SlideShare ist ein Scribd-Unternehmen logo
1 von 29
Downloaden Sie, um offline zu lesen
Gaussian	
  Ranking	
  	
  
by	
  	
  
Matrix	
  Factoriza5on	
  
Harald	
  	
  Steck	
  
hsteck@netflix.com
RecSys	
  	
  2015	
  
Overview	
  
•  Matrix	
  Factoriza<on	
  Model	
  
•  asymmetric	
  MF	
  
	
  
	
  
•  Objec<ve:	
  op<mize	
  various	
  Ranking	
  Metrics	
  
•  	
  exploit	
  proper<es	
  of	
  MF	
  model	
  &	
  implicit	
  data	
  
•  Training:	
  pointwise	
  &	
  listwise	
  
•  Related	
  Work	
  
•  Experiments	
  
 	
  	
  	
  	
  	
  Basic	
  Idea:	
  
	
  	
  
	
  
	
  
	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  data	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  .	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
items	
  
i	
  
users	
  	
  u	
  
≈
	
  users	
  	
  u	
  
Low-­‐rank	
  Matrix	
  Factoriza<on	
  Model	
  
Basic	
  Idea:	
  
	
  	
  
	
  
-­‐	
  	
  latent	
  user	
  vector:	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
-­‐	
  by	
  [Paterek	
  07],	
  extended	
  to	
  SVD++	
  [Koren	
  08]	
  
Asymmetric	
  Matrix	
  Factoriza<on	
  
Overview	
  
•  Matrix	
  Factoriza<on	
  Model	
  
•  asymmetric	
  MF	
  
	
  
	
  
•  Objec5ve:	
  op5mize	
  various	
  Ranking	
  Metrics	
  
•  	
  exploit	
  proper5es	
  of	
  MF	
  model	
  &	
  implicit	
  data	
  
•  Training:	
  pointwise	
  &	
  listwise	
  
•  Related	
  Work	
  
•  Experiments	
  
AMF	
  as	
  Neural	
  Network	
  	
  
	
  	
  
	
  
	
  
	
  
	
  
	
  
	
  	
  	
  	
  	
  	
  	
  rank	
  loss	
  	
  =	
  f	
  (ranks)	
  
items	
  	
  i	
  
…	
  click	
  history	
  
…	
  user	
  vec.	
  
…	
  scores	
  
…	
  ranks	
  
AMF	
  as	
  Neural	
  Network	
  	
  
	
  	
  
	
  
	
  
	
  
	
  
	
  
	
  	
  	
  	
  	
  	
  	
  rank	
  loss	
  	
  =	
  f	
  (ranks)	
  
items	
  	
  i	
  
…	
  click	
  history	
  
…	
  user	
  vec.	
  
…	
  scores	
  
…	
  ranks	
  
1st	
  	
  term:	
  Rank	
  Loss	
  	
  
example	
  1:	
  	
  	
  AUC	
  
	
  
	
  
	
  
	
  
•  pairwise	
  comparisons	
  !	
  (linear)	
  sum	
  of	
  ranks	
  
	
  
example	
  2:	
  	
  nDCG	
  (for	
  binary	
  relevance)	
  
	
  
	
  
	
  
	
  
•  emphasizes	
  top	
  of	
  ranked	
  list	
  
•  also	
  a	
  func<on	
  of	
  the	
  ranks	
  of	
  the	
  posi<ves	
  
1st	
  	
  term:	
  Rank	
  Loss	
  	
  
2nd	
  term:	
  Ac<va<on	
  Func<on	
  
T	
  
Scores	
  !	
  Ranks:	
  	
  
	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  +	
  +	
  +	
  	
  	
  	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  -­‐	
  binary	
  data:	
  	
  nega<ves	
  and	
  posi<ves	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  -­‐	
  sparse	
  data:	
  	
  	
  	
  	
  many	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  few	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  !	
  	
  	
  MF	
  scores:	
  	
  Gaussian	
  distrib.	
  assumed	
  
scores	
  i	
  
 
	
  
	
  
score	
  
rank	
  
1	
  
N	
  
Scores	
  !	
  Ranks:	
  
2nd	
  term:	
  Ac<va<on	
  Func<on	
  
score	
  
 
	
  
	
  
score	
  
…	
  piecewise	
  
	
  	
  	
  	
  quadra<c	
  
2nd	
  term:	
  Ac<va<on	
  Func<on	
  
3rd	
  term	
  	
  
	
  
•  score:	
  
•  deriva<ve:	
  
Pueng	
  it	
  All	
  Together	
  
training	
  objec<ve	
  func<on:	
  
	
  
	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  rank	
  	
  	
  	
  	
  	
  prior	
  on	
  param’s	
  	
  	
  	
  	
  	
  scores	
  of	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  loss	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  "	
  	
  lambda	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  nega<ves	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  "gamma	
  
-­‐	
  minimized	
  by	
  stochas<c	
  gradient	
  descent	
  	
  	
  	
  	
  	
  	
  	
  
	
  
Overview	
  
•  Matrix	
  Factoriza<on	
  Model	
  
•  asymmetric	
  MF	
  
	
  
•  Objec<ve:	
  op<mize	
  various	
  Ranking	
  Metrics	
  
•  	
  exploit	
  proper<es	
  of	
  MF	
  model	
  &	
  data	
  
•  Training:	
  pointwise	
  &	
  listwise	
  
•  Related	
  Work	
  
•  Experiments	
  
Listwise	
  Approach	
  
•  consider	
  ALL	
  items	
  for	
  each	
  user:	
  
-­‐  es<mate	
  standard	
  devia<on	
  of	
  scores	
  for	
  	
  
each	
  user	
  !	
  width	
  of	
  ac<va<on	
  func<on	
  
Listwise	
  Approach	
  
•  consider	
  ALL	
  items	
  for	
  each	
  user:	
  
	
  
	
  
	
  	
  	
  -­‐	
  sort	
  by	
  scores	
  !	
  exact	
  ranks	
  
	
  	
  	
  -­‐	
  using	
  logis<c	
  ac<va<on	
  func<on:	
  
	
  
	
  
	
  	
  	
  	
  	
  	
  2nd	
  term	
  in	
  chain	
  rule	
  
AUC	
  
nDCG	
  
Listwise	
  Approach	
  
	
  	
  deriva5ves	
  L’:	
  
	
  
	
  
	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  1st	
  	
  &	
  2nd	
  	
  terms	
  
	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  top	
  of	
  	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  ranked	
  list	
  	
  
 !	
  between	
  nDCG	
  and	
  AUC:	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  L’	
  =	
  constant	
  	
  
	
  !	
  	
  use	
  very	
  large	
  std.	
  
	
  	
  	
  	
  	
  	
  	
  for	
  ac<va<on	
  func<on	
  
	
  	
  	
  	
  	
  	
  	
  in	
  pointwise	
  approach	
  
AUC	
  
nDCG	
  
Pointwise	
  Approach	
  
	
  	
  deriva5ves	
  L’:	
  	
  
	
  
	
  
	
  
	
  
	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  top	
  of	
  	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  ranked	
  list	
  	
  
	
  
Overview	
  
•  Matrix	
  Factoriza<on	
  Model	
  
•  asymmetric	
  MF	
  	
  
	
  
	
  
•  Objec<ve:	
  op<mize	
  various	
  Ranking	
  Metrics	
  
•  	
  exploit	
  proper<es	
  of	
  MF	
  model	
  &	
  data	
  
•  Training	
  
•  Related	
  Work	
  
•  Experiments	
  
Related	
  Work	
  
•  various	
  learning-­‐to-­‐rank	
  approaches	
  exist	
  
•  ogen	
  tailored	
  to	
  specific	
  ranking	
  losses	
  
•  mostly	
  pairwise	
  approaches,	
  eg:	
  
•  AUC:	
  	
  BPR	
  [Rendle	
  et	
  al.	
  ’09]	
  
•  MRR:	
  	
  CLiMF	
  [Shi	
  et	
  al.	
  ’12]	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  used	
  as	
  
•  MAP:	
  TFMAP	
  [Shi	
  et	
  al.	
  ‘12]	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  baselines	
  
	
  
•  listwise	
  approaches,	
  eg:	
  
•  	
  	
  	
  	
  top-­‐1	
  [Shi	
  et	
  al.	
  ’10]	
  ...	
  like	
  neural	
  network	
  
•  …	
  addi<onal	
  references	
  in	
  the	
  paper	
  
	
  
Overview	
  
•  Matrix	
  Factoriza<on	
  Model	
  
•  basic	
  MF	
  !	
  asymmetric	
  MF	
  !	
  Neural	
  Network	
  
	
  
	
  
•  Objec<ve:	
  op<mize	
  various	
  Ranking	
  Metrics	
  
•  	
  exploit	
  proper<es	
  of	
  MF	
  model	
  &	
  data	
  
•  Training	
  
•  Related	
  Work	
  
•  Experiments	
  
10	
  m	
  MovieLens	
  	
  Data	
  
•  10k	
  movies	
  	
  &	
  	
  70k	
  users	
  
•  1%	
  dense	
  data	
  
•  binarized:	
  	
  	
  	
  	
  3+	
  star	
  ra<ng	
  !	
  1,	
  otherwise	
  0	
  
•  5-­‐fold	
  cross-­‐valida<on	
  
10	
  m	
  MovieLens	
  	
  Data	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  5-­‐fold	
  cross-­‐valida<on	
  	
  	
  	
  	
  std	
  	
  :	
  	
  	
  0.001	
  
	
  
10	
  m	
  MovieLens	
  	
  Data	
  
	
  	
  	
  	
  std=0.002	
  
Nellix	
  Play	
  Data	
  
•  Test	
  day:	
  	
  
	
  	
  	
  	
  4/9/2014	
  
	
  
•  rela(ve	
  	
  
	
  	
  	
  improvement	
  	
  
	
  	
  	
  to	
  RMSE	
  training	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  std=1%	
  
Nellix	
  Play	
  Data	
  
	
  
	
  
	
  
	
  
	
  
	
  
	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  std=2%	
  
Conclusions	
  
•  learning-­‐to-­‐rank	
  approach:	
  
– implicit	
  feedback	
  data	
  
– proper<es	
  of	
  MF	
  model	
  
! Gaussian	
  distribu<on	
  of	
  scores	
  
! non-­‐linear	
  ac<va<on	
  func<ons	
  derived	
  for	
  ranking	
  
•  pointwise	
  and	
  listwise	
  training	
  
•  various	
  ranking	
  metrics	
  can	
  be	
  used:	
  
– compe<<ve	
  for	
  op<mizing	
  AUC	
  
– par<cularly	
  effec<ve	
  at	
  head	
  of	
  ranked	
  list	
  
Thank	
  You	
  !	
  
Ques5ons	
  ?	
  

Weitere ähnliche Inhalte

Andere mochten auch

Participatory rural appraisal
Participatory rural appraisalParticipatory rural appraisal
Participatory rural appraisalDeekshit Geddam
 
Presentation on PRA & PLA Process
Presentation on PRA & PLA ProcessPresentation on PRA & PLA Process
Presentation on PRA & PLA ProcessGULZAR HUSSAIN
 
Interactive Recommender Systems with Netflix and Spotify
Interactive Recommender Systems with Netflix and SpotifyInteractive Recommender Systems with Netflix and Spotify
Interactive Recommender Systems with Netflix and SpotifyChris Johnson
 

Andere mochten auch (6)

Participatory rural appraisal
Participatory rural appraisalParticipatory rural appraisal
Participatory rural appraisal
 
Pra methods
Pra methodsPra methods
Pra methods
 
Final ppt
Final pptFinal ppt
Final ppt
 
Pra presentation
Pra presentationPra presentation
Pra presentation
 
Presentation on PRA & PLA Process
Presentation on PRA & PLA ProcessPresentation on PRA & PLA Process
Presentation on PRA & PLA Process
 
Interactive Recommender Systems with Netflix and Spotify
Interactive Recommender Systems with Netflix and SpotifyInteractive Recommender Systems with Netflix and Spotify
Interactive Recommender Systems with Netflix and Spotify
 

Ähnlich wie Gaussian Ranking by Matrix Factorization, ACM RecSys Conference 2015

Apache Spark Machine Learning
Apache Spark Machine LearningApache Spark Machine Learning
Apache Spark Machine LearningPraveen Devarao
 
Concept Location using Information Retrieval and Relevance Feedback
Concept Location using Information Retrieval and Relevance FeedbackConcept Location using Information Retrieval and Relevance Feedback
Concept Location using Information Retrieval and Relevance FeedbackSonia Haiduc
 
Irrf Presentation
Irrf PresentationIrrf Presentation
Irrf Presentationgregoryg
 
General Tips for participating Kaggle Competitions
General Tips for participating Kaggle CompetitionsGeneral Tips for participating Kaggle Competitions
General Tips for participating Kaggle CompetitionsMark Peng
 
Week 12 Dimensionality Reduction Bagian 1
Week 12 Dimensionality Reduction Bagian 1Week 12 Dimensionality Reduction Bagian 1
Week 12 Dimensionality Reduction Bagian 1khairulhuda242
 
Inference & Learning in Linear-Chain Conditional Random Fields (CRFs)
Inference & Learning in Linear-Chain Conditional Random Fields (CRFs)Inference & Learning in Linear-Chain Conditional Random Fields (CRFs)
Inference & Learning in Linear-Chain Conditional Random Fields (CRFs)Anmol Dwivedi
 
Predicting SPARQL query execution time and suggesting SPARQL queries based on...
Predicting SPARQL query execution time and suggesting SPARQL queries based on...Predicting SPARQL query execution time and suggesting SPARQL queries based on...
Predicting SPARQL query execution time and suggesting SPARQL queries based on...Rakebul Hasan
 
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 SystemsAravind Sesagiri Raamkumar
 
Top-k Exploration of Query Candidates for Efficient Keyword Search on Graph-S...
Top-k Exploration of Query Candidates for Efficient Keyword Search on Graph-S...Top-k Exploration of Query Candidates for Efficient Keyword Search on Graph-S...
Top-k Exploration of Query Candidates for Efficient Keyword Search on Graph-S...Thanh Tran
 
Customer Churn Analytics using Microsoft R Open
Customer Churn Analytics using Microsoft R OpenCustomer Churn Analytics using Microsoft R Open
Customer Churn Analytics using Microsoft R OpenPoo Kuan Hoong
 
LSH for
 Prediction Problem in Recommendation
LSH for
 Prediction Problem in RecommendationLSH for
 Prediction Problem in Recommendation
LSH for
 Prediction Problem in RecommendationMaruf Aytekin
 
Predicting Azure Churn with Deep Learning and Explaining Predictions with LIME
Predicting Azure Churn with Deep Learning and Explaining Predictions with LIMEPredicting Azure Churn with Deep Learning and Explaining Predictions with LIME
Predicting Azure Churn with Deep Learning and Explaining Predictions with LIMEFeng Zhu
 
Query dependent ranking using k nearest neighbor
Query dependent ranking using k nearest neighborQuery dependent ranking using k nearest neighbor
Query dependent ranking using k nearest neighboriyo
 
Continuous Evaluation of Collaborative Recommender Systems in Data Stream Man...
Continuous Evaluation of Collaborative Recommender Systems in Data Stream Man...Continuous Evaluation of Collaborative Recommender Systems in Data Stream Man...
Continuous Evaluation of Collaborative Recommender Systems in Data Stream Man...Dr. Cornelius Ludmann
 
KNN Algorithm using C++
KNN Algorithm using C++KNN Algorithm using C++
KNN Algorithm using C++Afraz Khan
 
Lecture 10 - DataMiningEngineering.ppt
Lecture 10 - DataMiningEngineering.pptLecture 10 - DataMiningEngineering.ppt
Lecture 10 - DataMiningEngineering.pptAsadkhan47384
 

Ähnlich wie Gaussian Ranking by Matrix Factorization, ACM RecSys Conference 2015 (20)

Apache Spark Machine Learning
Apache Spark Machine LearningApache Spark Machine Learning
Apache Spark Machine Learning
 
Concept Location using Information Retrieval and Relevance Feedback
Concept Location using Information Retrieval and Relevance FeedbackConcept Location using Information Retrieval and Relevance Feedback
Concept Location using Information Retrieval and Relevance Feedback
 
Irrf Presentation
Irrf PresentationIrrf Presentation
Irrf Presentation
 
General Tips for participating Kaggle Competitions
General Tips for participating Kaggle CompetitionsGeneral Tips for participating Kaggle Competitions
General Tips for participating Kaggle Competitions
 
Week 12 Dimensionality Reduction Bagian 1
Week 12 Dimensionality Reduction Bagian 1Week 12 Dimensionality Reduction Bagian 1
Week 12 Dimensionality Reduction Bagian 1
 
Inference & Learning in Linear-Chain Conditional Random Fields (CRFs)
Inference & Learning in Linear-Chain Conditional Random Fields (CRFs)Inference & Learning in Linear-Chain Conditional Random Fields (CRFs)
Inference & Learning in Linear-Chain Conditional Random Fields (CRFs)
 
Predicting SPARQL query execution time and suggesting SPARQL queries based on...
Predicting SPARQL query execution time and suggesting SPARQL queries based on...Predicting SPARQL query execution time and suggesting SPARQL queries based on...
Predicting SPARQL query execution time and suggesting SPARQL queries based on...
 
1710 track3 zhu
1710 track3 zhu1710 track3 zhu
1710 track3 zhu
 
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
 
Top-k Exploration of Query Candidates for Efficient Keyword Search on Graph-S...
Top-k Exploration of Query Candidates for Efficient Keyword Search on Graph-S...Top-k Exploration of Query Candidates for Efficient Keyword Search on Graph-S...
Top-k Exploration of Query Candidates for Efficient Keyword Search on Graph-S...
 
Customer Churn Analytics using Microsoft R Open
Customer Churn Analytics using Microsoft R OpenCustomer Churn Analytics using Microsoft R Open
Customer Churn Analytics using Microsoft R Open
 
LSH for
 Prediction Problem in Recommendation
LSH for
 Prediction Problem in RecommendationLSH for
 Prediction Problem in Recommendation
LSH for
 Prediction Problem in Recommendation
 
Predicting Azure Churn with Deep Learning and Explaining Predictions with LIME
Predicting Azure Churn with Deep Learning and Explaining Predictions with LIMEPredicting Azure Churn with Deep Learning and Explaining Predictions with LIME
Predicting Azure Churn with Deep Learning and Explaining Predictions with LIME
 
Query dependent ranking using k nearest neighbor
Query dependent ranking using k nearest neighborQuery dependent ranking using k nearest neighbor
Query dependent ranking using k nearest neighbor
 
Continuous Evaluation of Collaborative Recommender Systems in Data Stream Man...
Continuous Evaluation of Collaborative Recommender Systems in Data Stream Man...Continuous Evaluation of Collaborative Recommender Systems in Data Stream Man...
Continuous Evaluation of Collaborative Recommender Systems in Data Stream Man...
 
AlphaPy
AlphaPyAlphaPy
AlphaPy
 
MDE in Practice
MDE in PracticeMDE in Practice
MDE in Practice
 
KNN Algorithm using C++
KNN Algorithm using C++KNN Algorithm using C++
KNN Algorithm using C++
 
Search quality in practice
Search quality in practiceSearch quality in practice
Search quality in practice
 
Lecture 10 - DataMiningEngineering.ppt
Lecture 10 - DataMiningEngineering.pptLecture 10 - DataMiningEngineering.ppt
Lecture 10 - DataMiningEngineering.ppt
 

Kürzlich hochgeladen

Building on a FAIRly Strong Foundation to Connect Academic Research to Transl...
Building on a FAIRly Strong Foundation to Connect Academic Research to Transl...Building on a FAIRly Strong Foundation to Connect Academic Research to Transl...
Building on a FAIRly Strong Foundation to Connect Academic Research to Transl...Jack DiGiovanna
 
2006_GasProcessing_HB (1).pdf HYDROCARBON PROCESSING
2006_GasProcessing_HB (1).pdf HYDROCARBON PROCESSING2006_GasProcessing_HB (1).pdf HYDROCARBON PROCESSING
2006_GasProcessing_HB (1).pdf HYDROCARBON PROCESSINGmarianagonzalez07
 
From idea to production in a day – Leveraging Azure ML and Streamlit to build...
From idea to production in a day – Leveraging Azure ML and Streamlit to build...From idea to production in a day – Leveraging Azure ML and Streamlit to build...
From idea to production in a day – Leveraging Azure ML and Streamlit to build...Florian Roscheck
 
Advanced Machine Learning for Business Professionals
Advanced Machine Learning for Business ProfessionalsAdvanced Machine Learning for Business Professionals
Advanced Machine Learning for Business ProfessionalsVICTOR MAESTRE RAMIREZ
 
Heart Disease Classification Report: A Data Analysis Project
Heart Disease Classification Report: A Data Analysis ProjectHeart Disease Classification Report: A Data Analysis Project
Heart Disease Classification Report: A Data Analysis ProjectBoston Institute of Analytics
 
Call Girls In Dwarka 9654467111 Escorts Service
Call Girls In Dwarka 9654467111 Escorts ServiceCall Girls In Dwarka 9654467111 Escorts Service
Call Girls In Dwarka 9654467111 Escorts ServiceSapana Sha
 
科罗拉多大学波尔得分校毕业证学位证成绩单-可办理
科罗拉多大学波尔得分校毕业证学位证成绩单-可办理科罗拉多大学波尔得分校毕业证学位证成绩单-可办理
科罗拉多大学波尔得分校毕业证学位证成绩单-可办理e4aez8ss
 
Data Factory in Microsoft Fabric (MsBIP #82)
Data Factory in Microsoft Fabric (MsBIP #82)Data Factory in Microsoft Fabric (MsBIP #82)
Data Factory in Microsoft Fabric (MsBIP #82)Cathrine Wilhelmsen
 
Predictive Analysis for Loan Default Presentation : Data Analysis Project PPT
Predictive Analysis for Loan Default  Presentation : Data Analysis Project PPTPredictive Analysis for Loan Default  Presentation : Data Analysis Project PPT
Predictive Analysis for Loan Default Presentation : Data Analysis Project PPTBoston Institute of Analytics
 
Consent & Privacy Signals on Google *Pixels* - MeasureCamp Amsterdam 2024
Consent & Privacy Signals on Google *Pixels* - MeasureCamp Amsterdam 2024Consent & Privacy Signals on Google *Pixels* - MeasureCamp Amsterdam 2024
Consent & Privacy Signals on Google *Pixels* - MeasureCamp Amsterdam 2024thyngster
 
RABBIT: A CLI tool for identifying bots based on their GitHub events.
RABBIT: A CLI tool for identifying bots based on their GitHub events.RABBIT: A CLI tool for identifying bots based on their GitHub events.
RABBIT: A CLI tool for identifying bots based on their GitHub events.natarajan8993
 
Student Profile Sample report on improving academic performance by uniting gr...
Student Profile Sample report on improving academic performance by uniting gr...Student Profile Sample report on improving academic performance by uniting gr...
Student Profile Sample report on improving academic performance by uniting gr...Seán Kennedy
 
9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort service
9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort service9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort service
9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort servicejennyeacort
 
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...GQ Research
 
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...dajasot375
 
Generative AI for Social Good at Open Data Science East 2024
Generative AI for Social Good at Open Data Science East 2024Generative AI for Social Good at Open Data Science East 2024
Generative AI for Social Good at Open Data Science East 2024Colleen Farrelly
 
GA4 Without Cookies [Measure Camp AMS]
GA4 Without Cookies [Measure Camp AMS]GA4 Without Cookies [Measure Camp AMS]
GA4 Without Cookies [Measure Camp AMS]📊 Markus Baersch
 
RadioAdProWritingCinderellabyButleri.pdf
RadioAdProWritingCinderellabyButleri.pdfRadioAdProWritingCinderellabyButleri.pdf
RadioAdProWritingCinderellabyButleri.pdfgstagge
 
专业一比一美国俄亥俄大学毕业证成绩单pdf电子版制作修改
专业一比一美国俄亥俄大学毕业证成绩单pdf电子版制作修改专业一比一美国俄亥俄大学毕业证成绩单pdf电子版制作修改
专业一比一美国俄亥俄大学毕业证成绩单pdf电子版制作修改yuu sss
 

Kürzlich hochgeladen (20)

Building on a FAIRly Strong Foundation to Connect Academic Research to Transl...
Building on a FAIRly Strong Foundation to Connect Academic Research to Transl...Building on a FAIRly Strong Foundation to Connect Academic Research to Transl...
Building on a FAIRly Strong Foundation to Connect Academic Research to Transl...
 
2006_GasProcessing_HB (1).pdf HYDROCARBON PROCESSING
2006_GasProcessing_HB (1).pdf HYDROCARBON PROCESSING2006_GasProcessing_HB (1).pdf HYDROCARBON PROCESSING
2006_GasProcessing_HB (1).pdf HYDROCARBON PROCESSING
 
Call Girls in Saket 99530🔝 56974 Escort Service
Call Girls in Saket 99530🔝 56974 Escort ServiceCall Girls in Saket 99530🔝 56974 Escort Service
Call Girls in Saket 99530🔝 56974 Escort Service
 
From idea to production in a day – Leveraging Azure ML and Streamlit to build...
From idea to production in a day – Leveraging Azure ML and Streamlit to build...From idea to production in a day – Leveraging Azure ML and Streamlit to build...
From idea to production in a day – Leveraging Azure ML and Streamlit to build...
 
Advanced Machine Learning for Business Professionals
Advanced Machine Learning for Business ProfessionalsAdvanced Machine Learning for Business Professionals
Advanced Machine Learning for Business Professionals
 
Heart Disease Classification Report: A Data Analysis Project
Heart Disease Classification Report: A Data Analysis ProjectHeart Disease Classification Report: A Data Analysis Project
Heart Disease Classification Report: A Data Analysis Project
 
Call Girls In Dwarka 9654467111 Escorts Service
Call Girls In Dwarka 9654467111 Escorts ServiceCall Girls In Dwarka 9654467111 Escorts Service
Call Girls In Dwarka 9654467111 Escorts Service
 
科罗拉多大学波尔得分校毕业证学位证成绩单-可办理
科罗拉多大学波尔得分校毕业证学位证成绩单-可办理科罗拉多大学波尔得分校毕业证学位证成绩单-可办理
科罗拉多大学波尔得分校毕业证学位证成绩单-可办理
 
Data Factory in Microsoft Fabric (MsBIP #82)
Data Factory in Microsoft Fabric (MsBIP #82)Data Factory in Microsoft Fabric (MsBIP #82)
Data Factory in Microsoft Fabric (MsBIP #82)
 
Predictive Analysis for Loan Default Presentation : Data Analysis Project PPT
Predictive Analysis for Loan Default  Presentation : Data Analysis Project PPTPredictive Analysis for Loan Default  Presentation : Data Analysis Project PPT
Predictive Analysis for Loan Default Presentation : Data Analysis Project PPT
 
Consent & Privacy Signals on Google *Pixels* - MeasureCamp Amsterdam 2024
Consent & Privacy Signals on Google *Pixels* - MeasureCamp Amsterdam 2024Consent & Privacy Signals on Google *Pixels* - MeasureCamp Amsterdam 2024
Consent & Privacy Signals on Google *Pixels* - MeasureCamp Amsterdam 2024
 
RABBIT: A CLI tool for identifying bots based on their GitHub events.
RABBIT: A CLI tool for identifying bots based on their GitHub events.RABBIT: A CLI tool for identifying bots based on their GitHub events.
RABBIT: A CLI tool for identifying bots based on their GitHub events.
 
Student Profile Sample report on improving academic performance by uniting gr...
Student Profile Sample report on improving academic performance by uniting gr...Student Profile Sample report on improving academic performance by uniting gr...
Student Profile Sample report on improving academic performance by uniting gr...
 
9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort service
9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort service9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort service
9711147426✨Call In girls Gurgaon Sector 31. SCO 25 escort service
 
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
 
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...
Indian Call Girls in Abu Dhabi O5286O24O8 Call Girls in Abu Dhabi By Independ...
 
Generative AI for Social Good at Open Data Science East 2024
Generative AI for Social Good at Open Data Science East 2024Generative AI for Social Good at Open Data Science East 2024
Generative AI for Social Good at Open Data Science East 2024
 
GA4 Without Cookies [Measure Camp AMS]
GA4 Without Cookies [Measure Camp AMS]GA4 Without Cookies [Measure Camp AMS]
GA4 Without Cookies [Measure Camp AMS]
 
RadioAdProWritingCinderellabyButleri.pdf
RadioAdProWritingCinderellabyButleri.pdfRadioAdProWritingCinderellabyButleri.pdf
RadioAdProWritingCinderellabyButleri.pdf
 
专业一比一美国俄亥俄大学毕业证成绩单pdf电子版制作修改
专业一比一美国俄亥俄大学毕业证成绩单pdf电子版制作修改专业一比一美国俄亥俄大学毕业证成绩单pdf电子版制作修改
专业一比一美国俄亥俄大学毕业证成绩单pdf电子版制作修改
 

Gaussian Ranking by Matrix Factorization, ACM RecSys Conference 2015

  • 1. Gaussian  Ranking     by     Matrix  Factoriza5on   Harald    Steck   hsteck@netflix.com RecSys    2015  
  • 2. Overview   •  Matrix  Factoriza<on  Model   •  asymmetric  MF       •  Objec<ve:  op<mize  various  Ranking  Metrics   •   exploit  proper<es  of  MF  model  &  implicit  data   •  Training:  pointwise  &  listwise   •  Related  Work   •  Experiments  
  • 3.            Basic  Idea:                                                                      data                                                      .                                                                                                                                                                                                                                                                                                   items   i   users    u   ≈  users    u   Low-­‐rank  Matrix  Factoriza<on  Model  
  • 4. Basic  Idea:         -­‐    latent  user  vector:                                                                                                                                                                                                                                                   -­‐  by  [Paterek  07],  extended  to  SVD++  [Koren  08]   Asymmetric  Matrix  Factoriza<on  
  • 5. Overview   •  Matrix  Factoriza<on  Model   •  asymmetric  MF       •  Objec5ve:  op5mize  various  Ranking  Metrics   •   exploit  proper5es  of  MF  model  &  implicit  data   •  Training:  pointwise  &  listwise   •  Related  Work   •  Experiments  
  • 6. AMF  as  Neural  Network                                rank  loss    =  f  (ranks)   items    i   …  click  history   …  user  vec.   …  scores   …  ranks  
  • 7. AMF  as  Neural  Network                                rank  loss    =  f  (ranks)   items    i   …  click  history   …  user  vec.   …  scores   …  ranks  
  • 8. 1st    term:  Rank  Loss     example  1:      AUC           •  pairwise  comparisons  !  (linear)  sum  of  ranks    
  • 9. example  2:    nDCG  (for  binary  relevance)           •  emphasizes  top  of  ranked  list   •  also  a  func<on  of  the  ranks  of  the  posi<ves   1st    term:  Rank  Loss    
  • 10. 2nd  term:  Ac<va<on  Func<on   T   Scores  !  Ranks:                                                                                                                                +  +  +                                            -­‐  binary  data:    nega<ves  and  posi<ves                    -­‐  sparse  data:          many                                  few                    !      MF  scores:    Gaussian  distrib.  assumed   scores  i  
  • 11.       score   rank   1   N   Scores  !  Ranks:   2nd  term:  Ac<va<on  Func<on   score  
  • 12.       score   …  piecewise          quadra<c   2nd  term:  Ac<va<on  Func<on  
  • 13. 3rd  term       •  score:   •  deriva<ve:  
  • 14. Pueng  it  All  Together   training  objec<ve  func<on:                                      rank            prior  on  param’s            scores  of                                  loss                      "    lambda                        nega<ves                                                                                                                                      "gamma   -­‐  minimized  by  stochas<c  gradient  descent                  
  • 15. Overview   •  Matrix  Factoriza<on  Model   •  asymmetric  MF     •  Objec<ve:  op<mize  various  Ranking  Metrics   •   exploit  proper<es  of  MF  model  &  data   •  Training:  pointwise  &  listwise   •  Related  Work   •  Experiments  
  • 16. Listwise  Approach   •  consider  ALL  items  for  each  user:   -­‐  es<mate  standard  devia<on  of  scores  for     each  user  !  width  of  ac<va<on  func<on  
  • 17. Listwise  Approach   •  consider  ALL  items  for  each  user:            -­‐  sort  by  scores  !  exact  ranks        -­‐  using  logis<c  ac<va<on  func<on:                  2nd  term  in  chain  rule  
  • 18. AUC   nDCG   Listwise  Approach      deriva5ves  L’:                                                      1st    &  2nd    terms                                                                                                                                                    top  of                                                                                                                                            ranked  list    
  • 19.  !  between  nDCG  and  AUC:                            L’  =  constant      !    use  very  large  std.                for  ac<va<on  func<on                in  pointwise  approach   AUC   nDCG   Pointwise  Approach      deriva5ves  L’:                                                                                                                                                              top  of                                                                                                                                            ranked  list      
  • 20. Overview   •  Matrix  Factoriza<on  Model   •  asymmetric  MF         •  Objec<ve:  op<mize  various  Ranking  Metrics   •   exploit  proper<es  of  MF  model  &  data   •  Training   •  Related  Work   •  Experiments  
  • 21. Related  Work   •  various  learning-­‐to-­‐rank  approaches  exist   •  ogen  tailored  to  specific  ranking  losses   •  mostly  pairwise  approaches,  eg:   •  AUC:    BPR  [Rendle  et  al.  ’09]   •  MRR:    CLiMF  [Shi  et  al.  ’12]                                  used  as   •  MAP:  TFMAP  [Shi  et  al.  ‘12]                                baselines     •  listwise  approaches,  eg:   •         top-­‐1  [Shi  et  al.  ’10]  ...  like  neural  network   •  …  addi<onal  references  in  the  paper    
  • 22. Overview   •  Matrix  Factoriza<on  Model   •  basic  MF  !  asymmetric  MF  !  Neural  Network       •  Objec<ve:  op<mize  various  Ranking  Metrics   •   exploit  proper<es  of  MF  model  &  data   •  Training   •  Related  Work   •  Experiments  
  • 23. 10  m  MovieLens    Data   •  10k  movies    &    70k  users   •  1%  dense  data   •  binarized:          3+  star  ra<ng  !  1,  otherwise  0   •  5-­‐fold  cross-­‐valida<on  
  • 24. 10  m  MovieLens    Data                          5-­‐fold  cross-­‐valida<on          std    :      0.001    
  • 25. 10  m  MovieLens    Data          std=0.002  
  • 26. Nellix  Play  Data   •  Test  day:            4/9/2014     •  rela(ve          improvement          to  RMSE  training                                                                                                                                                                    std=1%  
  • 27. Nellix  Play  Data                                                        std=2%  
  • 28. Conclusions   •  learning-­‐to-­‐rank  approach:   – implicit  feedback  data   – proper<es  of  MF  model   ! Gaussian  distribu<on  of  scores   ! non-­‐linear  ac<va<on  func<ons  derived  for  ranking   •  pointwise  and  listwise  training   •  various  ranking  metrics  can  be  used:   – compe<<ve  for  op<mizing  AUC   – par<cularly  effec<ve  at  head  of  ranked  list  
  • 29. Thank  You  !   Ques5ons  ?