SlideShare a Scribd company logo
1 of 17
Download to read offline
Selective	Gradient	Boosting	
for	Effective	Learning	to	Rank
Claudio	Lucchese1
,	Franco	Maria	Nardini2
,	Raffaele	Perego2
,	
Salvatore	Orlando1,	Salvatore	Trani2
1 Ca’	Foscari University of	Venice,	Italy
2 ISTI-CNR,	Pisa,	Italy
l a b o r a t o r y
2018SIGIR
Scenario
• Learning	to	Rank	(LtR)	for	modern	IR	systems
• Supervised	approach	exploiting	graded-relevance	labels
• Multi-Stage	(2+)	IR	system
• First	stage	optimizes	Recall
• Following	stage(s)	deal	with	Ranking
• First	stage	of	production	systems	
outputs	thousands	of	documents
• Only	a	small	fraction	of	them	are	relevant	[1]
• Ranking	models	should	be	trained	accordingly
Credit:	“Efficient	 and	Effective	Retrieval	using	Selective	Pruning”
Tonellotto,	N.,	Macdonald,	 C.,	&	Ounis,	I.	ACM	WSDM	2013
[1] Yin,	D.	et	al.	Ranking	Relevance	 in	Yahoo	Search.	In ACM	SIGKDD	2016.
Problem
• With	a	high	number	of	negative	instances
• Higher	class	unbalance
• Better	coverage	of	negative	cases
• Research	Questions:
• Does	the	volume	of	negative	instances	impacts	the	performance	of	state-of-
the-art	LtR algorithms?
• Is	it	possible	to	improve	the	effectiveness	of	the	learned	model	by	exploiting	
the	diversity	and	variety	of	the	negative	instances?
Intuition
• Key	idea:	sampling	of	negative	instances	at	each	iterationof a	
gradient	boosting	process
• State-of-the-art	algorithms	learn	a	forest	of	regression	trees	
incrementally	through	gradient	boosting
• λ-MART	[1]	does	not	use	sampling
• Stochastic	Gradient	Boosting	[2]:	uniform	sampling
• Ground-truth	labels	are	not	taken	into	account
[1]	Christopher	JC	Burges.	2010.	From	ranknet to	lambdarank to	lambdamart:	 An	overview.	Learning	11,	23-581 (2010)
[2]	J.	H.	Friedman.	 2002.	Stochastic	gradient	boosting.	Computational	Statistics	&	Data	Analysis	38,	4 (2002)
• Smart	sampling	of	the	most	“useful”
instances	on	a	per-query	basis
• All	the	positive	instances
• A	fraction	of	the	top-ranked	negative	instances
• Rationale:
• The	negative	instances	with	the	highest	scores	are	exactly	those	being	more	likely	to	
be	ranked	above	relevant	instances,	thus	severely	hindering	model	effectiveness
• Selected	instance`s	are	close	to	the	decision	boundary	between	positives	and	
negatives
Proposal
Relevant	document
Irrelevant	document
Ranked	List
Selective	Gradient	Boosting	(SelGB)
• Identical	to	λ-MART	except	for	
the	red	part
• Boosting	approach
• λ-gradients	for	optimizing	list-wise	
metric	(e.g.,	NDCG)
• Selective	Sampling
• Repeated	every	n iterations
• Samples	p%	of	top	negative	
instances	plus	all	positives
• Selection	depends	on	the	model	
learned	so	far
Experimental Settings
• LtR algorithmextended:	λ-MART
• Metric:	NDCG@10
• Software:	QuickRank1
• Ensembleswith:
• Up	to	1,000	trees
• Each tree with	up	to	64	leaves
• Hyper-parameterstested:
• Sampling rate	in	{0.1%,	0.25%,	0,5%,	1%,	2.5%,	5%,	10%,	25%}
• Sampling frequency in	{1,	10,	50,	100}
1
http://quickrank.isti.cnr.it/
Dataset
• Novel	Istella-X	(eXtended)	dataset1:
• Up	to	5k	query/doc	pairs	ranked	according	to	BM25F
• Highly	unbalanced:
• Only	4.64	relevant	documents	per	query	on	average
1
Publicly	available	 at	http://quickrank.isti.cnr.it/istella-dataset/
Table 1: Datasets properties.
Properties IX5k IX2.5k IX1k IX500 IX100
# queries 10,000
# features 220
# query-doc pairs 26,791,447 15,778,399 7,363,902 3,995,063 906,782
max # docs/query 5,000 2,500 1,000 500 100
avg. # docs/query 2,679 1,578 736 400 91
# pos. query-doc pairs 46,371 (0.17%) 46,371 (0.29%) 46,371 (0.63%) 46,371 (1.16%) 46,371 (5.11%)
# neg. query-doc pairs 26,745,076 (99.83%) 15,732,028 (99.71%) 7,317,531 (99.37%) 3,948,692 (98,84%) 860,411 (94,89%)
results show that, by exploiting a large number of negative ex-
amples, SGB is able to build ranking models that result to be
more accurate than those learned with state-of-the-art ranking
algorithms.
This section is organized as follow. First we introduce the method-
ology used for the experimental evaluation, and the challenging
To investigate to which extent the presence of negative examples
may inuence the performance of LtR algorithms, we also created
three scaled-down variants of IX5k. For each query in the
training set, we rst sorted all negative examples in descending
order of BM25F scores, and then we discarded everything beyond a
given rank. We used this methodology to produce three datasets,
2.5k 1k 500
λ-MART
• Best	effectiveness	achieved	on	
Istella-X500
• λ-MART	is	not	able	to	exploit	the	
increased	number	of	negative	
instances
Stochastic	Gradient	Boosting
• Many	relevant	documents	are	
discarded	when	the	dataset	is	
highly	unbalanced
• Sampling	out	positive	instances	
severely	hurts	effectiveness
Negative	Stochastic	Gradient	Boosting
• SGB	performing	the	sampling	
only	on	the	irrelevant	docs
• Still	uniformly	at	random
• Effectiveness	is	not	improved
Selective	Gradient	Boosting
• Settings:
• Sampling	at	every	iteration
• Several	sampling	rates
• Gain	of	0,024	in	NDCG@10	
• SelGB is	able	of	exploiting	diversity	
and	variety	of	irrelevant	
documents
• Best	sampling	rate:	1%
• Improved	efficiency	of	training
Selective	Gradient	Boosting
• Settings:
• Fixed	sampling	rate:	1%
• Several	sampling	frequencies
• Best	performance	achieved	with	
frequent	re-sampling
• Best	sampling	frequency:	1
Summary	of	performance
• Comparison	against	λ-MART:
• +3.2%	in	terms	of	NDCG@10
• Similar	effectiveness	achieved	
with	less	than	15%	of	the	trees
• Higher	scoring	efficiency
Spatial	usage	of	negative	instances
• Analysis	on	which	documents	
have	been	selected	by	SelGB
• y-axis:	queries	sorted	by	list	size
• x-axis:	documents	ranked	by	BM25F
• color:	selection	frequency
• Outcomes:
• Most	selected	negative	instances	
have	high	BM25F	rank
• Especially	for	queries	with	longer	
list,	many	selected	negative	
instances	have	low	BM25F	rank
• Those	documents	are	important!
Conclusions and	Future	Work
• We	proposed	a	novel	LtR algorithm:
• Selectively	choose	the	most	informative	negative	instances
• Mitigate	dataset	class	unbalance
• Exploit	diversity	and	variety	of	negative	instances
• Outperform	λ-MART	both	in	of	effectiveness	and	efficiency
• Future	directions:
• Different	sampling	rates	on	a	per	query	basis
• Novel	selection	strategy
• Evaluate	improvement	in	terms	of	efficiency
17
Salvatore	Trani
salvatore.trani@isti.cnr.it

More Related Content

Similar to Selective Gradient Boosting for Effective Learning to Rank - SIGIR 2018

Similar to Selective Gradient Boosting for Effective Learning to Rank - SIGIR 2018 (20)

Nbvtalkonfeatureselection
NbvtalkonfeatureselectionNbvtalkonfeatureselection
Nbvtalkonfeatureselection
 
moraes-a2017ictir
moraes-a2017ictirmoraes-a2017ictir
moraes-a2017ictir
 
2014 IEEE JAVA DATA MINING PROJECT A similarity measure for text classificati...
2014 IEEE JAVA DATA MINING PROJECT A similarity measure for text classificati...2014 IEEE JAVA DATA MINING PROJECT A similarity measure for text classificati...
2014 IEEE JAVA DATA MINING PROJECT A similarity measure for text classificati...
 
IEEE 2014 JAVA DATA MINING PROJECTS A similarity measure for text classificat...
IEEE 2014 JAVA DATA MINING PROJECTS A similarity measure for text classificat...IEEE 2014 JAVA DATA MINING PROJECTS A similarity measure for text classificat...
IEEE 2014 JAVA DATA MINING PROJECTS A similarity measure for text classificat...
 
seminar.pptx
seminar.pptxseminar.pptx
seminar.pptx
 
A Combination of Simple Models by Forward Predictor Selection for Job Recomme...
A Combination of Simple Models by Forward Predictor Selection for Job Recomme...A Combination of Simple Models by Forward Predictor Selection for Job Recomme...
A Combination of Simple Models by Forward Predictor Selection for Job Recomme...
 
a deep reinforced model for abstractive summarization
a deep reinforced model for abstractive summarizationa deep reinforced model for abstractive summarization
a deep reinforced model for abstractive summarization
 
PR-231: A Simple Framework for Contrastive Learning of Visual Representations
PR-231: A Simple Framework for Contrastive Learning of Visual RepresentationsPR-231: A Simple Framework for Contrastive Learning of Visual Representations
PR-231: A Simple Framework for Contrastive Learning of Visual Representations
 
Network Based Intrusion Detection System using Filter Based Feature Selection...
Network Based Intrusion Detection System using Filter Based Feature Selection...Network Based Intrusion Detection System using Filter Based Feature Selection...
Network Based Intrusion Detection System using Filter Based Feature Selection...
 
Handling Missing Attributes using Matrix Factorization 
Handling Missing Attributes using Matrix Factorization Handling Missing Attributes using Matrix Factorization 
Handling Missing Attributes using Matrix Factorization 
 
Multi-method Evaluation in Scientific Paper Recommender Systems
Multi-method Evaluation in Scientific Paper Recommender SystemsMulti-method Evaluation in Scientific Paper Recommender Systems
Multi-method Evaluation in Scientific Paper Recommender Systems
 
Feature selection for classification
Feature selection for classificationFeature selection for classification
Feature selection for classification
 
Search quality in practice
Search quality in practiceSearch quality in practice
Search quality in practice
 
CrowdInG_learning_from_crowds.pptx
CrowdInG_learning_from_crowds.pptxCrowdInG_learning_from_crowds.pptx
CrowdInG_learning_from_crowds.pptx
 
Ijcatr04051005
Ijcatr04051005Ijcatr04051005
Ijcatr04051005
 
Recommendation engine Using Genetic Algorithm
Recommendation engine Using Genetic AlgorithmRecommendation engine Using Genetic Algorithm
Recommendation engine Using Genetic Algorithm
 
crossvalidation.pptx
crossvalidation.pptxcrossvalidation.pptx
crossvalidation.pptx
 
Optimization as a model for few shot learning
Optimization as a model for few shot learningOptimization as a model for few shot learning
Optimization as a model for few shot learning
 
Scalable Software Testing and Verification of Non-Functional Properties throu...
Scalable Software Testing and Verification of Non-Functional Properties throu...Scalable Software Testing and Verification of Non-Functional Properties throu...
Scalable Software Testing and Verification of Non-Functional Properties throu...
 
Regularization in deep learning
Regularization in deep learningRegularization in deep learning
Regularization in deep learning
 

Recently uploaded

Chiulli_Aurora_Oman_Raffaele_Beowulf.pptx
Chiulli_Aurora_Oman_Raffaele_Beowulf.pptxChiulli_Aurora_Oman_Raffaele_Beowulf.pptx
Chiulli_Aurora_Oman_Raffaele_Beowulf.pptx
raffaeleoman
 
Proofreading- Basics to Artificial Intelligence Integration - Presentation:Sl...
Proofreading- Basics to Artificial Intelligence Integration - Presentation:Sl...Proofreading- Basics to Artificial Intelligence Integration - Presentation:Sl...
Proofreading- Basics to Artificial Intelligence Integration - Presentation:Sl...
David Celestin
 
Uncommon Grace The Autobiography of Isaac Folorunso
Uncommon Grace The Autobiography of Isaac FolorunsoUncommon Grace The Autobiography of Isaac Folorunso
Uncommon Grace The Autobiography of Isaac Folorunso
Kayode Fayemi
 
Bring back lost lover in USA, Canada ,Uk ,Australia ,London Lost Love Spell C...
Bring back lost lover in USA, Canada ,Uk ,Australia ,London Lost Love Spell C...Bring back lost lover in USA, Canada ,Uk ,Australia ,London Lost Love Spell C...
Bring back lost lover in USA, Canada ,Uk ,Australia ,London Lost Love Spell C...
amilabibi1
 
If this Giant Must Walk: A Manifesto for a New Nigeria
If this Giant Must Walk: A Manifesto for a New NigeriaIf this Giant Must Walk: A Manifesto for a New Nigeria
If this Giant Must Walk: A Manifesto for a New Nigeria
Kayode Fayemi
 

Recently uploaded (15)

Chiulli_Aurora_Oman_Raffaele_Beowulf.pptx
Chiulli_Aurora_Oman_Raffaele_Beowulf.pptxChiulli_Aurora_Oman_Raffaele_Beowulf.pptx
Chiulli_Aurora_Oman_Raffaele_Beowulf.pptx
 
Report Writing Webinar Training
Report Writing Webinar TrainingReport Writing Webinar Training
Report Writing Webinar Training
 
SOLID WASTE MANAGEMENT SYSTEM OF FENI PAURASHAVA, BANGLADESH.pdf
SOLID WASTE MANAGEMENT SYSTEM OF FENI PAURASHAVA, BANGLADESH.pdfSOLID WASTE MANAGEMENT SYSTEM OF FENI PAURASHAVA, BANGLADESH.pdf
SOLID WASTE MANAGEMENT SYSTEM OF FENI PAURASHAVA, BANGLADESH.pdf
 
Proofreading- Basics to Artificial Intelligence Integration - Presentation:Sl...
Proofreading- Basics to Artificial Intelligence Integration - Presentation:Sl...Proofreading- Basics to Artificial Intelligence Integration - Presentation:Sl...
Proofreading- Basics to Artificial Intelligence Integration - Presentation:Sl...
 
Digital collaboration with Microsoft 365 as extension of Drupal
Digital collaboration with Microsoft 365 as extension of DrupalDigital collaboration with Microsoft 365 as extension of Drupal
Digital collaboration with Microsoft 365 as extension of Drupal
 
AWS Data Engineer Associate (DEA-C01) Exam Dumps 2024.pdf
AWS Data Engineer Associate (DEA-C01) Exam Dumps 2024.pdfAWS Data Engineer Associate (DEA-C01) Exam Dumps 2024.pdf
AWS Data Engineer Associate (DEA-C01) Exam Dumps 2024.pdf
 
ICT role in 21st century education and it's challenges.pdf
ICT role in 21st century education and it's challenges.pdfICT role in 21st century education and it's challenges.pdf
ICT role in 21st century education and it's challenges.pdf
 
The workplace ecosystem of the future 24.4.2024 Fabritius_share ii.pdf
The workplace ecosystem of the future 24.4.2024 Fabritius_share ii.pdfThe workplace ecosystem of the future 24.4.2024 Fabritius_share ii.pdf
The workplace ecosystem of the future 24.4.2024 Fabritius_share ii.pdf
 
Uncommon Grace The Autobiography of Isaac Folorunso
Uncommon Grace The Autobiography of Isaac FolorunsoUncommon Grace The Autobiography of Isaac Folorunso
Uncommon Grace The Autobiography of Isaac Folorunso
 
Bring back lost lover in USA, Canada ,Uk ,Australia ,London Lost Love Spell C...
Bring back lost lover in USA, Canada ,Uk ,Australia ,London Lost Love Spell C...Bring back lost lover in USA, Canada ,Uk ,Australia ,London Lost Love Spell C...
Bring back lost lover in USA, Canada ,Uk ,Australia ,London Lost Love Spell C...
 
If this Giant Must Walk: A Manifesto for a New Nigeria
If this Giant Must Walk: A Manifesto for a New NigeriaIf this Giant Must Walk: A Manifesto for a New Nigeria
If this Giant Must Walk: A Manifesto for a New Nigeria
 
Dreaming Marissa Sánchez Music Video Treatment
Dreaming Marissa Sánchez Music Video TreatmentDreaming Marissa Sánchez Music Video Treatment
Dreaming Marissa Sánchez Music Video Treatment
 
lONG QUESTION ANSWER PAKISTAN STUDIES10.
lONG QUESTION ANSWER PAKISTAN STUDIES10.lONG QUESTION ANSWER PAKISTAN STUDIES10.
lONG QUESTION ANSWER PAKISTAN STUDIES10.
 
Dreaming Music Video Treatment _ Project & Portfolio III
Dreaming Music Video Treatment _ Project & Portfolio IIIDreaming Music Video Treatment _ Project & Portfolio III
Dreaming Music Video Treatment _ Project & Portfolio III
 
My Presentation "In Your Hands" by Halle Bailey
My Presentation "In Your Hands" by Halle BaileyMy Presentation "In Your Hands" by Halle Bailey
My Presentation "In Your Hands" by Halle Bailey
 

Selective Gradient Boosting for Effective Learning to Rank - SIGIR 2018