attention please! a hybrid resource recommender mimicking attention-interpretation dynamics

14
h"p://LearningLayerseu – Scaling up Technologies for Informal Learning in SME Clusters – [email protected] h"p://LearningLayerseu – Scaling up Technologies for Informal Learning in SME Clusters – [email protected] Learning Layers Scaling up Technologies for Informal Learning in SME Clusters Attention Please! A Hybrid Resource Recommender Mimicking Attention-Interpretation Dynamics Paul Seitlinger, Dominik Kowald, Simone Kopeinik, Ilire Hasani-Mavriqi, Tobias Ley, Elisabeth Lex 1 Austrian Science Fund: P 25593-G22

Upload: elisabeth-lex

Post on 11-Aug-2015

154 views

Category:

Science


1 download

TRANSCRIPT

h"p://Learning-­‐Layers-­‐eu    –  Scaling  up  Technologies  for  Informal  Learning  in  SME  Clusters  –  layers@learning-­‐layers.eu  h"p://Learning-­‐Layers-­‐eu    –  Scaling  up  Technologies  for  Informal  Learning  in  SME  Clusters  –  layers@learning-­‐layers.eu  

Learning Layers Scaling up Technologies for Informal Learning in SME Clusters

Attention Please! A Hybrid Resource Recommender Mimicking Attention-Interpretation Dynamics Paul Seitlinger, Dominik Kowald, Simone Kopeinik, Ilire Hasani-Mavriqi, Tobias Ley, Elisabeth Lex

1  

Austrian Science Fund: P 25593-G22

h"p://Learning-­‐Layers-­‐eu    –  Scaling  up  Technologies  for  Informal  Learning  in  SME  Clusters  –  layers@learning-­‐layers.eu  

What will this talk be about?

•  Resource  RecommendaBon  (user-­‐based  CollaboraBve  Filtering)  

•  A  computaBonal  model  of  human  category  learning  (SUSTAIN)  

•  A  novel  hybrid  recommender  approach  that  combines  both  to  further  personalize  and  improve  CF  

2  

h"p://Learning-­‐Layers-­‐eu    –  Scaling  up  Technologies  for  Informal  Learning  in  SME  Clusters  –  layers@learning-­‐layers.eu  

Why?

•  Recommender  research  exploits  digital  traces  of  social  acBons  and  interacBons  – E.g.  CF  suggests  resources  of  most  similar  users  

•  EnBBes  of  different  quality  (e.g.,  users,  resources,  tags)  are  related  to  each  other  

•  In  CF,  users  just  another  enBty  •  Structuralist  simplificaBon  •  Neglects  nonlinear,  user-­‐resource  dynamics  that  shape  a"enBon  and  interpretaBon  

•  No  ranking  of  resources  in  CF  3  

h"p://Learning-­‐Layers-­‐eu    –  Scaling  up  Technologies  for  Informal  Learning  in  SME  Clusters  –  layers@learning-­‐layers.eu  

SUSTAIN (Love et al., 2004)

4  

•  Resource  represented  by  features    •  Cluster(s)  H    

–  Vector  of  values  along  the  n  feature  dimensions  

–  Fields  of  interest  •  A"enBonal  weights  wi:    

–  Importance  of  feature  for  user  •  Training  (for  each  resource  R)  

–  Start  with  one  cluster  –  Form  new  cluster  if  sim(R,H)  <  T  –  AdjusBng  Hj  and  wi  aYer  each  run  

•  TesBng  –  Compare  features  of  candidate  to  best  cluster  

h"p://Learning-­‐Layers-­‐eu    –  Scaling  up  Technologies  for  Informal  Learning  in  SME  Clusters  –  layers@learning-­‐layers.eu  

Our Approach: SUSTAIN+CFU

•  Step  1:  Create  candidate  set  Cu  for  target  user  u  (top  100  resources  of  CFU  

•  Step  2:  Train  SUSTAIN  network  of  target  user  u  •  Step  3:  Apply  each  candidate  c  of  Cu  to  network  •  Step  4:  Hybrid  approach  

5  

h"p://Learning-­‐Layers-­‐eu    –  Scaling  up  Technologies  for  Informal  Learning  in  SME  Clusters  –  layers@learning-­‐layers.eu  

Evaluation: Datasets

•  Social  tagging  systems  –  Freely  available  for  scienBfic  purposes  –  Topics  can  be  easily  derived  from  tagging  data    

 (e.g.,  Krestel  et  al.,  2010)  à Latent  Dirichlet  AllocaBon  (LDA)  with  500  topics  

•  No  p-­‐core  pruning  but  deleted  unique  resources  

6  

h"p://Learning-­‐Layers-­‐eu    –  Scaling  up  Technologies  for  Informal  Learning  in  SME  Clusters  –  layers@learning-­‐layers.eu  

Evaluation: Method and Metrics

•  Training  and  test-­‐set  splits  •  Per  user:  20%  most  recent  for  tesBng,  80%  for  training  •  Retains  chronological  order    

à  predict  future  based  on  the  past  •  Comparison  of  top-­‐20  recommended  resources  with  relevant  resources  from  test-­‐set  

•  Metrics  –  nDCG@20  – MAP@20  –  Precision  /  Recall  plots  (k  =  1  –  20)  

7  

h"p://Learning-­‐Layers-­‐eu    –  Scaling  up  Technologies  for  Informal  Learning  in  SME  Clusters  –  layers@learning-­‐layers.eu  

Baseline Algorithms

•  Most  Popular  (MP)  •  User-­‐Based  CollaboraBve  Filtering  (CFU)  •  Resource-­‐Based  CollaboraBve  Filtering  (CFR)  •  Content-­‐based  Filtering  using  Topics  (CBT)  •  SUSTAIN+CFU  à Available  in  the  open  source  TagRec  framework  

•  Weighted  Regularized  Matrix  FactorizaBon  (WRMF)  à  MyMediaLite  

8  

h"p://Learning-­‐Layers-­‐eu    –  Scaling  up  Technologies  for  Informal  Learning  in  SME  Clusters  –  layers@learning-­‐layers.eu  

Results

9  

•  SUSTAIN+CFU  improves  CFU  on  all  three  datasets    •  CiteULike:  High  average  topic  similarity  per  user,  CFR  wins  •  Delicious:  Mutual-­‐fan  crawling  strategy,  WRMF  wins  

h"p://Learning-­‐Layers-­‐eu    –  Scaling  up  Technologies  for  Informal  Learning  in  SME  Clusters  –  layers@learning-­‐layers.eu  

Evaluation: Open Issues •  Datasets  

–  Other  Delicious  dataset,  LastFM,  MovieLens  –  External/other  feature  (not  dependent  on  LDA)  

•  Other  metrics  –  Diversity,  Serendipity,  Coverage  

•  ComputaBonal  Costs  –  Our  experiments  showed  that  our  approach  is  much  faster  than  CFR  and  especially  WRMF  

•  Although  LDA  is  needed    –  RunBme  experiment  is  needed  +  computaBonal  complexity  

•  Online  evaluaBon  –  Learning  Layers  field  study  

10  

h"p://Learning-­‐Layers-­‐eu    –  Scaling  up  Technologies  for  Informal  Learning  in  SME  Clusters  –  layers@learning-­‐layers.eu  

Future Work

•  Technical  – CF-­‐independent  variant  

•  RecommendaBons  solely  based  on  user-­‐specific  SUSTAIN  network  

– Detailed  analysis  of  computaBonal  costs  •  Conceptual  

– Dynamic  recommendaBon  logic  •  Exploring  relaBonship  between  a"enBonal  focus  and  novelty  seeking  and  use  this  for  recommendaBon  

11  

h"p://Learning-­‐Layers-­‐eu    –  Scaling  up  Technologies  for  Informal  Learning  in  SME  Clusters  –  layers@learning-­‐layers.eu  

Take Away Messages

•  Our  approach  SUSTAIN  +  CFU  can  improve  CF  predicBons  – More  robust  in  terms  of  accuracy  esBmates  –  From  our  observaBon:  less  complex  in  terms  of  computaBonal  efforts  

•  User-­‐resource  dynamics,  if  modelled  with  a  connecBonist  approach,  can  help  gain  a  deeper  understanding  of  Web  interacBons  in  terms  of  a"enBon,  categorizaBon  and  decision  making  

12  

h"p://Learning-­‐Layers-­‐eu    –  Scaling  up  Technologies  for  Informal  Learning  in  SME  Clusters  –  layers@learning-­‐layers.eu  

Code and Framework

13  

•  TagRec  framework  •  hAps://github.com/learning-­‐layers/TagRec/  

 •  Framework  for  developing  and  evaluaBng  new            recommender  algorithms  in  folksonomies  •  Contains  our  approach,  the  baseline  algorithms              and  the  evaluaBon  protocol  and  metrics  •  Capable  of  tag,  resource  and  user  recommendaBons  

•  Used  as  recommender  engine  in  the  Learning  Layers  EU  project  •  Links  to  the  datasets  we  used:  

–  BibSonomy  (2013-­‐07-­‐01):  h"p://www.kde.cs.uni-­‐  kassel.de/bibsonomy/dumps/    –  CiteULike  (2013-­‐03-­‐10):  h"p://www.citeulike.org/faq/data.adp    –  Delicious  (2011-­‐05-­‐01):  h"p://files.grouplens.org/datasets/hetrec2011/  hetrec2011-­‐

delicious-­‐2k.zip    

h"p://Learning-­‐Layers-­‐eu    –  Scaling  up  Technologies  for  Informal  Learning  in  SME  Clusters  –  layers@learning-­‐layers.eu  

Thank you for your attention!

 QuesJons?  

 Elisabeth  Lex  

[email protected]  Ass.  Prof.  at  Graz  University  of  Technology  (Austria)  Head  of  Social  CompuBng  at  Know-­‐Center  (Austria)  

 

14