public static ICModel runSparsifier( SocialNetwork socNet, ObservationsReader observations, ICModel originalModel, int sparseSize, Sparsifier sparse, int numOfChunks, boolean reportPartial) { LOGGER.info( "Arcs: total in social network=" + socNet.getArcs().size() + ", with non-zero probability=" + originalModel.getProbs().cardinality() + ", target k=" + sparseSize); LOGGER.info("Begin creating sparsified model"); ICModel sparseModel = sparse.sparsify(sparseSize, numOfChunks, observations, reportPartial); LOGGER.info("Done creating sparsified model"); LOGGER.info("Begin measuring the resulting sparsified model"); // Report change in log likelihood double sparsifiedLogL = sparseModel.getLogLikelihoodIgnoringParentInformation(sparse.getAuxiliary()); LOGGER.info("SPARSIFIED log likelihood (ignoring parent information)=" + sparsifiedLogL); double originalLogL = originalModel.getLogLikelihoodIgnoringParentInformation(sparse.getAuxiliary()); LOGGER.info("ORIGINAL log likelihood (ignoring parent information)=" + originalLogL); LOGGER.info("End measuring the resulting sparsified model"); return sparseModel; }
/** * Computes and stores the partial fraction of covered propagations. * * <p>Works only if {@link #computeFractionOfPropagations} is true, otherwise does nothing. * * @param observations the observations * @param newProbs the probabilities of the current model * @param i a number indicating the number of edges so far */ protected void computeAndStorePartialFractionOfPropagations( ObservationsReader observations, SparseDoubleMatrix2D newProbs, int i) { if (computeFractionOfPropagations) { ICModel tempModel = new ICModel(originalModel.getSn(), newProbs); double sparsifiedFraction = tempModel.getTotalFraction(observations); storePartialResult(Measure.FRACTION_OF_PROPAGATIONS, i, sparsifiedFraction); } }
public void computeAuxiliary( ObservationsReader observations, CandidateSelectionPolicy candidateSelectionPolicy) { if (this.auxiliary != null) { throw new IllegalStateException("Auxiliary was already set"); } this.auxiliary = new ICEstimateAuxiliary(originalModel.getSn(), observations, candidateSelectionPolicy); }
public static void main(String[] args) throws Exception { final SimpleJSAP jsap = new SimpleJSAP( Sparsifier.class.getName(), "Estimates and sparsifies a propagation model from a set of observations.", new Parameter[] { new FlaggedOption( "social-network", JSAP.STRING_PARSER, JSAP.NO_DEFAULT, JSAP.REQUIRED, 's', "social-network", "The file containing the social network graph"), new FlaggedOption( "probabilities", JSAP.STRING_PARSER, JSAP.NO_DEFAULT, JSAP.REQUIRED, 'p', "probabilities", "The file containing the propagation probabilities"), new FlaggedOption( "candidate-selection-policy", JSAP.STRING_PARSER, CandidateSelectionPolicy.DEFAULT_CANDIDATE_SELECTION_POLICY .getClass() .getSimpleName(), JSAP.REQUIRED, 'c', "candidate-selection-policy", "The name of the candidate selection policy"), new FlaggedOption( "auxiliary-basename", JSAP.STRING_PARSER, JSAP.NO_DEFAULT, JSAP.NOT_REQUIRED, JSAP.NO_SHORTFLAG, "auxiliary-basename", "The base name for reading a pre-computed auxiliary structure"), new FlaggedOption( "input", JSAP.STRING_PARSER, JSAP.NO_DEFAULT, JSAP.REQUIRED, 'i', "input", "The file containing the observations"), new FlaggedOption( "sparsifier", JSAP.STRING_PARSER, DEFAULT_SPARSIFIER.getSimpleName(), JSAP.NOT_REQUIRED, 'f', "sparsifier", "The sparsifier to run, from this list: " + StringUtils.join(Reflection.subClasses(Sparsifier.class), ',')), new FlaggedOption( "sparse-model-size", JSAP.INTEGER_PARSER, JSAP.NO_DEFAULT, JSAP.NOT_REQUIRED, 'k', "sparse-model-size", "The size of the sparse model"), new FlaggedOption( "number-of-chunks", JSAP.INTEGER_PARSER, JSAP.NO_DEFAULT, JSAP.NOT_REQUIRED, 'r', "number-of-chunks", "The number of chunks to be sparsified in parralel"), new FlaggedOption( "output", JSAP.STRING_PARSER, JSAP.NO_DEFAULT, JSAP.NOT_REQUIRED, 'o', "output", "File to dump sparsified model to"), new FlaggedOption( "measures-file", JSAP.STRING_PARSER, JSAP.NO_DEFAULT, JSAP.NOT_REQUIRED, 'z', "measures-file", "Save measures of partial models to file"), new FlaggedOption( "debug-file", JSAP.STRING_PARSER, JSAP.NO_DEFAULT, JSAP.NOT_REQUIRED, 'd', "debug-file", "Save debug information to file"), new Switch( "with-fraction", 'n', "with-fraction", "Disable the computation of the 'fraction of covered propagations'."), new Switch( "incremental-likelihood", JSAP.NO_SHORTFLAG, "incremental-likelihood", "Performs incremental computation of likelihood, for sparsifications methods that support this option (faster, experimental)."), }); final JSAPResult jsapResult = jsap.parse(args); if (jsap.messagePrinted()) { return; } // Load social network and input String snFilename = jsapResult.getString("social-network"); SocialNetwork socNet = new SocialNetwork(Utilities.getIterator(snFilename)); String obsFilename = jsapResult.getString("input"); ObservationsReader observations = new ObservationsReader(obsFilename); // Load original model ICModel originalModel = new ICModel(socNet, Utilities.getIterator(jsapResult.getString("probabilities"))); // Load candidate selection policy String selectionPolicyName = jsapResult.getString("candidate-selection-policy"); Class<?> selectionPolicyClass = Class.forName( CandidateSelectionPolicy.class.getPackage().getName() + "." + selectionPolicyName); CandidateSelectionPolicy candidateSelectionPolicy = (CandidateSelectionPolicy) selectionPolicyClass.getConstructor(new Class[] {}).newInstance(new Object[] {}); // Create sparsifier String sparsifierName = jsapResult.getString("sparsifier"); Class<?> sparsifierClass = Class.forName(Sparsifier.class.getPackage().getName() + "." + sparsifierName); Sparsifier sparsifier = (Sparsifier) sparsifierClass .getConstructor(new Class[] {originalModel.getClass()}) .newInstance(new Object[] {originalModel}); LOGGER.info("Created a " + sparsifier.getClass().getSimpleName()); // Set sparsifier options if (!jsapResult.getBoolean("with-fraction")) { sparsifier.disableComputationOfPartialFractionOfPropagations(); LOGGER.info("Disabled the computation of fraction of propagations"); } if (jsapResult.getBoolean("incremental-likelihood")) { if (sparsifier instanceof GreedySparsifier) { ((GreedySparsifier) sparsifier).setIncrementalLikelihoodComputation(); LOGGER.info("Enabled incremental computation of likelihood (faster, experimental)"); } else { LOGGER.warn( "This type of sparsifier does not accept the --incrementa-likelihood switch, ignoring"); } } if (jsapResult.userSpecified("auxiliary-basename")) { // Use existing auxiliary file String auxiliaryBasename = jsapResult.getString("auxiliary-basename"); ICEstimateAuxiliary auxiliary = new ICEstimateAuxiliary(socNet, observations, null); LOGGER.info("Loading pre-computed auxiliary variables"); auxiliary.read(auxiliaryBasename); sparsifier.useAuxiliary(auxiliary); if (auxiliary .getCandidateSelectionPolicy() .toSpec() .equals(candidateSelectionPolicy.toSpec())) { LOGGER.info( "Candidate selection policy: " + auxiliary.getCandidateSelectionPolicy().toSpec()); } else { throw new IllegalArgumentException( "The candidate selection policies do not match: auxiliary has '" + auxiliary.getCandidateSelectionPolicy().toSpec() + "', sparsifier has '" + candidateSelectionPolicy.toSpec() + "'"); } } else { // Compute auxiliary variables LOGGER.info("Computing auxiliary variables"); sparsifier.computeAuxiliary(observations, candidateSelectionPolicy); } int maxSparseSize; if (jsapResult.userSpecified("sparse-model-size")) { maxSparseSize = jsapResult.getInt("sparse-model-size"); } else { maxSparseSize = originalModel.getProbs().cardinality(); LOGGER.info( "Setting target number of arcs to number of arcs with non-zero probability in the original model"); } // Open debug file if (jsapResult.userSpecified("debug-file")) { String debugFilename = jsapResult.getString("debug-file"); LOGGER.info("Will write debug output to " + debugFilename); sparsifier.openDebugFile(debugFilename); } int numOfChunks = 1; if (jsapResult.userSpecified("number-of-chunks")) { numOfChunks = jsapResult.getInt("number-of-chunks"); } ICModel sparseModel = runSparsifier( socNet, observations, originalModel, maxSparseSize, sparsifier, numOfChunks, true); // Write partial results to file if necessary if (jsapResult.userSpecified("measures-file")) { for (Measure m : sparsifier.partialResults.keySet()) { String logFilename = jsapResult.getString("measures-file"); switch (m) { case LOG_L: logFilename = logFilename + ".logL"; break; case FRACTION_OF_PROPAGATIONS: logFilename = logFilename + ".frac"; break; default: break; } PrintWriter report = new PrintWriter(new BufferedWriter(new FileWriter(new File(logFilename)))); LOGGER.info("Writing partial " + m.toString() + " results to " + logFilename); report.println("#k\t" + m.toString()); int[] ks = sparsifier.partialResults.get(m).keySet().toArray(new int[] {}); Arrays.sort(ks); for (int k : ks) { report.println(k + "\t" + sparsifier.partialResults.get(m).get(k)); } report.close(); } } // Dump probabilities if (jsapResult.userSpecified("output")) { String probsFilename = jsapResult.getString("output"); PrintWriter pw = Utilities.getPW(probsFilename); LOGGER.info("Dumping probabilities to " + probsFilename); sparseModel.dumpProbabilities(pw); pw.close(); } sparsifier.closeDebugFile(); }
protected double computeAndStorePartialLogLikelihood(int k, SparseDoubleMatrix2D newProbs) { ICModel model = new ICModel(originalModel.getSn(), newProbs); double logLikelihood = model.getLogLikelihoodIgnoringParentInformation(auxiliary); storePartialResult(Measure.LOG_L, k, logLikelihood); return logLikelihood; }