Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Recommender Systems in TEL

2,377 views

Published on

Nikos Manouselis

Published in: Education, Technology
  • Be the first to comment

Recommender Systems in TEL

  1. 1. Recommender Systems in TEL Nikos Manouselis Greek Research & Technology Network (GRNET) nikosm@ieee.org
  2. 2. about me <ul><li>Computer Engineer </li></ul><ul><li>MSc on Operational Research </li></ul><ul><li>PhD from Informatics Lab of an Agricultural University </li></ul><ul><li>working on services for agricultural & rural communities </li></ul><ul><ul><li>learning repositories </li></ul></ul><ul><ul><li>social information retrieval </li></ul></ul><ul><ul><li>Organic.Edunet e Content plus </li></ul></ul>
  3. 3. (promised)aim of this lecture <ul><li>introduce recommender systems </li></ul><ul><li>discuss how they relate to TEL </li></ul><ul><li>identify open research issues </li></ul>
  4. 4. (actual)aim of this lecture <ul><li>share some concerns about TEL and recommender systems </li></ul>
  5. 5. structure <ul><li>tale of 3 friends </li></ul><ul><li>tasks </li></ul><ul><li>modeling & techniques </li></ul><ul><li>evaluation </li></ul><ul><li>wrap up </li></ul>
  6. 6. intro: tale of 3 friends
  7. 7. which movie?
  8. 8. lets ask some friend <ul><li>“ Guys, heard about the last Batman movie… should I watch it?” </li></ul>“ You will definitely like it” “ Maybe not, the scenario is too weak”
  9. 9. lets ask some friend <ul><li>“ Wait – did you like the previous one?” </li></ul>
  10. 10. … so, which movie? <ul><li>taking advantage of knowledge or experience from people in the social circle or network </li></ul><ul><ul><li>e.g. colleagues, friends, peers </li></ul></ul><ul><li>need to answer several questions </li></ul><ul><ul><li>how to identify like-minded people? </li></ul></ul><ul><ul><li>on which dimensions? </li></ul></ul><ul><ul><li>for which types of items? </li></ul></ul><ul><ul><li>does context matter? </li></ul></ul><ul><ul><li>… </li></ul></ul>
  11. 11. recommender systems
  12. 14. <ul><li>using the opinions of a community of users </li></ul><ul><ul><li>to help individuals in that community to identify more effectively content of interest </li></ul></ul><ul><ul><li>from a potentially overwhelming set of choices </li></ul></ul><ul><li>Resnick P. & Varian H.R., “Recommender Systems”, Communications of the ACM, 40(3),1997 </li></ul>definition (1/2)
  13. 15. definition (2/2) <ul><li>any system that </li></ul><ul><ul><li>produces individualized recommendations as output </li></ul></ul><ul><ul><li>or has the effect of guiding the user in a personalized way to interesting or useful objects in a large space of possible options </li></ul></ul><ul><li>Burke R. “Hybrid Recommender Systems: Survey and Experiments”, User Modeling & User-Adapted Interaction, 12, 331-370, 2002 </li></ul>
  14. 16. why do we need them? <ul><li>A trip to a local supermarket [F. Ricci] : </li></ul><ul><ul><li>85 different varieties and brands of crackers </li></ul></ul><ul><ul><li>285 varieties of cookies </li></ul></ul><ul><ul><li>165 varieties of “juice drinks” </li></ul></ul><ul><ul><li>75 iced teas </li></ul></ul><ul><ul><li>275 varieties of cereal </li></ul></ul><ul><ul><li>120 different pasta sauces </li></ul></ul><ul><ul><li>80 different pain relievers </li></ul></ul><ul><ul><li>40 options for toothpaste </li></ul></ul><ul><ul><li>95 varieties of snacks (chips, pretzels, etc.) </li></ul></ul><ul><ul><li>61 varieties of sun tan oil and sunblock </li></ul></ul><ul><ul><li>360 types of shampoo, conditioner, gel, and mousse. </li></ul></ul><ul><ul><li>90 different cold remedies and decongestants. </li></ul></ul><ul><ul><li>230 soups, including 29 different chicken soups </li></ul></ul><ul><ul><li>175 different salad dressings </li></ul></ul>
  15. 17. wait a second <ul><li>is TEL like a super market?? </li></ul>
  16. 18. large number of options
  17. 19. tasks for recommender systems
  18. 20. tasks usually supported <ul><li>annotation in context </li></ul><ul><li>find good items </li></ul><ul><li>find all good items </li></ul><ul><li>receive sequence of items </li></ul><ul><li>(+some less important ones) </li></ul><ul><li>Herlocker et al., “Evaluating Collaborative Filtering Recommender Systems” ACM Transactions on Information Systems, 22(1), 5-53, 2004. </li></ul>
  19. 21. 1. annotation in context <ul><li>integrated in existing working environment to provide additional support or information, e.g. </li></ul><ul><ul><li>predicted usefulness of an item that the user is currently viewing </li></ul></ul><ul><ul><li>links within a Web page that the user is recommended to follow </li></ul></ul>
  20. 22. annotation in context <ul><li>Screenshot/example </li></ul>
  21. 23. 2. find good items <ul><li>suggesting specific item(s) to a user </li></ul><ul><ul><li>characterized as core recommendation task, since occurring in most systems </li></ul></ul><ul><ul><li>e.g. presenting a ranked list of recommended items </li></ul></ul>
  22. 24. find good items <ul><li>Screenshot/example </li></ul>
  23. 25. 3. find all good items <ul><li>user wants to identify all items that might be interesting </li></ul><ul><ul><li>when its important not to overlook any potentially relevant case </li></ul></ul><ul><ul><li>e.g. medical or legal cases </li></ul></ul>
  24. 26. find all good items
  25. 27. 4. sequence of items <ul><li>sequence of related items is recommended to the user </li></ul><ul><ul><li>e.g. entertainment applications such as TV or radio programs </li></ul></ul>
  26. 28. sequence of items
  27. 29. and what about TEL? <ul><li>informal reminder: </li></ul><ul><ul><li>technology enhanced learning is generally dealing with the ways ICT can be used to support learning , teaching , and competence development </li></ul></ul><ul><li>[http://cordis.europa.eu/fp7/ict/telearn-digicult/telearn_en.html] </li></ul>
  28. 30. break2think <ul><li>bring yourself in one typical learning situation that occurs very often to YOU </li></ul>
  29. 31. break2think <ul><li>imagine that some magic TEL system is there to support you </li></ul><ul><ul><li>it could make some great suggestions about something to you </li></ul></ul><ul><li>name one learning task where a recommender system would be useful </li></ul>
  30. 32. modeling & techniques
  31. 33. typical classification <ul><li>content-based: information needs of user and characteristics of items are represented in some (usually textual) form </li></ul><ul><li>collaborative filtering: user is recommended items that people with similar tastes and preferences liked </li></ul><ul><li>hybrid: methods that combine content-based and collaborative methods </li></ul><ul><li>… other categorizations also exist (Burke, 2002) </li></ul>
  32. 34. example: content-based
  33. 35. example: collaborative filtering
  34. 36. generally speaking: some user <ul><li>has a profile with some user characteristics, e.g. </li></ul><ul><ul><li>past ratings [collaborative filtering] </li></ul></ul><ul><ul><li>keywords describing past selections [content-based recommendation] </li></ul></ul>
  35. 37. generally speaking: some items <ul><li>are represented using some dimensions, e.g. </li></ul><ul><ul><li>satisfaction over one (or more) criteria [collaborative filtering] </li></ul></ul><ul><ul><li>item attributes/features [content-based recommendation] </li></ul></ul>
  36. 38. generally speaking: a mechanism <ul><li>is taking advantage of the user profile and the item representations </li></ul><ul><ul><li>it provides personalised recommendations of items to users </li></ul></ul>
  37. 39. rings some bell? <ul><ul><li>for TEL, this sounds so… </li></ul></ul><ul><ul><li>adaptive educational hypermedia systems ( AEHS ) </li></ul></ul>
  38. 40. a generic architecture [Karampiperis & Sampson, 2005]
  39. 41. an example [Karampiperis & Sampson, 2005]
  40. 42. <ul><li>enhanced version of [Hanani et al., &quot;Information Filtering: Overview of Issues, Research and Systems&quot;, User Modeling and User-Adapted Interaction, 11, 2001] </li></ul>classification/analysis
  41. 43. recommend in TEL based on what? <ul><li>on learner models/profiles </li></ul><ul><ul><li>e.g. learning styles, competence gaps </li></ul></ul><ul><ul><li>… other ideas? </li></ul></ul><ul><li>on item characteristics </li></ul><ul><ul><li>e.g. interactivity, granularity, accessibility </li></ul></ul><ul><ul><li>… other ideas? </li></ul></ul>
  42. 44. evaluation
  43. 45. evaluating recommendation <ul><li>currently based on performance </li></ul><ul><li>“ how good are your algorithms?” </li></ul><ul><li>e.g. </li></ul><ul><ul><li>how accurate are they in predictions? </li></ul></ul><ul><ul><li>for how many unknown items can they produce a prediction? </li></ul></ul><ul><ul><li>… mainly information retrieval evaluation approaches </li></ul></ul><ul><li>[Herlocker et al., “Evaluating Collaborative Filtering Recommender Systems” ACM Transactions on Information Systems, 22(1), 5-53, 2004] </li></ul>
  44. 46. typical results means that a prediction could be 4,6 stars instead of 4 or 5 … does this really matter in TEL?
  45. 47. other issues <ul><li>live experiments vs. offline analyses </li></ul><ul><li>synthesized vs. natural data sets </li></ul><ul><ul><li>properties of data sets </li></ul></ul><ul><ul><li>existing data sets </li></ul></ul>
  46. 48. metrics (popular) <ul><li>accuracy </li></ul><ul><ul><li>predictive accuracy (MAE) </li></ul></ul><ul><ul><li>classification accuracy </li></ul></ul><ul><li>Precision and Recall </li></ul><ul><ul><li>probability that a selected item is relevant </li></ul></ul><ul><ul><li>probability that a relevant item will be selected </li></ul></ul><ul><li>ad hoc </li></ul><ul><ul><li>Rank Accuracy Metrics </li></ul></ul><ul><ul><li>Prediction-Rating Correlation </li></ul></ul><ul><li>coverage </li></ul><ul><ul><li>percentage of items for which prediction is possible </li></ul></ul>
  47. 49. metrics (not popular) <ul><li>novelty </li></ul><ul><li>serendipity </li></ul><ul><li>confidence </li></ul><ul><li>user evaluation </li></ul><ul><ul><li>explicit (ask) vs. implicit (observe) </li></ul></ul><ul><ul><li>laboratory studies vs. field studies </li></ul></ul><ul><ul><li>outcome vs. process </li></ul></ul><ul><ul><li>short-term vs. long-term </li></ul></ul>
  48. 50. evaluation in TEL recommenders <ul><li>few systems actually evaluated </li></ul><ul><ul><li>even fewer actually tried with users </li></ul></ul><ul><li>recent analysis of 15 TEL recommender systems: </li></ul><ul><ul><li>half of the systems (8/15) still at design or prototyping stage </li></ul></ul><ul><ul><li>only 5 systems evaluated through trials with human users </li></ul></ul><ul><ul><li>[N.Manouselis, H.Drachsler, R.Vuorikari, H.Hummel, R.Koper, “Recommender Systems in Technology Enhanced Learning”, Handbook of Recommender Systems (under review)] </li></ul></ul>
  49. 51. example: Altered Vista <ul><li>evaluate the effectiveness and usefulness </li></ul><ul><ul><li>system usability and performance </li></ul></ul><ul><ul><li>predictive accuracy of recommender engine </li></ul></ul><ul><ul><li>extent to which reviewing Web resources within a community of users supports and promotes collaborative and c ommunity-building activities </li></ul></ul><ul><ul><li>extent to which critical review of Web resources leads to improvements in user’s information literacy skills </li></ul></ul><ul><ul><li>[Walker et al., “ Collaborative Information Filtering: a review and an educational application”, International Journal of Artificial Intelligence in Education 14, 2004 ] </li></ul></ul>
  50. 52. another look at it <ul><li>e.g. using Kirckpatrick’s model on evaluating training programs </li></ul><ul><ul><li>reaction of student - what they thought and felt about the training </li></ul></ul><ul><ul><li>learning - the resulting increase in knowledge or capability </li></ul></ul><ul><ul><li>behaviour - extent of behaviour and capability improvement and implementation/application </li></ul></ul><ul><ul><li>results - the effects on the business or environment resulting from the trainee's performance </li></ul></ul>
  51. 53. what else could be evaluated? <ul><li>when deploying a recommender system in a TEL setting </li></ul><ul><li>… what could we evaluate and how to measure it? </li></ul>
  52. 54. wrap up & directions
  53. 55. basic conclusion <ul><li>assuming an information overload problem in TEL </li></ul><ul><ul><li>recommender systems are good </li></ul></ul><ul><ul><li>need to think out of the box </li></ul></ul><ul><ul><li>connect with existing research </li></ul></ul><ul><ul><li>focus on TEL particularities </li></ul></ul><ul><ul><li>explore alternative uses </li></ul></ul><ul><ul><li>integrate with existing theories </li></ul></ul>
  54. 56. interesting (?) issues <ul><li>recommendation of peers </li></ul><ul><li>criteria for expressing learner satisfaction (no more 5-stars) </li></ul><ul><li>study actual usage/acceptance </li></ul><ul><li>assess performance/learning improvement </li></ul><ul><li>… implement, deploy, pilot! </li></ul>
  55. 57. but do they exist?? <ul><li>http://www.oerrecommender.org </li></ul>
  56. 58. interested in more? <ul><li>Journal of Digital Information (JoDI) </li></ul><ul><ul><li>Special Issue on Social Information Retrieval for Technology-Enhanced Learning, 10(2), 2009 </li></ul></ul><ul><li>Workshop on Social Information Retrieval for Technology Enhanced Learning (SIRTEL) </li></ul><ul><ul><li>SIRTEL 2007 (http://ceur-ws.org/Vol-307) </li></ul></ul><ul><ul><li>SIRTEL 2008 (http://ceur-ws.org/Vol-382) </li></ul></ul><ul><ul><li>SIRTEL 2009 (http://celstec.org/sirtel) </li></ul></ul><ul><ul><ul><li>co-located with ICWL’09, Aachen, Germany, August 21 st - deadline: 12/6 </li></ul></ul></ul>
  57. 59. thank you! questions? ideas?

×