Logo ROOT  
Reference Guide
 
Loading...
Searching...
No Matches
MethodRuleFit.cxx
Go to the documentation of this file.
1// @(#)root/tmva $Id$
2// Author: Fredrik Tegenfeldt
3
4/**********************************************************************************
5 * Project: TMVA - a Root-integrated toolkit for multivariate data analysis *
6 * Package: TMVA *
7 * Class : MethodRuleFit *
8 * Web : http://tmva.sourceforge.net *
9 * *
10 * Description: *
11 * Implementation (see header file for description) *
12 * *
13 * Authors (alphabetical): *
14 * Fredrik Tegenfeldt <Fredrik.Tegenfeldt@cern.ch> - Iowa State U., USA *
15 * *
16 * Copyright (c) 2005: *
17 * CERN, Switzerland *
18 * Iowa State U. *
19 * MPI-K Heidelberg, Germany *
20 * *
21 * Redistribution and use in source and binary forms, with or without *
22 * modification, are permitted according to the terms listed in LICENSE *
23 * (http://tmva.sourceforge.net/LICENSE) *
24 **********************************************************************************/
25
26/*! \class TMVA::MethodRuleFit
27\ingroup TMVA
28J Friedman's RuleFit method
29*/
30
31#include "TMVA/MethodRuleFit.h"
32
34#include "TMVA/Config.h"
35#include "TMVA/Configurable.h"
36#include "TMVA/CrossEntropy.h"
37#include "TMVA/DataSet.h"
38#include "TMVA/DecisionTree.h"
39#include "TMVA/GiniIndex.h"
40#include "TMVA/IMethod.h"
41#include "TMVA/MethodBase.h"
43#include "TMVA/MsgLogger.h"
44#include "TMVA/Ranking.h"
45#include "TMVA/RuleFitAPI.h"
46#include "TMVA/SdivSqrtSplusB.h"
47#include "TMVA/SeparationBase.h"
48#include "TMVA/Timer.h"
49#include "TMVA/Tools.h"
50#include "TMVA/Types.h"
51
52#include "TRandom3.h"
53#include "TMatrix.h"
54
55#include <iostream>
56#include <iomanip>
57#include <algorithm>
58#include <list>
59#include <random>
60
61using std::min;
62
63REGISTER_METHOD(RuleFit)
64
66
67////////////////////////////////////////////////////////////////////////////////
68/// standard constructor
69
71 const TString& methodTitle,
72 DataSetInfo& theData,
73 const TString& theOption) :
74 MethodBase( jobName, Types::kRuleFit, methodTitle, theData, theOption)
75 , fSignalFraction(0)
76 , fNTImportance(0)
77 , fNTCoefficient(0)
78 , fNTSupport(0)
79 , fNTNcuts(0)
80 , fNTNvars(0)
81 , fNTPtag(0)
82 , fNTPss(0)
83 , fNTPsb(0)
84 , fNTPbs(0)
85 , fNTPbb(0)
86 , fNTSSB(0)
87 , fNTType(0)
88 , fUseRuleFitJF(kFALSE)
89 , fRFNrules(0)
90 , fRFNendnodes(0)
91 , fNTrees(0)
92 , fTreeEveFrac(0)
93 , fSepType(0)
94 , fMinFracNEve(0)
95 , fMaxFracNEve(0)
96 , fNCuts(0)
97 , fPruneMethod(TMVA::DecisionTree::kCostComplexityPruning)
98 , fPruneStrength(0)
99 , fUseBoost(kFALSE)
100 , fGDPathEveFrac(0)
101 , fGDValidEveFrac(0)
102 , fGDTau(0)
103 , fGDTauPrec(0)
104 , fGDTauMin(0)
105 , fGDTauMax(0)
106 , fGDTauScan(0)
107 , fGDPathStep(0)
108 , fGDNPathSteps(0)
109 , fGDErrScale(0)
110 , fMinimp(0)
111 , fRuleMinDist(0)
112 , fLinQuantile(0)
113{
114 fMonitorNtuple = NULL;
115}
116
117////////////////////////////////////////////////////////////////////////////////
118/// constructor from weight file
119
121 const TString& theWeightFile) :
122 MethodBase( Types::kRuleFit, theData, theWeightFile)
123 , fSignalFraction(0)
124 , fNTImportance(0)
125 , fNTCoefficient(0)
126 , fNTSupport(0)
127 , fNTNcuts(0)
128 , fNTNvars(0)
129 , fNTPtag(0)
130 , fNTPss(0)
131 , fNTPsb(0)
132 , fNTPbs(0)
133 , fNTPbb(0)
134 , fNTSSB(0)
135 , fNTType(0)
136 , fUseRuleFitJF(kFALSE)
137 , fRFNrules(0)
138 , fRFNendnodes(0)
139 , fNTrees(0)
140 , fTreeEveFrac(0)
141 , fSepType(0)
142 , fMinFracNEve(0)
143 , fMaxFracNEve(0)
144 , fNCuts(0)
145 , fPruneMethod(TMVA::DecisionTree::kCostComplexityPruning)
146 , fPruneStrength(0)
147 , fUseBoost(kFALSE)
148 , fGDPathEveFrac(0)
149 , fGDValidEveFrac(0)
150 , fGDTau(0)
151 , fGDTauPrec(0)
152 , fGDTauMin(0)
153 , fGDTauMax(0)
154 , fGDTauScan(0)
155 , fGDPathStep(0)
156 , fGDNPathSteps(0)
157 , fGDErrScale(0)
158 , fMinimp(0)
159 , fRuleMinDist(0)
160 , fLinQuantile(0)
161{
162 fMonitorNtuple = NULL;
163}
164
165////////////////////////////////////////////////////////////////////////////////
166/// destructor
167
169{
170 for (UInt_t i=0; i<fEventSample.size(); i++) delete fEventSample[i];
171 for (UInt_t i=0; i<fForest.size(); i++) delete fForest[i];
172}
173
174////////////////////////////////////////////////////////////////////////////////
175/// RuleFit can handle classification with 2 classes
176
178{
179 if (type == Types::kClassification && numberClasses == 2) return kTRUE;
180 return kFALSE;
181}
182
183////////////////////////////////////////////////////////////////////////////////
184/// define the options (their key words) that can be set in the option string
185/// know options.
186///
187/// #### general
188///
189/// - RuleFitModule <string>
190/// available values are:
191/// - RFTMVA - use TMVA implementation
192/// - RFFriedman - use Friedmans original implementation
193///
194/// #### Path search (fitting)
195///
196/// - GDTau <float> gradient-directed path: fit threshold, default
197/// - GDTauPrec <float> gradient-directed path: precision of estimated tau
198/// - GDStep <float> gradient-directed path: step size
199/// - GDNSteps <float> gradient-directed path: number of steps
200/// - GDErrScale <float> stop scan when error>scale*errmin
201///
202/// #### Tree generation
203///
204/// - fEventsMin <float> minimum fraction of events in a splittable node
205/// - fEventsMax <float> maximum fraction of events in a splittable node
206/// - nTrees <float> number of trees in forest.
207/// - ForestType <string>
208/// available values are:
209/// - Random - create forest using random subsample and only random variables subset at each node
210/// - AdaBoost - create forest with boosted events
211///
212/// #### Model creation
213///
214/// - RuleMinDist <float> min distance allowed between rules
215/// - MinImp <float> minimum rule importance accepted
216/// - Model <string> model to be used
217/// available values are:
218/// - ModRuleLinear <default>
219/// - ModRule
220/// - ModLinear
221///
222/// #### Friedmans module
223///
224/// - RFWorkDir <string> directory where Friedmans module (rf_go.exe) is installed
225/// - RFNrules <int> maximum number of rules allowed
226/// - RFNendnodes <int> average number of end nodes in the forest of trees
227
229{
230 DeclareOptionRef(fGDTau=-1, "GDTau", "Gradient-directed (GD) path: default fit cut-off");
231 DeclareOptionRef(fGDTauPrec=0.01, "GDTauPrec", "GD path: precision of tau");
232 DeclareOptionRef(fGDPathStep=0.01, "GDStep", "GD path: step size");
233 DeclareOptionRef(fGDNPathSteps=10000, "GDNSteps", "GD path: number of steps");
234 DeclareOptionRef(fGDErrScale=1.1, "GDErrScale", "Stop scan when error > scale*errmin");
235 DeclareOptionRef(fLinQuantile, "LinQuantile", "Quantile of linear terms (removes outliers)");
236 DeclareOptionRef(fGDPathEveFrac=0.5, "GDPathEveFrac", "Fraction of events used for the path search");
237 DeclareOptionRef(fGDValidEveFrac=0.5, "GDValidEveFrac", "Fraction of events used for the validation");
238 // tree options
239 DeclareOptionRef(fMinFracNEve=0.1, "fEventsMin", "Minimum fraction of events in a splittable node");
240 DeclareOptionRef(fMaxFracNEve=0.9, "fEventsMax", "Maximum fraction of events in a splittable node");
241 DeclareOptionRef(fNTrees=20, "nTrees", "Number of trees in forest.");
242
243 DeclareOptionRef(fForestTypeS="AdaBoost", "ForestType", "Method to use for forest generation (AdaBoost or RandomForest)");
244 AddPreDefVal(TString("AdaBoost"));
245 AddPreDefVal(TString("Random"));
246 // rule cleanup options
247 DeclareOptionRef(fRuleMinDist=0.001, "RuleMinDist", "Minimum distance between rules");
248 DeclareOptionRef(fMinimp=0.01, "MinImp", "Minimum rule importance accepted");
249 // rule model option
250 DeclareOptionRef(fModelTypeS="ModRuleLinear", "Model", "Model to be used");
251 AddPreDefVal(TString("ModRule"));
252 AddPreDefVal(TString("ModRuleLinear"));
253 AddPreDefVal(TString("ModLinear"));
254 DeclareOptionRef(fRuleFitModuleS="RFTMVA", "RuleFitModule","Which RuleFit module to use");
255 AddPreDefVal(TString("RFTMVA"));
256 AddPreDefVal(TString("RFFriedman"));
257
258 DeclareOptionRef(fRFWorkDir="./rulefit", "RFWorkDir", "Friedman\'s RuleFit module (RFF): working dir");
259 DeclareOptionRef(fRFNrules=2000, "RFNrules", "RFF: Mximum number of rules");
260 DeclareOptionRef(fRFNendnodes=4, "RFNendnodes", "RFF: Average number of end nodes");
261}
262
263////////////////////////////////////////////////////////////////////////////////
264/// process the options specified by the user
265
267{
268 if (IgnoreEventsWithNegWeightsInTraining()) {
269 Log() << kFATAL << "Mechanism to ignore events with negative weights in training not yet available for method: "
270 << GetMethodTypeName()
271 << " --> please remove \"IgnoreNegWeightsInTraining\" option from booking string."
272 << Endl;
273 }
274
275 fRuleFitModuleS.ToLower();
276 if (fRuleFitModuleS == "rftmva") fUseRuleFitJF = kFALSE;
277 else if (fRuleFitModuleS == "rffriedman") fUseRuleFitJF = kTRUE;
278 else fUseRuleFitJF = kTRUE;
279
280 fSepTypeS.ToLower();
281 if (fSepTypeS == "misclassificationerror") fSepType = new MisClassificationError();
282 else if (fSepTypeS == "giniindex") fSepType = new GiniIndex();
283 else if (fSepTypeS == "crossentropy") fSepType = new CrossEntropy();
284 else fSepType = new SdivSqrtSplusB();
285
286 fModelTypeS.ToLower();
287 if (fModelTypeS == "modlinear" ) fRuleFit.SetModelLinear();
288 else if (fModelTypeS == "modrule" ) fRuleFit.SetModelRules();
289 else fRuleFit.SetModelFull();
290
291 fPruneMethodS.ToLower();
292 if (fPruneMethodS == "expectederror" ) fPruneMethod = DecisionTree::kExpectedErrorPruning;
293 else if (fPruneMethodS == "costcomplexity" ) fPruneMethod = DecisionTree::kCostComplexityPruning;
294 else fPruneMethod = DecisionTree::kNoPruning;
295
296 fForestTypeS.ToLower();
297 if (fForestTypeS == "random" ) fUseBoost = kFALSE;
298 else if (fForestTypeS == "adaboost" ) fUseBoost = kTRUE;
299 else fUseBoost = kTRUE;
300 //
301 // if creating the forest by boosting the events
302 // the full training sample is used per tree
303 // -> only true for the TMVA version of RuleFit.
304 if (fUseBoost && (!fUseRuleFitJF)) fTreeEveFrac = 1.0;
305
306 // check event fraction for tree generation
307 // if <0 set to automatic number
308 if (fTreeEveFrac<=0) {
309 Int_t nevents = Data()->GetNTrainingEvents();
310 Double_t n = static_cast<Double_t>(nevents);
311 fTreeEveFrac = min( 0.5, (100.0 +6.0*sqrt(n))/n);
312 }
313 // verify ranges of options
314 VerifyRange(Log(), "nTrees", fNTrees,0,100000,20);
315 VerifyRange(Log(), "MinImp", fMinimp,0.0,1.0,0.0);
316 VerifyRange(Log(), "GDTauPrec", fGDTauPrec,1e-5,5e-1);
317 VerifyRange(Log(), "GDTauMin", fGDTauMin,0.0,1.0);
318 VerifyRange(Log(), "GDTauMax", fGDTauMax,fGDTauMin,1.0);
319 VerifyRange(Log(), "GDPathStep", fGDPathStep,0.0,100.0,0.01);
320 VerifyRange(Log(), "GDErrScale", fGDErrScale,1.0,100.0,1.1);
321 VerifyRange(Log(), "GDPathEveFrac", fGDPathEveFrac,0.01,0.9,0.5);
322 VerifyRange(Log(), "GDValidEveFrac",fGDValidEveFrac,0.01,1.0-fGDPathEveFrac,1.0-fGDPathEveFrac);
323 VerifyRange(Log(), "fEventsMin", fMinFracNEve,0.0,1.0);
324 VerifyRange(Log(), "fEventsMax", fMaxFracNEve,fMinFracNEve,1.0);
325
326 fRuleFit.GetRuleEnsemblePtr()->SetLinQuantile(fLinQuantile);
327 fRuleFit.GetRuleFitParamsPtr()->SetGDTauRange(fGDTauMin,fGDTauMax);
328 fRuleFit.GetRuleFitParamsPtr()->SetGDTau(fGDTau);
329 fRuleFit.GetRuleFitParamsPtr()->SetGDTauPrec(fGDTauPrec);
330 fRuleFit.GetRuleFitParamsPtr()->SetGDTauScan(fGDTauScan);
331 fRuleFit.GetRuleFitParamsPtr()->SetGDPathStep(fGDPathStep);
332 fRuleFit.GetRuleFitParamsPtr()->SetGDNPathSteps(fGDNPathSteps);
333 fRuleFit.GetRuleFitParamsPtr()->SetGDErrScale(fGDErrScale);
334 fRuleFit.SetImportanceCut(fMinimp);
335 fRuleFit.SetRuleMinDist(fRuleMinDist);
336
337
338 // check if Friedmans module is used.
339 // print a message concerning the options.
340 if (fUseRuleFitJF) {
341 Log() << kINFO << "" << Endl;
342 Log() << kINFO << "--------------------------------------" <<Endl;
343 Log() << kINFO << "Friedmans RuleFit module is selected." << Endl;
344 Log() << kINFO << "Only the following options are used:" << Endl;
345 Log() << kINFO << Endl;
346 Log() << kINFO << gTools().Color("bold") << " Model" << gTools().Color("reset") << Endl;
347 Log() << kINFO << gTools().Color("bold") << " RFWorkDir" << gTools().Color("reset") << Endl;
348 Log() << kINFO << gTools().Color("bold") << " RFNrules" << gTools().Color("reset") << Endl;
349 Log() << kINFO << gTools().Color("bold") << " RFNendnodes" << gTools().Color("reset") << Endl;
350 Log() << kINFO << gTools().Color("bold") << " GDNPathSteps" << gTools().Color("reset") << Endl;
351 Log() << kINFO << gTools().Color("bold") << " GDPathStep" << gTools().Color("reset") << Endl;
352 Log() << kINFO << gTools().Color("bold") << " GDErrScale" << gTools().Color("reset") << Endl;
353 Log() << kINFO << "--------------------------------------" <<Endl;
354 Log() << kINFO << Endl;
355 }
356
357 // Select what weight to use in the 'importance' rule visualisation plots.
358 // Note that if UseCoefficientsVisHists() is selected, the following weight is used:
359 // w = rule coefficient * rule support
360 // The support is a positive number which is 0 if no events are accepted by the rule.
361 // Normally the importance gives more useful information.
362 //
363 //fRuleFit.UseCoefficientsVisHists();
364 fRuleFit.UseImportanceVisHists();
365
366 fRuleFit.SetMsgType( Log().GetMinType() );
367
368 if (HasTrainingTree()) InitEventSample();
369
370}
371
372////////////////////////////////////////////////////////////////////////////////
373/// initialize the monitoring ntuple
374
376{
377 BaseDir()->cd();
378 fMonitorNtuple= new TTree("MonitorNtuple_RuleFit","RuleFit variables");
379 fMonitorNtuple->Branch("importance",&fNTImportance,"importance/D");
380 fMonitorNtuple->Branch("support",&fNTSupport,"support/D");
381 fMonitorNtuple->Branch("coefficient",&fNTCoefficient,"coefficient/D");
382 fMonitorNtuple->Branch("ncuts",&fNTNcuts,"ncuts/I");
383 fMonitorNtuple->Branch("nvars",&fNTNvars,"nvars/I");
384 fMonitorNtuple->Branch("type",&fNTType,"type/I");
385 fMonitorNtuple->Branch("ptag",&fNTPtag,"ptag/D");
386 fMonitorNtuple->Branch("pss",&fNTPss,"pss/D");
387 fMonitorNtuple->Branch("psb",&fNTPsb,"psb/D");
388 fMonitorNtuple->Branch("pbs",&fNTPbs,"pbs/D");
389 fMonitorNtuple->Branch("pbb",&fNTPbb,"pbb/D");
390 fMonitorNtuple->Branch("soversb",&fNTSSB,"soversb/D");
391}
392
393////////////////////////////////////////////////////////////////////////////////
394/// default initialization
395
397{
398 // the minimum requirement to declare an event signal-like
399 SetSignalReferenceCut( 0.0 );
400
401 // set variables that used to be options
402 // any modifications are then made in ProcessOptions()
403 fLinQuantile = 0.025; // Quantile of linear terms (remove outliers)
404 fTreeEveFrac = -1.0; // Fraction of events used to train each tree
405 fNCuts = 20; // Number of steps during node cut optimisation
406 fSepTypeS = "GiniIndex"; // Separation criterion for node splitting; see BDT
407 fPruneMethodS = "NONE"; // Pruning method; see BDT
408 fPruneStrength = 3.5; // Pruning strength; see BDT
409 fGDTauMin = 0.0; // Gradient-directed path: min fit threshold (tau)
410 fGDTauMax = 1.0; // Gradient-directed path: max fit threshold (tau)
411 fGDTauScan = 1000; // Gradient-directed path: number of points scanning for best tau
412
413}
414
415////////////////////////////////////////////////////////////////////////////////
416/// write all Events from the Tree into a vector of Events, that are
417/// more easily manipulated.
418/// This method should never be called without existing trainingTree, as it
419/// the vector of events from the ROOT training tree
420
422{
423 if (Data()->GetNEvents()==0) Log() << kFATAL << "<Init> Data().TrainingTree() is zero pointer" << Endl;
424
425 Int_t nevents = Data()->GetNEvents();
426 for (Int_t ievt=0; ievt<nevents; ievt++){
427 const Event * ev = GetEvent(ievt);
428 fEventSample.push_back( new Event(*ev));
429 }
430 if (fTreeEveFrac<=0) {
431 Double_t n = static_cast<Double_t>(nevents);
432 fTreeEveFrac = min( 0.5, (100.0 +6.0*sqrt(n))/n);
433 }
434 if (fTreeEveFrac>1.0) fTreeEveFrac=1.0;
435 //
436 std::shuffle(fEventSample.begin(), fEventSample.end(), std::default_random_engine{});
437 //
438 Log() << kDEBUG << "Set sub-sample fraction to " << fTreeEveFrac << Endl;
439}
440
441////////////////////////////////////////////////////////////////////////////////
442
444{
446 // training of rules
447
448 if(!IsSilentFile()) InitMonitorNtuple();
449
450 // fill the STL Vector with the event sample
451 this->InitEventSample();
452
453 if (fUseRuleFitJF) {
454 TrainJFRuleFit();
455 }
456 else {
457 TrainTMVARuleFit();
458 }
459 fRuleFit.GetRuleEnsemblePtr()->ClearRuleMap();
461 ExitFromTraining();
462}
463
464////////////////////////////////////////////////////////////////////////////////
465/// training of rules using TMVA implementation
466
468{
469 if (IsNormalised()) Log() << kFATAL << "\"Normalise\" option cannot be used with RuleFit; "
470 << "please remove the option from the configuration string, or "
471 << "use \"!Normalise\""
472 << Endl;
473
474 // timer
475 Timer timer( 1, GetName() );
476
477 // test tree nmin cut -> for debug purposes
478 // the routine will generate trees with stopping cut on N(eve) given by
479 // a fraction between [20,N(eve)-1].
480 //
481 // MakeForestRnd();
482 // exit(1);
483 //
484
485 // Init RuleFit object and create rule ensemble
486 // + make forest & rules
487 fRuleFit.Initialize( this );
488
489 // Make forest of decision trees
490 // if (fRuleFit.GetRuleEnsemble().DoRules()) fRuleFit.MakeForest();
491
492 // Fit the rules
493 Log() << kDEBUG << "Fitting rule coefficients ..." << Endl;
494 fRuleFit.FitCoefficients();
495
496 // Calculate importance
497 Log() << kDEBUG << "Computing rule and variable importance" << Endl;
498 fRuleFit.CalcImportance();
499
500 // Output results and fill monitor ntuple
501 fRuleFit.GetRuleEnsemblePtr()->Print();
502 //
503 if(!IsSilentFile())
504 {
505 Log() << kDEBUG << "Filling rule ntuple" << Endl;
506 UInt_t nrules = fRuleFit.GetRuleEnsemble().GetRulesConst().size();
507 const Rule *rule;
508 for (UInt_t i=0; i<nrules; i++ ) {
509 rule = fRuleFit.GetRuleEnsemble().GetRulesConst(i);
510 fNTImportance = rule->GetRelImportance();
511 fNTSupport = rule->GetSupport();
512 fNTCoefficient = rule->GetCoefficient();
513 fNTType = (rule->IsSignalRule() ? 1:-1 );
514 fNTNvars = rule->GetRuleCut()->GetNvars();
515 fNTNcuts = rule->GetRuleCut()->GetNcuts();
516 fNTPtag = fRuleFit.GetRuleEnsemble().GetRulePTag(i); // should be identical with support
517 fNTPss = fRuleFit.GetRuleEnsemble().GetRulePSS(i);
518 fNTPsb = fRuleFit.GetRuleEnsemble().GetRulePSB(i);
519 fNTPbs = fRuleFit.GetRuleEnsemble().GetRulePBS(i);
520 fNTPbb = fRuleFit.GetRuleEnsemble().GetRulePBB(i);
521 fNTSSB = rule->GetSSB();
522 fMonitorNtuple->Fill();
523 }
524
525 fRuleFit.MakeVisHists();
526 fRuleFit.MakeDebugHists();
527 }
528 Log() << kDEBUG << "Training done" << Endl;
529
530}
531
532////////////////////////////////////////////////////////////////////////////////
533/// training of rules using Jerome Friedmans implementation
534
536{
537 fRuleFit.InitPtrs( this );
538 Data()->SetCurrentType(Types::kTraining);
539 UInt_t nevents = Data()->GetNTrainingEvents();
540 std::vector<const TMVA::Event*> tmp;
541 for (Long64_t ievt=0; ievt<nevents; ievt++) {
542 const Event *event = GetEvent(ievt);
543 tmp.push_back(event);
544 }
545 fRuleFit.SetTrainingEvents( tmp );
546
547 RuleFitAPI *rfAPI = new RuleFitAPI( this, &fRuleFit, Log().GetMinType() );
548
549 rfAPI->WelcomeMessage();
550
551 // timer
552 Timer timer( 1, GetName() );
553
554 Log() << kINFO << "Training ..." << Endl;
555 rfAPI->TrainRuleFit();
556
557 Log() << kDEBUG << "reading model summary from rf_go.exe output" << Endl;
558 rfAPI->ReadModelSum();
559
560 // fRuleFit.GetRuleEnsemblePtr()->MakeRuleMap();
561
562 Log() << kDEBUG << "calculating rule and variable importance" << Endl;
563 fRuleFit.CalcImportance();
564
565 // Output results and fill monitor ntuple
566 fRuleFit.GetRuleEnsemblePtr()->Print();
567 //
568 if(!IsSilentFile())fRuleFit.MakeVisHists();
569
570 delete rfAPI;
571
572 Log() << kDEBUG << "done training" << Endl;
573}
574
575////////////////////////////////////////////////////////////////////////////////
576/// computes ranking of input variables
577
579{
580 // create the ranking object
581 fRanking = new Ranking( GetName(), "Importance" );
582
583 for (UInt_t ivar=0; ivar<GetNvar(); ivar++) {
584 fRanking->AddRank( Rank( GetInputLabel(ivar), fRuleFit.GetRuleEnsemble().GetVarImportance(ivar) ) );
585 }
586
587 return fRanking;
588}
589
590////////////////////////////////////////////////////////////////////////////////
591/// add the rules to XML node
592
593void TMVA::MethodRuleFit::AddWeightsXMLTo( void* parent ) const
594{
595 fRuleFit.GetRuleEnsemble().AddXMLTo( parent );
596}
597
598////////////////////////////////////////////////////////////////////////////////
599/// read rules from an std::istream
600
602{
603 fRuleFit.GetRuleEnsemblePtr()->ReadRaw( istr );
604}
605
606////////////////////////////////////////////////////////////////////////////////
607/// read rules from XML node
608
610{
611 fRuleFit.GetRuleEnsemblePtr()->ReadFromXML( wghtnode );
612}
613
614////////////////////////////////////////////////////////////////////////////////
615/// returns MVA value for given event
616
618{
619 // cannot determine error
620 NoErrorCalc(err, errUpper);
621
622 return fRuleFit.EvalEvent( *GetEvent() );
623}
624
625////////////////////////////////////////////////////////////////////////////////
626/// write special monitoring histograms to file (here ntuple)
627
629{
630 BaseDir()->cd();
631 Log() << kINFO << "Write monitoring ntuple to file: " << BaseDir()->GetPath() << Endl;
632 fMonitorNtuple->Write();
633}
634
635////////////////////////////////////////////////////////////////////////////////
636/// write specific classifier response
637
638void TMVA::MethodRuleFit::MakeClassSpecific( std::ostream& fout, const TString& className ) const
639{
640 Int_t dp = fout.precision();
641 fout << " // not implemented for class: \"" << className << "\"" << std::endl;
642 fout << "};" << std::endl;
643 fout << "void " << className << "::Initialize(){}" << std::endl;
644 fout << "void " << className << "::Clear(){}" << std::endl;
645 fout << "double " << className << "::GetMvaValue__( const std::vector<double>& inputValues ) const {" << std::endl;
646 fout << " double rval=" << std::setprecision(10) << fRuleFit.GetRuleEnsemble().GetOffset() << ";" << std::endl;
647 MakeClassRuleCuts(fout);
648 MakeClassLinear(fout);
649 fout << " return rval;" << std::endl;
650 fout << "}" << std::endl;
651 fout << std::setprecision(dp);
652}
653
654////////////////////////////////////////////////////////////////////////////////
655/// print out the rule cuts
656
657void TMVA::MethodRuleFit::MakeClassRuleCuts( std::ostream& fout ) const
658{
659 Int_t dp = fout.precision();
660 if (!fRuleFit.GetRuleEnsemble().DoRules()) {
661 fout << " //" << std::endl;
662 fout << " // ==> MODEL CONTAINS NO RULES <==" << std::endl;
663 fout << " //" << std::endl;
664 return;
665 }
666 const RuleEnsemble *rens = &(fRuleFit.GetRuleEnsemble());
667 const std::vector< Rule* > *rules = &(rens->GetRulesConst());
668 const RuleCut *ruleCut;
669 //
670 std::list< std::pair<Double_t,Int_t> > sortedRules;
671 for (UInt_t ir=0; ir<rules->size(); ir++) {
672 sortedRules.push_back( std::pair<Double_t,Int_t>( (*rules)[ir]->GetImportance()/rens->GetImportanceRef(),ir ) );
673 }
674 sortedRules.sort();
675 //
676 fout << " //" << std::endl;
677 fout << " // here follows all rules ordered in importance (most important first)" << std::endl;
678 fout << " // at the end of each line, the relative importance of the rule is given" << std::endl;
679 fout << " //" << std::endl;
680 //
681 for ( std::list< std::pair<double,int> >::reverse_iterator itpair = sortedRules.rbegin();
682 itpair != sortedRules.rend(); ++itpair ) {
683 UInt_t ir = itpair->second;
684 Double_t impr = itpair->first;
685 ruleCut = (*rules)[ir]->GetRuleCut();
686 if (impr<rens->GetImportanceCut()) fout << " //" << std::endl;
687 fout << " if (" << std::flush;
688 for (UInt_t ic=0; ic<ruleCut->GetNvars(); ic++) {
689 Double_t sel = ruleCut->GetSelector(ic);
690 Double_t valmin = ruleCut->GetCutMin(ic);
691 Double_t valmax = ruleCut->GetCutMax(ic);
692 Bool_t domin = ruleCut->GetCutDoMin(ic);
693 Bool_t domax = ruleCut->GetCutDoMax(ic);
694 //
695 if (ic>0) fout << "&&" << std::flush;
696 if (domin) {
697 fout << "(" << std::setprecision(10) << valmin << std::flush;
698 fout << "<inputValues[" << sel << "])" << std::flush;
699 }
700 if (domax) {
701 if (domin) fout << "&&" << std::flush;
702 fout << "(inputValues[" << sel << "]" << std::flush;
703 fout << "<" << std::setprecision(10) << valmax << ")" <<std::flush;
704 }
705 }
706 fout << ") rval+=" << std::setprecision(10) << (*rules)[ir]->GetCoefficient() << ";" << std::flush;
707 fout << " // importance = " << Form("%3.3f",impr) << std::endl;
708 }
709 fout << std::setprecision(dp);
710}
711
712////////////////////////////////////////////////////////////////////////////////
713/// print out the linear terms
714
715void TMVA::MethodRuleFit::MakeClassLinear( std::ostream& fout ) const
716{
717 if (!fRuleFit.GetRuleEnsemble().DoLinear()) {
718 fout << " //" << std::endl;
719 fout << " // ==> MODEL CONTAINS NO LINEAR TERMS <==" << std::endl;
720 fout << " //" << std::endl;
721 return;
722 }
723 fout << " //" << std::endl;
724 fout << " // here follows all linear terms" << std::endl;
725 fout << " // at the end of each line, the relative importance of the term is given" << std::endl;
726 fout << " //" << std::endl;
727 const RuleEnsemble *rens = &(fRuleFit.GetRuleEnsemble());
728 UInt_t nlin = rens->GetNLinear();
729 for (UInt_t il=0; il<nlin; il++) {
730 if (rens->IsLinTermOK(il)) {
731 Double_t norm = rens->GetLinNorm(il);
732 Double_t imp = rens->GetLinImportance(il)/rens->GetImportanceRef();
733 fout << " rval+="
734 // << std::setprecision(10) << rens->GetLinCoefficients(il)*norm << "*std::min(" << setprecision(10) << rens->GetLinDP(il)
735 // << ", std::max( inputValues[" << il << "]," << std::setprecision(10) << rens->GetLinDM(il) << "));"
736 << std::setprecision(10) << rens->GetLinCoefficients(il)*norm
737 << "*std::min( double(" << std::setprecision(10) << rens->GetLinDP(il)
738 << "), std::max( double(inputValues[" << il << "]), double(" << std::setprecision(10) << rens->GetLinDM(il) << ")));"
739 << std::flush;
740 fout << " // importance = " << Form("%3.3f",imp) << std::endl;
741 }
742 }
743}
744
745////////////////////////////////////////////////////////////////////////////////
746/// get help message text
747///
748/// typical length of text line:
749/// "|--------------------------------------------------------------|"
750
752{
753 TString col = gConfig().WriteOptionsReference() ? TString() : gTools().Color("bold");
754 TString colres = gConfig().WriteOptionsReference() ? TString() : gTools().Color("reset");
755 TString brk = gConfig().WriteOptionsReference() ? "<br>" : "";
756
757 Log() << Endl;
758 Log() << col << "--- Short description:" << colres << Endl;
759 Log() << Endl;
760 Log() << "This method uses a collection of so called rules to create a" << Endl;
761 Log() << "discriminating scoring function. Each rule consists of a series" << Endl;
762 Log() << "of cuts in parameter space. The ensemble of rules are created" << Endl;
763 Log() << "from a forest of decision trees, trained using the training data." << Endl;
764 Log() << "Each node (apart from the root) corresponds to one rule." << Endl;
765 Log() << "The scoring function is then obtained by linearly combining" << Endl;
766 Log() << "the rules. A fitting procedure is applied to find the optimum" << Endl;
767 Log() << "set of coefficients. The goal is to find a model with few rules" << Endl;
768 Log() << "but with a strong discriminating power." << Endl;
769 Log() << Endl;
770 Log() << col << "--- Performance optimisation:" << colres << Endl;
771 Log() << Endl;
772 Log() << "There are two important considerations to make when optimising:" << Endl;
773 Log() << Endl;
774 Log() << " 1. Topology of the decision tree forest" << brk << Endl;
775 Log() << " 2. Fitting of the coefficients" << Endl;
776 Log() << Endl;
777 Log() << "The maximum complexity of the rules is defined by the size of" << Endl;
778 Log() << "the trees. Large trees will yield many complex rules and capture" << Endl;
779 Log() << "higher order correlations. On the other hand, small trees will" << Endl;
780 Log() << "lead to a smaller ensemble with simple rules, only capable of" << Endl;
781 Log() << "modeling simple structures." << Endl;
782 Log() << "Several parameters exists for controlling the complexity of the" << Endl;
783 Log() << "rule ensemble." << Endl;
784 Log() << Endl;
785 Log() << "The fitting procedure searches for a minimum using a gradient" << Endl;
786 Log() << "directed path. Apart from step size and number of steps, the" << Endl;
787 Log() << "evolution of the path is defined by a cut-off parameter, tau." << Endl;
788 Log() << "This parameter is unknown and depends on the training data." << Endl;
789 Log() << "A large value will tend to give large weights to a few rules." << Endl;
790 Log() << "Similarly, a small value will lead to a large set of rules" << Endl;
791 Log() << "with similar weights." << Endl;
792 Log() << Endl;
793 Log() << "A final point is the model used; rules and/or linear terms." << Endl;
794 Log() << "For a given training sample, the result may improve by adding" << Endl;
795 Log() << "linear terms. If best performance is obtained using only linear" << Endl;
796 Log() << "terms, it is very likely that the Fisher discriminant would be" << Endl;
797 Log() << "a better choice. Ideally the fitting procedure should be able to" << Endl;
798 Log() << "make this choice by giving appropriate weights for either terms." << Endl;
799 Log() << Endl;
800 Log() << col << "--- Performance tuning via configuration options:" << colres << Endl;
801 Log() << Endl;
802 Log() << "I. TUNING OF RULE ENSEMBLE:" << Endl;
803 Log() << Endl;
804 Log() << " " << col << "ForestType " << colres
805 << ": Recommended is to use the default \"AdaBoost\"." << brk << Endl;
806 Log() << " " << col << "nTrees " << colres
807 << ": More trees leads to more rules but also slow" << Endl;
808 Log() << " performance. With too few trees the risk is" << Endl;
809 Log() << " that the rule ensemble becomes too simple." << brk << Endl;
810 Log() << " " << col << "fEventsMin " << colres << brk << Endl;
811 Log() << " " << col << "fEventsMax " << colres
812 << ": With a lower min, more large trees will be generated" << Endl;
813 Log() << " leading to more complex rules." << Endl;
814 Log() << " With a higher max, more small trees will be" << Endl;
815 Log() << " generated leading to more simple rules." << Endl;
816 Log() << " By changing this range, the average complexity" << Endl;
817 Log() << " of the rule ensemble can be controlled." << brk << Endl;
818 Log() << " " << col << "RuleMinDist " << colres
819 << ": By increasing the minimum distance between" << Endl;
820 Log() << " rules, fewer and more diverse rules will remain." << Endl;
821 Log() << " Initially it is a good idea to keep this small" << Endl;
822 Log() << " or zero and let the fitting do the selection of" << Endl;
823 Log() << " rules. In order to reduce the ensemble size," << Endl;
824 Log() << " the value can then be increased." << Endl;
825 Log() << Endl;
826 // "|--------------------------------------------------------------|"
827 Log() << "II. TUNING OF THE FITTING:" << Endl;
828 Log() << Endl;
829 Log() << " " << col << "GDPathEveFrac " << colres
830 << ": fraction of events in path evaluation" << Endl;
831 Log() << " Increasing this fraction will improve the path" << Endl;
832 Log() << " finding. However, a too high value will give few" << Endl;
833 Log() << " unique events available for error estimation." << Endl;
834 Log() << " It is recommended to use the default = 0.5." << brk << Endl;
835 Log() << " " << col << "GDTau " << colres
836 << ": cutoff parameter tau" << Endl;
837 Log() << " By default this value is set to -1.0." << Endl;
838 // "|----------------|---------------------------------------------|"
839 Log() << " This means that the cut off parameter is" << Endl;
840 Log() << " automatically estimated. In most cases" << Endl;
841 Log() << " this should be fine. However, you may want" << Endl;
842 Log() << " to fix this value if you already know it" << Endl;
843 Log() << " and want to reduce on training time." << brk << Endl;
844 Log() << " " << col << "GDTauPrec " << colres
845 << ": precision of estimated tau" << Endl;
846 Log() << " Increase this precision to find a more" << Endl;
847 Log() << " optimum cut-off parameter." << brk << Endl;
848 Log() << " " << col << "GDNStep " << colres
849 << ": number of steps in path search" << Endl;
850 Log() << " If the number of steps is too small, then" << Endl;
851 Log() << " the program will give a warning message." << Endl;
852 Log() << Endl;
853 Log() << "III. WARNING MESSAGES" << Endl;
854 Log() << Endl;
855 Log() << col << "Risk(i+1)>=Risk(i) in path" << colres << brk << Endl;
856 Log() << col << "Chaotic behaviour of risk evolution." << colres << Endl;
857 // "|----------------|---------------------------------------------|"
858 Log() << " The error rate was still decreasing at the end" << Endl;
859 Log() << " By construction the Risk should always decrease." << Endl;
860 Log() << " However, if the training sample is too small or" << Endl;
861 Log() << " the model is overtrained, such warnings can" << Endl;
862 Log() << " occur." << Endl;
863 Log() << " The warnings can safely be ignored if only a" << Endl;
864 Log() << " few (<3) occur. If more warnings are generated," << Endl;
865 Log() << " the fitting fails." << Endl;
866 Log() << " A remedy may be to increase the value" << brk << Endl;
867 Log() << " "
868 << col << "GDValidEveFrac" << colres
869 << " to 1.0 (or a larger value)." << brk << Endl;
870 Log() << " In addition, if "
871 << col << "GDPathEveFrac" << colres
872 << " is too high" << Endl;
873 Log() << " the same warnings may occur since the events" << Endl;
874 Log() << " used for error estimation are also used for" << Endl;
875 Log() << " path estimation." << Endl;
876 Log() << " Another possibility is to modify the model - " << Endl;
877 Log() << " See above on tuning the rule ensemble." << Endl;
878 Log() << Endl;
879 Log() << col << "The error rate was still decreasing at the end of the path"
880 << colres << Endl;
881 Log() << " Too few steps in path! Increase "
882 << col << "GDNSteps" << colres << "." << Endl;
883 Log() << Endl;
884 Log() << col << "Reached minimum early in the search" << colres << Endl;
885
886 Log() << " Minimum was found early in the fitting. This" << Endl;
887 Log() << " may indicate that the used step size "
888 << col << "GDStep" << colres << "." << Endl;
889 Log() << " was too large. Reduce it and rerun." << Endl;
890 Log() << " If the results still are not OK, modify the" << Endl;
891 Log() << " model either by modifying the rule ensemble" << Endl;
892 Log() << " or add/remove linear terms" << Endl;
893}
#define REGISTER_METHOD(CLASS)
for example
#define e(i)
Definition RSha256.hxx:103
const Bool_t kFALSE
Definition RtypesCore.h:101
double Double_t
Definition RtypesCore.h:59
long long Long64_t
Definition RtypesCore.h:80
const Bool_t kTRUE
Definition RtypesCore.h:100
#define ClassImp(name)
Definition Rtypes.h:364
int type
Definition TGX11.cxx:121
char * Form(const char *fmt,...)
Bool_t WriteOptionsReference() const
Definition Config.h:65
Implementation of the CrossEntropy as separation criterion.
Class that contains all the data information.
Definition DataSetInfo.h:62
static void SetIsTraining(bool on)
Implementation of a Decision Tree.
Implementation of the GiniIndex as separation criterion.
Definition GiniIndex.h:63
Virtual base Class for all MVA method.
Definition MethodBase.h:111
J Friedman's RuleFit method.
Double_t GetMvaValue(Double_t *err=0, Double_t *errUpper=0)
returns MVA value for given event
void MakeClassSpecific(std::ostream &, const TString &) const
write specific classifier response
void MakeClassLinear(std::ostream &) const
print out the linear terms
void GetHelpMessage() const
get help message text
void TrainJFRuleFit()
training of rules using Jerome Friedmans implementation
virtual Bool_t HasAnalysisType(Types::EAnalysisType type, UInt_t numberClasses, UInt_t)
RuleFit can handle classification with 2 classes.
void ProcessOptions()
process the options specified by the user
void ReadWeightsFromStream(std::istream &istr)
read rules from an std::istream
void AddWeightsXMLTo(void *parent) const
add the rules to XML node
void InitEventSample(void)
write all Events from the Tree into a vector of Events, that are more easily manipulated.
void MakeClassRuleCuts(std::ostream &) const
print out the rule cuts
void InitMonitorNtuple()
initialize the monitoring ntuple
virtual ~MethodRuleFit(void)
destructor
void Init(void)
default initialization
void WriteMonitoringHistosToFile(void) const
write special monitoring histograms to file (here ntuple)
MethodRuleFit(const TString &jobName, const TString &methodTitle, DataSetInfo &theData, const TString &theOption="")
standard constructor
void ReadWeightsFromXML(void *wghtnode)
read rules from XML node
void DeclareOptions()
define the options (their key words) that can be set in the option string know options.
const Ranking * CreateRanking()
computes ranking of input variables
void TrainTMVARuleFit()
training of rules using TMVA implementation
Implementation of the MisClassificationError as separation criterion.
Ranking for variables in method (implementation)
Definition Ranking.h:48
A class describing a 'rule cut'.
Definition RuleCut.h:36
UInt_t GetNvars() const
Definition RuleCut.h:72
Double_t GetCutMin(Int_t is) const
Definition RuleCut.h:74
UInt_t GetSelector(Int_t is) const
Definition RuleCut.h:73
Char_t GetCutDoMin(Int_t is) const
Definition RuleCut.h:76
Char_t GetCutDoMax(Int_t is) const
Definition RuleCut.h:77
UInt_t GetNcuts() const
get number of cuts
Definition RuleCut.cxx:164
Double_t GetCutMax(Int_t is) const
Definition RuleCut.h:75
Double_t GetLinDP(int i) const
Double_t GetLinDM(int i) const
const std::vector< Double_t > & GetLinCoefficients() const
Double_t GetImportanceRef() const
const std::vector< Double_t > & GetLinNorm() const
UInt_t GetNLinear() const
const std::vector< TMVA::Rule * > & GetRulesConst() const
const std::vector< Double_t > & GetLinImportance() const
Bool_t IsLinTermOK(int i) const
J Friedman's RuleFit method.
Definition RuleFitAPI.h:51
Bool_t ReadModelSum()
read model from rulefit.sum
void WelcomeMessage()
welcome message
Implementation of a rule.
Definition Rule.h:50
Double_t GetSupport() const
Definition Rule.h:142
const RuleCut * GetRuleCut() const
Definition Rule.h:139
Bool_t IsSignalRule() const
Definition Rule.h:119
Double_t GetCoefficient() const
Definition Rule.h:141
const RuleEnsemble * GetRuleEnsemble() const
Definition Rule.h:140
Double_t GetSSB() const
Definition Rule.h:117
Double_t GetRelImportance() const
Definition Rule.h:102
Implementation of the SdivSqrtSplusB as separation criterion.
Timing information for training and evaluation of MVA methods.
Definition Timer.h:58
const TString & Color(const TString &)
human readable color strings
Definition Tools.cxx:828
Singleton class for Global types used by TMVA.
Definition Types.h:71
@ kClassification
Definition Types.h:127
@ kTraining
Definition Types.h:143
virtual Int_t Write(const char *name=0, Int_t option=0, Int_t bufsize=0)
Write this object to the current directory.
Definition TObject.cxx:868
virtual void Print(Option_t *option="") const
This method must be overridden when a class wants to print itself.
Definition TObject.cxx:622
Basic string class.
Definition TString.h:136
A TTree represents a columnar dataset.
Definition TTree.h:79
const Int_t n
Definition legend1.C:16
create variable transformations
Config & gConfig()
Tools & gTools()
MsgLogger & Endl(MsgLogger &ml)
Definition MsgLogger.h:148