Skip to content

Latest commit

 

History

History
182 lines (127 loc) · 6.03 KB

README.md

File metadata and controls

182 lines (127 loc) · 6.03 KB
Raw

Spreading Activation

Based on http://dx.doi.org/10.14279/depositonce-8408 and http://dx.doi.org/10.14279/depositonce-7136
Implementation for apache/tinkerpop

Usage

  • Pulse 0 refer to the initial activated vertices (i.e. no output, edge nor input activation).
  • The direction of edges does not matter, activation will be spread both ways (in back and forth direction).
  • Invalid or non-existent activation values (i.e. infinity, NaN, 0) will not be considered for further calculation nor stored.

The API heavily uses builder pattern, including many sensible defaults. The following examples show almost all possible options available.

Spreading Activation

Build a spreading activation configuration:

int pulses = 4;

SpreadingActivation spreadingActivation = SpreadingActivation.build(pulses)
		.pulseInception(PulseInception.Default.MINIMUM_ACTIVATION(0.1d))
		.activationMode(ActivationMode.Default.LOG10)
		.attenuationMode(AttenuationMode.Default.FIXED(0.99d))
		.branchMode(BranchMode.Default.NONE)
		.sendMode(SpreadingMode.Default.FORWARD)
		.edgeWeight(EdgeWeight.Default.CONSTANT)
		.create();

Build an execution task of a spreading activation for a graph:

Graph graph = TinkerFactory.createModern();
		
Execution execution = spreadingActivation.execution(graph.traversal())
		.executor(Executors.newCachedThreadPool()) // default
		.parallelTasks(Runtime.getRuntime().availableProcessors()) // default
		.propertyKeyFactory(new Execution.DefaultPropertyKeyFactory(UUID.randomUUID().toString())) // default, see Execution.PropertyKeyFactory
		.create();

Activate initial vertices and start execution:

execution.activate(__.has("name", "marko"), 1d);
execution.run();

Interrupt a running execution:

execution.interrupt();

Get the state of an execution task (always finished==true at the end, whether interrupted or not):

boolean started = execution.isStarted();
boolean interrupted = execution.isInterrupted();
boolean finished = execution.isFinished();

Get the result of the spreading activation, including activates results, last pulse and specific activation values for a certain vertex at a certain pulse:

ExecutionResult result = execution.getResult();

int lastPulse = result.pulse();
GraphTraversal<?, Vertex> activatedVertices = result.activatedVertices(lastPulse);
double activation = result.activation(vertex, pulse);

Remove all spreading activation properties:

result.cleanup();

Hint: Those properties are necessary for the following spread graphs, cleanup last but not least.

Spread Graphs

Build an generation task to create a spread graph from a spreading activation result:

Generation generation = result.generateSpreadGraph()
		.into(TinkerGraph.open().traversal()) // default
		.startPulse(0) // default
		.endPulse(lastPulse) // default
		.propertyKeyFactory(SpreadGraph.DefaultPropertyKeyFactory.INSTANCE) // default, see SpreadGraph.PropertyKeyFactory
		.create();

Starting, interrupting and querying state of generation is equivalent to execution:

generation.run();

generation.interrupt();

boolean started = generation.isStarted();
boolean interrupted = generation.isInterrupted();
boolean finished = generation.isFinished();

Interrupt a running generation:

generation.interrupt();

Get the state of an generation task (always finished==true at the end, whether interrupted or not):

boolean started = generation.isStarted();
boolean interrupted = generation.isInterrupted();
boolean finished = generation.isFinished();

Get the vertices of the spread graph. Start vertex do not have incoming edges (sources), end vertices have no outgoing edges (sinks).

SpreadGraph spreadGraph = generation.getSpreadGraph();

GraphTraversal<Vertex,Vertex> vertices = spreadGraph.vertices(lastPulse);
GraphTraversal<Vertex,Vertex> startVertices = spreadGraph.startVertices();
GraphTraversal<Vertex,Vertex> endVertices = spreadGraph.endVertices();

Build a minimization task to create a spread graph for a single relevant vertex from the original spreading activation (i.e. remove all edges and vertices, which did not contribute to activate this vertex while spreading activation)

Vertex relevantVertex = activatedVertices.next();

RelevantMinimization minimization = RelevantMinimization.build(spreadGraph, relevantVertex.id())
		.into(TinkerGraph.open().traversal()) // default
		.startPulse(0) // default
		.endPulse(lastPulse) // default
		.propertyKeyFactory(SpreadGraph.DefaultPropertyKeyFactory.INSTANCE) // default, see SpreadGraph.PropertyKeyFactory
		.create();

Starting, interrupting and querying state of minimization is equivalent to execution:

minimization.run();

minimization.interrupt();

boolean started = minimization.isStarted();
boolean interrupted = minimization.isInterrupted();
boolean finished = minimization.isFinished();

The result of a minimization is again a spread graph.

SpreadGraph minimizedSpreadGraph = minimization.getSpreadGraph();

GraphTraversal<Vertex,Vertex> vertices = minimizedSpreadGraph.vertices(lastPulse);
GraphTraversal<Vertex,Vertex> startVertices = minimizedSpreadGraph.startVertices();
GraphTraversal<Vertex,Vertex> endVertices = minimizedSpreadGraph.endVertices();

Extension Points

  • Spreading Activation Modes: See interfaces ActivationMode, AttenuationMode, BranchMode, EdgeWeight and SendMode
  • Property Keys: See interface Execution.PropertyKeyFactory and SpreadGraph.PropertyKeyFactory
  • Spread Graph Transformation (like RevelantMinimization): Extend Transformation (and Transformation.Builder)

Current Implementation Features

  • extendable configuration via interfaces (all predefined spreading activation options implemented)
  • parallel execution within spreading activation calculation
  • spread graph for detailed analysis of spreading activation process
  • all long term processes interruptible
  • activation values as double, stored in arbitrary properties (parallel execution of multiple spreading activations on a single graph)