Experience Discovery: Hybrid Recommendation of Student Activities - PowerPoint PPT Presentation

Experience Discovery: Hybrid Recommendation of Student Activities using Social Network Data Robin Burke, Yong Zheng, Scott Riley Web Intelligence Laboratory College of Computing and Digital Media DePaul University

Problem  Service organizations offer many educational programs and activities for youth  Participation (especially by underprivileged youth) is low  Even though these are the individuals who would benefit the most  How to get better participation?  not just a recommendation problem

The Role of Recommendation  Need for personalization  Many diverse activities  from basketball to poetry to robots to knitting  Low tolerance for imprecise results  Need for system initiative  user research shows that students are unlikely to search and browse  To “pull” opportunities  system should “push” suggestions  we are considering mobile platforms

Partners  Digital Youth Network  service organization focused on the creation of digital media  Nichole Pinkard  YouMedia  school-based online social network  affiliated with DYN  Chicago Learning Network  consortium of museums and non-profits  Chicago Public Schools  Funders  MacArthur Foundation  Gates Foundation

Experience Discovery: Research opportunities  Full cycle observation  activity enrollments  activity attendance  click-through  post-activity rating, tagging, reviewing  Social network data  uploading of digital media  browsing / commenting behavior  friend / follower connections

Research question 1  There are multiple important knowledge sources  past enrollment history  content data  social network data  log data  Mixed vs integrated hybrid recommendation  should different knowledge sources be integrated in making recommendations?  or should recommendations of different types be presented side-by-side?

Research question 2  Activities sometimes have a logical planned sequence  Video editing I -> Video editing II  Sometimes they are sequenced idiosyncratically  Digital photography -> Zoo explorer I  Educational goal  increase both depth and breadth of student participation  The role of “curricula”  how can recommendations be used to increase both breadth and depth of student involvement?  what is the role of top-down vs bottom-up sequences in recommendation?

Research question 3  Dynamics of interest  students mature a lot between 11 and 18  old activities may lose their appeal  Dynamics of offerings  activities change from year to year and season to season  may not be explicit  Coping with change  how can we ensure that recommendations don’t lag student interest?  how to detect and respond to program changes?

Research question 4  Students aren’t the only ones with questions  Service providers can get value, too  what activities should I offer and where?  how do my offerings compare to other groups?  what needs are not being met?  Analytics and recommendations for service providers  what can we provide that is helpful and comprehensible?

Architecture ! "#$% & $( 0$,- $01' ' $( 2*31( ,4+ *51% ' , ?+ & $( ), D#$% *31( *+ , 6037& )8,9*)*, 6+ A1% & )@ ' ,E& F% *% 8, 6##+ & 0*31( , > ( )$% C *0$, 6G$( 2*( 0$, 9*)*, - $01' ' $( 2*31( , - $./+ ), > ( #/), ! ( A& ( $, ?*0@ $, ! "#$% & ' $( )*+ , ?*0@ $, ?1( BA/% *31( ., : 10& *+ , ; $)< 1% =, ! 7*+ /*31( , ! "#$% & ' $( )*+ , 9*)*, > ( )$% C *0$, - $./+ ).,

Initial experiments  Data (2 schools)  226 students  32 activities  3800 records  (now adding ~2000 enrollments and ~50 activities / month)  Algorithms  collaborative / binary  collaborative / pseudo rating  content-collaborative meta-level hybrid  plus behavioral descriptors

Pseudo-ratings  Some activities are attended multiple times  evidence of strong interest  Example  book discussion group  Normalize to user’s profile  weight for activity a = # of times attending a / total attendances  Can we normalize in other ways?  take into account how often something was offered

Meta-level hybrid  Use course topic descriptors  13 choices  health, music, visual arts, etc.  activities may include several topics  Build a topic profile by summing over descriptions of all activities  Compare users based on topic profiles  rather than attendance data

Adding social network data  Extracted 10 features from the social network  counts of uploaded media types  overall level of activity  Used feature combination  content profile 8$93& (: . 4) $& , 0 .  behavior profile ! , (, . 5"(6 ) % 7. ! "#$% & ' () % #. 1"2, 3& ) % .- , (, . ! "#$% & ' () % * +, #"- . 1"2, 3& ) % * +, #"- . ' % ) /0 ". ' % ) /0 ".

Results  Temporal leave-one-out evaluation  see Burke, 2010  Look at a user’s experience over time  looking at users divided by  # of enrollments (profile size)  profile diversity (# of different enrollments)  Need to do more research  Hybrid 2 works best for large, diverse users  Doesn’t matter what you do for non - diverse users

Conclusions  We are in the early stages here  Eager to get our hands on bigger data  Many research questions  Would like to hear ideas

Thanks  Questions / Comments / Ideas