementation of Graphs
16
Adjacency
List
For each node a list
of neighbors.
If the graph is
directed, adjacency
list of i contains
only the outgoing
nodes of i.
Cheaper for
obtaining the
neighbors of a
node.
Not suitable for
checking if there
is an edge
between two
nodes.
Incidence
List
Vertices and edges
are stored as records
of objects.
Each vertex stores
incident edges.
Each edge stores
incident nodes.
Adjacency
Matrix
Bidimensional
graph
representation.
Rows represent
source vertices.
Columns represent
destination vertices.
Each non-null
entry represents
that there is an
edge from the
source node to the
destination node.
Incidence
Matrix
Bi-dimensional
representation
of graph.
Rows
represent
Vertices.
Columns
represent
edges
An entry of 1 represents
that the Source Vertex is
incident to the Edge.
Compressed
Adjacency
Matrix
Differential
encoding
between two
consecutive
nodes
[Sakr and Pardede2012]
18. Implementation of Graphs
18
Adjacency
List
For each node a list
of neighbors.
If the graph is
directed, adjacency
list of i contains
only the outgoing
nodes of i.
Cheaper for
obtaining the
neighbors of a
node.
Not suitable for
checking if there
is an edge
between two
nodes.
Incidence
List
Vertices and edges
are stored as records
of objects.
Each vertex stores
incident edges.
Each edge stores
incident nodes.
Adjacency
Matrix
Bidimensional
graph
representation.
Rows represent
source vertices.
Columns represent
destination vertices.
Each non-null
entry represents
that there is an
edge from the
source node to the
destination node.
Incidence
Matrix
Bidimensional
graph
representation.
Rows
represent
vertices.
Columns
represent
edges
A non-null entry represents
that the source vertex is
incident to the edge.
Compressed
Adjacency
Matrix
Differential
encoding
between two
consecutive
nodes
[Sakr and Pardede2012]
20. Implementation of Graphs
20
Adjacency
List
For each node a list
of neighbors.
If the graph is
directed, adjacency
list of i contains
only the outgoing
nodes of i.
Cheaper for
obtaining the
neighbors of a
node.
Not suitable for
checking if there
is an edge
between two
nodes.
Incidence
List
Vertices and edges
are stored as records
of objects.
Each vertex stores
incident edges.
Each edge stores
incident nodes.
Adjacency
Matrix
Bidimensional
graph
representation.
Rows represent
source vertices.
Columns represent
destination vertices.
Each non-null
entry represents
that there is an
edge from the
source node to the
destination node.
Incidence
Matrix
Bidimensional
graph
representation.
Rows
represent
Vertices.
Columns
represent
edges
A non-null entry represents
that the source vertex is
incident to the Edge.
Compressed
Adjacency
Matrix
Differential
encoding
between two
consecutive
nodes
[Sakr and Pardede2012]
21. Compressed Adjacency List
Gap Encoding
21
Node Neighbors
1 20,23,24,25,27
2 30,32,33,34
3 40,42,45,46,47,48
…..
Node Successors
1 38,2,0,0,1
2 56,1,0,0
3 74,1,2,0,0,0
…..
v(x) =
2x if x ³ 0
2 x -1 if x < 0
ì
í
ï
îï
The ordered adjacent list:
A(x)=(a1,a2,a3,…,an)
is encoded as
(v(a1-x),a2-a1-1,a3-a2-1,…, an-an-1 -1)
Where:
Properties:
Storage: O(|V|+|E|+|L|)
Adjacent(G,x,y): O(|E|)
Neighbors(G,x): O(|E|)
AdjacentEdges(G,x,y): O(|E|)
Add(G,x,y,l): O(|E|)
Delete(G,x,y,l): O(|E|)
Original
AdjacencyList
Compressed
AdjacencyList
22. Traversal Search
Breadth First
Search
Expands shallowest
unexpanded nodes
first.
Search is
complete.
Depth First
Search
Expands deepest
unexpanded nodes
first.
Search may be not
complete: may fail in
loops.
22
26. The Graph Data Management
Paradigm
Graph database models:
Data models for schema and instances are
graph models or generalizations of them.
Data manipulation is expressed as graph-
based operations.
26
27. Survey of Graph Database Models
27
A graph data model
represents data and
schema:
Graphs or
More general, the
notion of graphs:
Hypergraphs,
Digraphs.
[Renzo and Gutiérrez 2008]
28. Survey of Graph Database Models
28
[Renzo and Gutiérrez 2008]
Types of relationship supported by graph data models:
• Properties,
• Mono- or multi-
valued.
• Groups of real-
world objects.
• Structures to
represent
neighborhoods
of an entity.
Attributes Entities
Neighborhood
relations
• Part-of,
composed-by,
n-ary
associations.
• Subclasses and
superclasses,
• Relations of
instantiations.
• Recursively
specified
relations.
Standard
abstractions
Derivation and
inheritance
Nested
relations
29. Representative Approaches
29
No Index
Only works
for “toy”
graphs.
Vertices are
on the scale
of 1K.
Node/Edge
Index
Create index
in different
nodes/edges
Hexastore,
RDF3x,
BitMat,Neoj4,
Frequent
Subgraph Index
Indices on
frequently
queried
subgraphs
Reachability and
Neighborhood
Indices
2-hop
labeling
schemas for
index
structures.
Every two
nodes within
distance L.
33. Existing General Graph Database
Engines
33
DEX
Java library for
management of
persistent and
temporary
graphs.
Implementation
relies on bitmaps
and secondary
structures (B+-
tree)
HyperGraphDB
Implements the
hyper graph data
model.
Neo4j
Network oriented
model where
relations are first-
class objects.
Native disk-based
storage manager
for graphs.
Framework for
graph traversal.
Most graph databases implement an API instead of a query language.
34. DEX (1)
Labeled attribute multi-graph.
All node and edge objects belong to a type.
Edges have a direction:
Tail: corresponds to the source of the edge.
Head: corresponds to the destination of the edge.
Node and edge objects can have one or
more attributes.
There are no restrictions on the
number of edges between two nodes.
Loops are allowed.
Attributes are single valued.
34
[Martínez-Basan et al. 2007]
http://www.sparsity-technologies.com/dex_tutorials
35. DEX (2)
Attributes can have different indexes:
Basic attributes.
Indexed attributes.
Unique attributes.
Edges can index neighborhoods.
Neighborhood index.
The persistent database of DEX is a single file.
DEX can manage very large graphs.
35
[Martínez-Basan et al. 2007]
http://www.sparsity-technologies.com/dex_tutorials
37. DEX Internal Representation (1)
37
[Martínez-Basan et al. 2007]
DEX Approach: Map + Bitmaps Links
Link: bidirectional association between values and OIDs
Given a value a set of OIDs
Given an OID the value
http://www.sparsity-technologies.com/dex_tutorials
38. DEX Internal Representation (2)
38
[Martínez-Basan et al. 2007]
A DEX Graph is a combination of
Bitmaps:
Bitmap for each node or edge
type.
One link for each attribute.
Two links for each type: Out-
going and in-going edges.
Maps are B+trees
A compressed UTF-8 storage
for UNICODE string.
http://www.sparsity-technologies.com/dex_tutorials
39. DEX API
39
Operations of the Abstract Data Type Graph:
Graph construction.
Definition of node and edge types.
Creation of node and edge objects.
Definition and use of attributes.
Query nodes and edges.
Management of edges and nodes.
http://www.sparsity-technologies.com/dex_tutorials
40. DEX API (2)
Graph construction
40
DexConfigcfg = new DexConfig();
cfg.setCacheMaxSize(2048); // 2 GB
cfg.setLogFile(“PUBLICATIONSDex.log");
Dexdex = new Dex(cfg);
Database db = dex.create(”Publications.dex",
”PUBLICATIONSDex");
Session sess = db.newSession();
Graph graph = sess.getGraph();
……
sess.close();
db.close();
dex.close();
Operations on the graph
database will be performed
on the object graph
41. Wrote Cites
Wrote
Paper1 Paper2
Paper3
Peter Smith
41
Conference
ESWC
Paper4
Conference
Cites
Paper5
ISWC
Conference
Conference
Cites
Conference
Juan Perez
Wrote
John Smith
Wrote
Wrote
Wrote
Cites
Wrote
Example
42. DEX API (3)
Adding nodes to the graph
42
intauthorTypeId = graph.newNodeType(‛AUTHOR");
intpaperTypeId = graph.newNodeType(‛PAPER");
intconferenceTypeId = graph.newNodeType(‛CONFERENCE");
long author1 = graph.newNode(authorTypeId);
long author2 = graph.newNode(authorTypeId);
long author3 = graph.newNode(authorTypeId);
long paper1 = graph.newNode(paperTypeId);
long paper2 = graph.newNode(paperTypeId);
long paper3 = graph.newNode(paperTypeId);
long paper4 = graph.newNode(paperTypeId);
long paper5 = graph.newNode(paperTypeId);
long conference1 = graph.newNode(conferenceTypeId);
long conference2 = graph.newNode(conferenceTypeId);
43. DEX API (4)
Adding edges to the graph
43
intwroteTypeId = graph.newEdgeType(‛WROTE", true,true);
intisWrittenTypeId = graph.newEdgeType(‛Is-Written", true,true);
intciteTypeId = graph.newEdgeType(‛CITE", true,true);
intisCitedByTypeId = graph.newEdgeType(‛IsCitedBy", true,true);
intconferenceTypeId = graph.newEdgeType(‛IsConference", true,true);
long wrote1 = graph.newEdge(wroteTypeId, author1, paper1);
long wrote2 = graph.newEdge(wroteTypeId, author1, paper3);
long wrote3 = graph.newEdge(wroteTypeId, author1, paper5);
long wrote4 = graph.newEdge(wroteTypeId, author2, paper3);
long wrote5 = graph.newEdge(wroteTypeId, author2, paper4);
long wrote4 = graph.newEdge(wroteTypeId, author3, paper2);
long wrote5 = graph.newEdge(wroteTypeId, author3, paper4);
44. DEX API (5)
Indexing
Basic attributes:There is no index associated with the attribute.
Indexed attributes:There is an index automatically maintained by the
system associated with the attribute.
Unique attributes:The same as for indexed attributes but with an added
integrity restriction: two different objects cannot have the same value, with
the exception of the null value.
DEX operations:Accessing the graph through an attribute will
automatically use the defined index, significantly improving the performance
of the operation.
A specific indexcan also be defined to improve certain navigational
operations, for example, Neighborhood.
There is theAttributeKindenum class which includes basic, indexed and
unique attributes.
44
45. DEX API (6)
Adding attributes to nodes of the graph
45
nameAttrId = graph.newAttribute(authorTypeId, "Name",
DataType.String, AttributeKind.Indexed);
conferenceNameAttrId = graph.newAttribute(conferenceTypeId,
‛ConferenceName", DataType.String, AttributeKind.Indexed);
Value v = new Value();
graph.setAttribute(author1, nameAttrId, v.setString(Peter Smith"));
graph.setAttribute(author2, nameAttrId, v.setString(”Juan Perez"));
graph.setAttribute(author2, nameAttrId, v.setString(”John Smith"));
graph.setAttribute(conference1, conferenceNameAttrId,
v.setString(“ESWC”)));
graph.setAttribute(conference2, conferenceNameAttrId,
v.setString(“ISWC"));
46. DEX API (7)
Searching in the graph
46
Value v = new Value();
// retrieve all ’AUTHOR’ node objects
nodeObjects authorObjs1 = graph.select(authorTypeId);
...// retrieve Peter Smith from the graph, which is a ”AUTHOR"
nodeObjects authorObjs2 = graph.select(nameAttrId, Condition.Equal,
v.setString(‛Peter Smith"));
...// retrieve all ’author' node objects having ”Smith" in the name.
It would retrieve
// Peter Smith and John Smith
nodeObjects authorObjs3 = graph.select(nameAttrId, Condition.Like,
v.setString(‛Smith"));
...
authorObjs1.close();
authorObjs2.close();
authorObjs3.close();
47. DEX API (8)
Searching in the graph
47
Graph graph = sess.getGraph();
Value v = new Value();
...
// retrieve all ’author' node objects having a value for the 'name'
attribute
// satisfying the ’^J[^]*n$'
regular expressionObjectsauthorObjs =
graph.select(nameAttrId, Condition.RegExp,
v.setString(‛^J[^]*n$"));
...
authorObjs.close();
48. DEX API (9)
Searching in the graph
48
Value v = new Value();
Graph graph = sess.getGraph();
sess.begin();
intauthorTypeId = graph.findType(‛AUTHOR");
intnameAttrId = graph.findAttribute(authorTypeId, "NAME");
v.setString(‛Peter Smith");
Long peterSmith = graph.findObject(nameAttrId, v);
sess.commit();
49. DEX API (10)
Traversing the graph
Navigational directioncan be restricted through the
edges
The EdgesDirectionenumclass:
Outgoing: Only out-going edges, starting from the source node,
will be valid for the navigation.
Ingoing: Only in-going edges will be valid for the navigation.
Any: Both out-going and in-going edges will be valid for the
navigation.
49
50. Graph graph = sess.getGraph();
...
Objects authorObjs1 = graph.select(authorTypeId);
...// retrieve Peter Smith from the graph, which is a ”AUTHOR"
nodeObjects node1 = graph.select(nameAttrId, Condition.Equal,
v.setString(‛Peter Smith"));
nodeObjects node2 = graph.select(nameAttrId, Condition.Like,
v.setString(‛Paper"));
IntwroteTypeId = graph.findType(‛WROTE");
IntciteTypeId = graph.findType(‛CITE");
...// retrieve all in-comings WROTE
edgesObjects edges = graph.explode(node1, wroteTypeId,
EdgesDirection.Ingoing);
...// retrieve all nodes through CITE
edgesObjects cites = graph.neighbors(node2, citeTypeId,
EdgesDirection.Any);
...
edges.close();
cites.close();
DEX API (11)
Traversing the graph
50
Explode-basedmethods visit
the edges of a given node
Neighbor-basedmethods visit
the neighbor nodes of a given
node identifier
52. DEX API Summary
52
Operation Description
CreateGraph Creates a empty graph
DropGraph Removes a graph
RegisterType Registers a new node or edge type
GetTypes Returns the list of all node or edge types
FindType Returns the id of a node or edge type
NewNode Creates a new empty node of a given node type
AddNode Copies a node
AddEdge Adds a new edge of some edge type
DropObject Removes a node or an edge
Scan Returns all nodes or edges of a given type
Select Returns the nodes or edges of a given type which have an attribute that
evaluates true to a comparison operator over a value
Related Returns all neighbors of a node following edges members of an edge type
Neighbors Returns all neighbors
53. HyperGraphDB(1)
Represents graph and hypergraphstructures, a mathematical
generalization of a graph, where several edges correspond to
a hyperedge.
Hyperedges connect an arbitrary set of nodes (n-ary
relations) and can point to other hyperedges(higher-order
relations).
This representation model saves space.
53
[Iordanov 2010]
Researcher:Bob Researcher:Mary
Area:Database Area:AI Area:SW
Researcher:Bob Researcher:Mary
Area:Database Area:AI Area:SW
http://www.kobrix.com/index.jsp
54. HyperGraphDB(2)
Edgesare represented as ordered sets (directed graphs with
hyperedges).
Nodesand edges are unified into something called atom,
which is a typed tuple.
The elements of a tuple are known as the Target Set.
The size of a Target Set is the arity of a tuple.
If a tuple x has 0 arity, it’s a node, otherwise it’s a link (or edge).
The set of atoms pointing to a tuple is called the incidence
set of the tuple.
Every atom has a Handle (unique identifier).
Each entity has a type and value.
A type is an atom conforming to a special interface, values are
data managed and stored by a type atom.
54
[Iordanov 2010]
http://www.kobrix.com/index.jsp
55. HyperGraphDB Storage
Data is stored in the form of key-value pairs.
Berkeley DB is used for storage of key-value structures.
Provides three access methods for data:
Hashes:Linear hash.
B-Trees:Data stored in leaves of a balanced tree.
Recno:Assigns a logical record identifier to each pair (indexed for direct access).
Based on a key-value layer with duplicate support.
Handles are custom generated UUIDs, ints, longs, etc.
Several key-value stores: some core, some are user defined indices.
55http://www.kobrix.com/index.jsp
56. HyperGraphDB Operations And
Indexing
56
Predefined Indexers Description
ByPartIndexer Indexes atoms with compound
types along a given
dimension.
ByTargetIndexer Indexes atoms by a specific
target (a position in the target
tuple).
CompositeIndexer Combines any two indexers
into a single one. This allows
to essentially precompute and
maintain an arbitrary join.
LinkIndexer Indexes atoms by their full
target tuple.
TargetToTargetIndexer Given a link, index one of its
target by another.
Essential graph queries:
Node/Edge adjacency.
Pattern matching.
Graph traversal.
Indexing by:
Properties.
Targets.
User defined.
Predefined indexers:
ByPartIndexer.
ByTargetIndexer.
CompositeIndexer.
LinkIndexer.
TargetToTargetIndexer.
http://www.kobrix.com/index.jsp
57. Wrote Cites
Wrote
Paper1 Paper2
Paper3
Peter Smith
57
Conference
ESWC
Paper4
Conference
Cites
Paper5
ISWC
Conference
Conference
Cites
Conference
Juan Perez
Wrote
John Smith
Wrote
Wrote
Wrote
Cites
Wrote
Example
58. HyperGraphDB API (1)
Adding nodes to the graph
HyperGraph graph = new HyperGraph(databaseLocation);
author author1 = new AUTHOR(‛Peter Smith‛);
author author2 = new AUTHOR(‛Juan Perez‛);
author author3 = new AUTHOR(‛John Smith‛);
PAPER paper1 = new PAPER(‚Paper1‛);
PAPER paper2 = new PAPER(‚Paper2‛);
PAPER paper3 = new PAPER(‚Paper3‛);
PAPER paper4 = new PAPER(‚Paper4‛);
PAPER paper5 = new PAPER(‚Paper5‛);
CONFERENCE conference1 = new CONFERENCE(‚ESWC‛);
CONFERENCE conference2 = new CONFERENCE(‚ISWC‛);
HGHandle authorHandle1 = graph.add(paper1);
HGHandle authorHandle2 = graph.add(paper2);
…
HGHandle conferenceHandle2 = graph.add(conference2);
…
graph.close();
58
59. HyperGraphDB API (2)
Adding edges to the graph
59
HyperGraph graph = new HyperGraph(databaseLocation);
author author1 = new AUTHOR(‛Peter Smith‛);
author author2 = new AUTHOR(‛Juan Perez‛);
author author3 = new AUTHOR(‛John Smith‛);
PAPER paper1 = new PAPER(‚Paper1‛);
HGHandle authorHandle1 = graph.add(paper1);
HGHandleyearHandle = graph.add(2013);
HGValueLink link1 = new HGValueLink(‛publication year‛,
authorHandle1, yearHandle);
HGHandle linkHandle1 = graph.add(link1);
graph.close();
60. HyperGraphDB API (3)
Searching node properties
60
HGQueryCondition condition= new And(
new AtomTypeCondition(AUTHOR.class),
new AtomPartCondition(newString[]{‛name"}, ‛Peter Smith",
ComparisonOperator.EQ));
HGSearchResult<HGHandle>rs = graph.find(condition);
while (rs.hasNext()) {
HGHandle current = rs.next();
AUTHOR author = graph.get(current);
System.out.println(AUTHOR.getBithDate());
}
61. HyperGraphDB API (4)
61
List<author> authors = hg.getAll(hg.and(hg.type(AUTHOR.class),
hg.eq("author", ‛Peter Smith")));
for (author p : authors)
System.out.println(author.getBirthDate());
}
Searching node properties
62. HyperGraphDB API (5)
Searching node properties
62
DefaultALGeneratoralgen = new DefaultALGenerator(graph,
hg.type(CitedBy.class),
hg.and(hg.type(PAPER.class),
hg.eq(‛track‛,‛Semantic Data Management‛),
true, false, false);
HGTraversal traversal = new HGBreadthFirstTraversal(
startingPaper, algen);
Paper currentArticle = startingPaper;
while (traversal.hasNext()) {
Pair<HGHandle, HGHandle> next = traversal.next();
PAPER nextPaper = graph.get(next.getSecond());
System.out.println(‛Paper " + current + " quotes " + nextPaper);
currentPaper = nextPaper;
}
HGBreadthFirstTraversalmeth
od to traverse a graph returns a
startAtom and
adjListGenerator
63. HyperGraphDB API (7)
Searching node properties
63
author author1 = newAUTHOR(‛Peter Smith‛);
HGDepthFirstTraversal traversal = new HGDepthFirstTraversal(
author1,
new SimpleALGenerator(graph));
while (traversal.hasNext()) {
Pair<HGHandle, HGHandle> current = traversal.next();
HGLink l = (HGLink)graph.get(current.getFirst());
Object atom = graph.get(current.getSecond());
System.out.println("Visiting atom " + atom +
‚ pointed to by " + l);
}
SimpleAlGeneratormethod to
produce all atoms link to a given
atom
64. HyperGraphDB API Summary
64
Class/Operation Description
HGEnvironment Creates a hypergraph from a file
Add Adds a new object to the hypergraph
Update Updates an object in the hypergraph
Remove Removes an object from the hypergraph
HGHandle A reference to a hypergraph atom
HGPersistentHandle a HGHandle that survives system downtime
HGValueLink Adds a newhyperedge
get Returns the object referred by a given HGHandle
HGQueryCondition Returns the nodes or edges of a given type which
have an attribute that evaluates true to a
comparison operator over a value
HGQuery.hg.getAll Returns all the objects that meet a condition
65. Neo4J
Labeled attribute multigraph.
Nodes and edges can have properties.
There are norestrictions on the number of edges
between two nodes.
Loops are allowed.
Attributes are single valued.
Different types of indexes: Nodes & relationships.
Different types of traversal strategies.
API for Java, Python.
65
[Robinson et al. 2013]
http://neo4j.org/
68. Wrote Cites
Wrote
Paper1 Paper2
Paper3
Peter Smith
68
Conference
ESWC
Paper4
Conference
Cites
Paper5
ISWC
Conference
Conference
Cites
Conference
Juan Perez
Wrote
John Smith
Wrote
Wrote
Wrote
Cites
Wrote
Example
74. Neo4J: Cypher Query Language (1)
START:
Specifies one or more starting points in the graph.
Starting points can be defined as index lookups or just by
an element ID.
MATCH:
Pattern matching to match based on the starting point(s)
WHERE:
Filtering criteria.
RETURN:
What is projected out from the evaluation of the query.
74
75. Neo4J: Cypher Query Language (2)
Query:IsPeter Smith connected to John Smith.
75
START author=node:authors(
'firstname:‛Peter" AND lastname:‛Smith"')
MATCH (author)-[R*]->(m)
RETURNcount(R)
76. Neo4J: Cypher Query Language (2)
Query: Papers written by Peter Smith.
76
START author=node:authors(
'firstname:‛Peter" AND lastname:‛Smith"')
MATCH (author)-[:WROTE]->(papers)
RETURN papers
77. Neo4J: Cypher Query Language (3)
Query: Papers cited by a paper written by Peter
Smith.
77
START author=node:authors(
'firstname:‛Peter" AND lastname:‛Smith"')
MATCH (author)-[:WROTE]->()-[:CITES](papers)
RETURN papers
78. Neo4J: Cypher Query Language (3)
Query: Papers cited by a paper written by Peter
Smith that have at most 20 cites.
78
START author=node:authors(
'firstname:‛Peter" AND lastname:‛Smith"')
MATCH (author)-[:WROTE]->()-[:CITES]->()-[:CITES]->(papers)
WITH COUNT(papers) as cites
WHERE cites < 21
RETURN papers
79. Neo4J: Cypher Query Language (4)
Query: Papers cited by a paper written by Peter
Smith or cited by papers cited by a paper written
by Peter Smith.
79
START author=node:authors(
'firstname:‛Peter" AND lastname:‛Smith"')
MATCH (author)-[:WROTE]->()-[:CITES*1..2](papers)
RETURN papers
80. Neo4J: Cypher Query Language (5)
Query: Number of papers cited by a paper written
by Peter Smith or cited by papers cited by a paper
written by Peter Smith.
80
START author=node:authors(
'firstname:‛Peter" AND lastname:‛Smith"')
MATCH (author)-[:WROTE]->()-[:CITES*1..2](papers)
RETURN count(papers)
81. Neo4J: Cypher Query Language (6)
Query: Number of papers cited by a paper written
by Peter Smith or cited by papers cited by a paper
written by Peter Smith, and have been published in
ESWC.
81
START author=node:authors(
'firstname:‛Peter" AND lastname:‛Smith"')
MATCH (author)-[:WROTE]->()-[:CITES*1..2](papers)
WHERE papers.conference! =‘ESWC’
RETURN count(papers)
Exclamation mark in papers.conference ensures that papers
without the conference attribute are not included in the output.
82. Neo4J: Cypher Query Language (7)
Query: Number of papers cited by a paper written
by Peter Smith or cited by papers cited by a paper
written by Peter Smith, have been published in
ESWC and have at most 4 co-authors .
82
START author=node:authors(
'firstname:‛Peter" AND lastname:‛Smith"')
MATCH (author)-[:WROTE]->()-[:CITES*1..2](papers)-
[:Was-Written]->authorFinal
WHERE papers.conference! =‘ESWC’
WITH author, count(authorFinal) as authorFinalCount
WHERE authorFinalCount< 4
RETURN author, authorFinalCount
Exclamation mark in papers.conference ensures that papers
without the conference attribute are not included in the output.
83. Neo4J API-Cypher
Query: Peter Smith Neighborhood
83
static Neo4j TestGraph;
String Query = "start n=node:node_auto_index(idNode='PeterSmith')" +
"match n-[r]->m return m.idNode";
ExecutionResultresult = TestGraph.query(Query);
booleanres = false;
for ( Map<String, Object> row : result ){
for ( Map.Entry<String, Object> column : row.entrySet() ){
print(column.getValue().toString());
res = true;
}
}
if (!res) print(" No result");
84. Operation Description
GraphDatabaseFactory Creates a empty graph
registerShutdownHook Removes a graph
createNode Creates a new empty node of a given node type
AddNode Copies a node
createRelationshipTo Adds a new edge of some edge type
forNodes Index for a Node
IndexManager Index Manager
forRelationships Index for a Relationship
add Add an object to an index
TraversalDescription Different types of strategies to traverse a graph
Traversal
Branch Selector
preorderDepthFirst, postorderDepthFirst,
preorderBreadthFirst, postorderBreadthFirst
Neo4J API Summary
84
89. In the beginnings …
89
• Traditionally:
– Use Relational DBMS’ to store data
– Early Approach: onelargeSPOtable
(physical)
• Pros: simple schema, fast updates (when no index
present), fast single triple pattern queries
• Cons: not suitable for multi-join queries, results in
multiple self-joins
90. In the beginnings …
90
• Traditionally:
– Use Relational DBMS’ to store data
– Further Approaches: propertytables
(multiple flavors)
• Pros: fast retrieval of subject / object pairs for
given predicates
• Cons: not easy to solve queries where predicate is
unbound & harder to maintain schema (a table for
each known predicate)
91. RDF-tailored management approaches
91
• Same conceptual approach: onelargeSPOtable (but
with RDF-specific physical organization):
– Triples are indexed in all 3!=6 possible permutations:
SPO, PSO, OSP, OPS, PSO, POS
– Logical extension of the vertical partitioning proposed
in [Abadi et al 2008] (the PSO index = generalization
of the SO table)
– Each index (i.e. SPO) stores partial triple pattern
cardinalities
– Resulting in 6 tree-based indexes, each 3 levels deep
• First proposed in Hexastore
[Weiss et al. 2008]
92. Hexastore / Conceptual Model
92
• 2 main operations: (input = tp: partially bounded triple pattern, i.e.
<bound_subject><some predicate> ?o)
• cardinality = getCardinality(tp)
• term_set = getSet(tp, set)
• Resulting term_set is sorted
93. Hexastore / Conceptual Model
93
• Pros:
• The optimal index permutation & level is used (i.e. <bound_s> ?p ?o : SPO level
1)
• Ability to make use of fast merge-joins
• Cons:
• Higher upkeep cost: a theoretical upper bound of 5x space increase over original
(encoded) data
94. 94
Wrote Cite
s
Wrote
Paper1 Paper2
Paper3
Peter Smith
Conference
ESWC
Paper4
Conference
Cite
s
Paper5
ISWC
Conference
Conference
Cite
s
Conference
Juan Perez
Wrote
John Smith
Wrote
Wrote
Wrote
Cite
s
Wrote
An Example (Publications):
logical representation
101. Hexastore / Physical implementation.
DBMS (managed) storage
101
• Pros:
– Flexible design: each index can
be configured for a different
physical data-structure:
• i.e. PSO=B+Tree, OSP=LSM Tree
– Separation of concerns (the
underlying DBMS manages and
optimizes for access patterns,
cache, etc…)
• Cons:
– Less control over fine-grained
data storage (up to the level
permitted by the DBMS API)
– Usually less potential for native
optimization
2
2
4
3
3
4
4
9
8
1
5
2
4
3
S
1
1
1
1
11
12
1
1
13 1
14 1
1 11 9 -
1 -1812
1 13 5 -
1 14 2 -
2
2
2
1
13
1
12
11
1
2 11 8 -
2 12 20 -
2 13 15 -
3
3
3
3
11
12
1
2
13 1
14 2
3 11 9 -
12 -3 18
3 13 15 -
3 -14 1
4
4
4
4 11
12
1
2
13 1
14 1
4 11 9 -
4 -12 20
4 13 15 -
4 14 2 -
5
5
5 12 1
111
13 1
5 11 8 -
5 12 18 -
5 13 15 -
8
8 17 1
116
8 16 10 -
8 17 7 -
9
9 17
1
1
16
9 16 10 -
9 17 6 -
SP SPO
-3 1912
43 -14
12 19 -4
102. RDF3x (1)
RISC style operators.
Implements the traditionalOptimize-then-Execute
paradigm, to identify left-linear plans of star-shaped sub-
queries of basic triple patterns.
Makes use of secondary memory structures to locally
store RDF data and reduce the overhead of I/O
operations.
Registers aggregated count or number of occurrences,
and is able to detect if a query will return an empty
answer.
102
[Neumann and Weikum 2008]
103. RDF3x (2)
Triple table, each row represents a triple.
Mapping dictionary, replacing all the literal
strings by an id; two dictionary indices.
Compressed clustered B+-tree to index all triples:
Six permutations of subject, predicate and object;
each permutation in a different index.
Triples are sorted lexicographically in each B+-tree.
103
[Neumann and Weikum 2008]
108. Wrote Cites
Wrote
Paper1 Paper2
Paper3
Peter Smith
108
Conference
ESWC
Paper4
Conference
Cites
Paper5
ISWC
Conference
Conference
Cites
Conference
Juan Perez
Wrote
John Smith
Wrote
Wrote
Wrote
Cites
Wrote
Example
111. BitMat(1)
3D bit-cube representation where each dimension
represents subjects, predicates and objects.
Each cell of the matrix is 0 or 1, and represents an RDF
triple that can be formed by the combination of S, P, O.
3D bit-cube is slicedalong a dimension to obtain 2D
matrices.
A gap compression schema on each bit-row is
implemented:
Basic operations over BitMat are defined to join basic triple
patterns.
111
[Atre et al. 2010]
114. Wrote Cites
Wrote
Paper1 Paper2
Paper3
Peter Smith
114
Conference
ESWC
Paper4
Conference
Cites
Paper5
ISWC
Conference
Conference
Cites
Conference
Juan Perez
Wrote
John Smith
Wrote
Wrote
Wrote
Cites
Wrote
Example
119. BHyper(1)
A hypergraph-basedstructure is used to represent RDF documents.
Mapping dictionary, replacing all the literal strings by an id, two
dictionary indices, indexed with a Hash Index.
Encodings of RDF resources are implemented as nodes, and set of
triple patterns corresponds to hyperedges.
A role function is used to denote the role played by a resource in a
hyperedge. Roles are represented as labels in the hyperedges.
Indicesare used to index the hyperdges and the different roles
played by the resources in the hyperedges:
B+-tree are used to index hyperedges.
BHyper is built on top of Tokyo Cabinet1.4.451.
119
1 http://sourceforge.net/projects/tokyocabinet/
[Vidal and Martínez 2011]
123. Experimental Set-Up
The experiments presented in this section were executed on a
machine with the following characteristics:
Model: Sun Fire x4100
Quantity of processors: 2
Processor specification: Dual-Core AMD Opteron™ Processor 2218 64
bits
RAM memory: 16GB
Disk capacity: 2 disks, 135GB e/o
Network interface: 2 + ALOM
Operating System:CentOS 5.5. 64 bits
RDF3x version: 0.3.7
DEX version: 4.8 Very Large Databases
HyperGraphDB version: 1.2
Neo4J version: 1.8.2
Timeout: 3,600 secs. (Value returned in the portal -2)
123
124. 124
Graph Name #Nodes #Edges Description
DSJC1000.1 1,000 49,629 Graph Density:0.1
[Johnson91]
DSJC1000.5 1,000 249,826 Graph Density:0.5
[Johnson91]
DSJC1000.9 1,000 449,449 Graph Density:0.9
[Johnson91]
SSCA2-17 131,072 3,907,909 GTgraph 1)
[Johnson91] Johnson, D., Aragon, C., McGeoch, L., and Schevon, C. Optimization by simulated
annealing: an experimental evaluation; part ii, graph coloring and number partitioning. Operations
research 39, 3 (1991), 378–406.
GTgraph: A suite of three synthetic graph generators:
1) SSCA2: generates graphs used as input instances for the DARPA. High Productivity Computing
Systems SSCA#2 graph theory benchmark
http://www.cse.psu.edu/~madduri/software/GTgraph/index.html
Benchmark Graphs
125. 125
Task Description
adjacentXP Given a node X and an edge label P, find
adjacentnodes Y.
adjacentX Given a node X,find adjacentnodes Y.
edgeBetween Given two nodes X and Y, find the labels of the
edges between X and Y.
2-hopX Given a node X, find the 2-hop Y nodes.
3-hopX Given a node X, find the 3-hop Y nodes.
4-hopX Given a node X, find the 4-hop Y nodes.
Benchmark Tasks
The benchmark tasks evaluate the engine performance while executing
the following graph operations:
126. 126
Task SPARQL Query
adjacentXP Select ?y
where {
<http://graphdatabase.ldc.usb.ve/resource/6><http://graphdatabase.ldc.usb.ve/resource/pr>
?y}
adjacentX Select ?y
where {
<http://graphdatabase.ldc.usb.ve/resource/6> ?p ?y}
edgeBetween Select ?p
where {
<http://graphdatabase.ldc.usb.ve/resource/6> ?p<http://graphdatabase.ldc.usb.ve/resource/8>
2-hopX Select ?z
where {
<http://graphdatabase.ldc.usb.ve/resource/6> ?p1 ?y1.
?y1 ?p2 ?z}
3-hopX Select ?z
where {
<http://graphdatabase.ldc.usb.ve/resource/6> ?p1 ?y1.
?y1 ?p2 ?y2.
?y2 ?p3 ?z}
4-hopX Select ?z
where {
<http://graphdatabase.ldc.usb.ve/resource/6> ?p1 ?y1.
?y1 ?p2 ?y2.
?y2 ?p3 ?y3.
?y3 ?p4 ?z}
Benchmark Tasks: RDF Engines
The benchmark tasks were implemented in the RDF engines with the
following SPARQL queries:
127. Graph Mining Tasks
These tasks evaluate the engine performance
while executing the following graph mining:
Graph summarization.
Dense subgraph.
Shortest path.
127
130. A
B
C
D
E
F
G
H
1
2 1
1
1
3
3
2
3
1
1
1
2
Given a subset of nodes E’ compute the densest subgraph containing
them, i.e., density of E’ is maximized:
density(E') =
weight(a,b)
| E'|(a,b) in E'
å
[Khuller et al. 2010]
Dense Subgraph
130
131. Shortest Path
• Given two nodes finds the shortest path
between them.
• Implemented using Iterative Depth First Search
(DFID).
• Ten node pairs are taken at random from each
graph, and then the Shortest Path algorithm is
executed on each pair of nodes
131
134. 134
Task Description
adjacentXP Given a node X and an edge label P, find
adjacentnodes Y.
adjacentX Given a node X,find adjacentnodes Y.
edgeBetween Given two nodes X and Y, find the labels of the
edges between X and Y.
2-hopX Given a node X, find the 2-hop Y nodes.
3-hopX Given a node X, find the 3-hop Y nodes.
4-hopX Given a node X, find the 4-hop Y nodes.
Benchmark Tasks
The benchmark tasks evaluate the engine performance while executing
the following graph operations:
135. 135
Task SPARQL Query
adjacentXP Select ?y
where {
<http://graphdatabase.ldc.usb.ve/resource/6><http://graphdatabase.ldc.usb.ve/resource/pr>
?y}
adjacentX Select ?y
where {
<http://graphdatabase.ldc.usb.ve/resource/6> ?p ?y}
edgeBetween Select ?p
where {
<http://graphdatabase.ldc.usb.ve/resource/6> ?p<http://graphdatabase.ldc.usb.ve/resource/8>
2-hopX Select ?z
where {
<http://graphdatabase.ldc.usb.ve/resource/6> ?p1 ?y1.
?y1 ?p2 ?z}
3-hopX Select ?z
where {
<http://graphdatabase.ldc.usb.ve/resource/6> ?p1 ?y1.
?y1 ?p2 ?y2.
?y2 ?p3 ?z}
4-hopX Select ?z
where {
<http://graphdatabase.ldc.usb.ve/resource/6> ?p1 ?y1.
?y1 ?p2 ?y2.
?y2 ?p3 ?y3.
?y3 ?p4 ?z}
Benchmark Tasks: RDF Engines
The benchmark tasks were implemented in the RDF engines with the
following SPARQL queries:
145. A
B
C
D
E
F
G
H
1
2 1
1
1
3
3
2
3
1
1
1
2
Given a subset of nodes E’ compute the densest subgraph containing
them, i.e., density of E’ is maximized:
density(E') =
weight(a,b)
| E'|(a,b) in E'
å
[Khuller et al. 2010]
Dense Subgraph
145
146. Shortest Path
• Given two nodes finds the shortest path
between them.
• Implemented using Iterative Depth First Search
(DFID).
• Ten node pairs are taken at random from each
graph, and then the Shortest Path algorithm is
executed on each pair of nodes
146
151. Graph Data Storing Features
151
Graph Database Main
Memory
Persistent
Memory
Backend
Storage
Indexing
DEX X X X
HyperGraphDB1 X X X X
Neo4j X X X
Hexastore2 X X X X
RDF3x X X
Bitmat X X
Bypher2 X X X X
2Backend Storage: Tokyo Cabinet
1Backend Storage: Berkeley DB
152. Graph Data Management Features
152
Graph Database Graph API Query
Language
DataVi
suali-
zationData
Definition
Data
Manipulation
Standard
Compliance
DEX X X X X
HyperGraphDB X X X
Neo4j X X X X X
Hexastore X X X X
RDF3x X
BitMat X X X
Bypher X X X
156. Data Storing Features:
Engines implement a wide variety of
structures to efficiently represent large graphs.
RDF engines implement special-purpose
structures to efficiently store RDF graphs.
156
Conclusions (1)
157. Data Management Features:
General-purpose graph database engines
provide APIs to manage and query data.
RDF engines support general data management
operations.
157
Conclusions (2)
158. Graph-Based Operations Support:
General-purpose graph database engines provide API
methods to implement a wide variety of graph-based
operations.
Only few offer query languages to express pattern
matching.
RDF engines are not customized for basic graph-based
operations, however,
They efficiently implement the pattern matching-
based language SPARQL.
158
Conclusions (3)
159. Extensionof existing enginesto support specific
tasks of graph traversal and mining.
Definition of benchmarksto study the performance
and quality of existing graph database engines, e.g.,
Graph traversal, link prediction, pattern discovery,
graph mining.
Empirical evaluation of the performance and quality
of existing graph database engines
159
Future Directions
160. P. Anderson, A. Thor, J. Benik, L. Raschid, and M.-E. Vidal. Pang: finding patterns in
anno- tation graphs. In SIGMOD Conference, pages 677–680, 2012.
R. Angles and C. Gutiérrez. Survey of graph database models. ACM Comput. Surv.,
40(1), 2008.
M. Atre, V. Chaoji, M. Zaki, and J. Hendler. Matrix Bit loaded: a scalable lightweight join
query processor for RDF data. In International Conference on World Wide Web,
Raleigh, USA, 2010. ACM.
B.Iordanov.Hypergraphdb:Ageneralizedgraphdatabase.InWAIMWorkshops,pages25–36,
2010.
N. Martíınez-Bazan, V. Muntés-Mulero, S. G ómez-Villamor, J. Nin, M.-A. nchez-
Martíınez, and J.-L. Larriba-Pey. Dex: high-performance exploration on large graphs
for information retrieval. In CIKM, pages 573–582, 2007.
T. Neumann and G. Weikum. RDF-3X: a RISC-style engine for RDF. Proc. VLDB, 1(1),
2008.
160
References
161. M. K. Nguyen, C. Basca, and A. Bernstein. B+hash tree: optimizing query execution
times for on-disk semantic web data structures. In Proceedings Of The 6th
International Workshop On Scalable Semantic Web Knowledge Base Systems
(SSWS2010), 2010.
I Robinson, J. Webber, and E. Eifrem. Graph Databases. O’Reilly Media, Incorporated,
2013.
S. Sakr and E. Pardede, editors. Graph Data Management: Techniques and Applications.
IGI Global, 2011.
D.ShimpiandS.Chaudhari. An overview of graph databases .In IJCA Proceedings on
International Conference on Recent Trends in Information Technology and
Computer Science 2012, 2013.
M.-E. Vidal, A. Martínez, E. Ruckhaus, T. Lampo, and J. Sierra. On the efficiency of
querying and storing rdf documents. In Graph Data Management, pages 354–385.
2011.
C. Weiss, P. Karras, and A. Bernstein. Hexastore: Sextuple indexing for semantic web
data management. Proc. of the 34th Intl Conf. on Very Large Data Bases (VLDB),
2008.
161
References
162. C Weiss, A Bernstein, On-disk storage techniques for Semantic Web data - Are B-Trees
always the optimal solution?, Proceedings of the 5th International Workshop on
Scalable Semantic Web Knowledge Base Systems, October 2009.
162
References