Fostering Friendships - Enhancing Social Bonds in the Classroom
A Hybrid Peer Recommender System for a Online Community Teachers
1. A Hybrid Peer Recommender
System for a Online Community
Teachers
Cristian Miranda*, Julio Guerra**, Denis Parra**, Eliana
Scheihing*
Southern University of Chile*
University of Pittsburgh**
SRS: 3rd International Workshop on Social Recommender
Systems ~ at UMAP 2012
2. Outline
• Introduction to Kelluwen Project
• What is Being Recommended?
• Recommendation Model
• Evaluation/Results
• Conclusions/Future Work
4. Kelluwen Project
• Kelluwen is a mapudungun word that means
“Group of People” that we used as
“Collaborative Work”.
• The Kelluwen project address deficits in socio-
communicate skills of vulnerable students in
ages 12~17 from
• Technology involved: The creation of a
community of students, teachers and
researchers supported by web 2.0 tools.
5. What is being recommended?
• Our RecSys recommends “comments and
suggestions” made by teachers when they
conduct these educational activities.
• Educational Activities. These are activities
performed in class by students. They report them
and comment them in the Kelluwen platform.
• Teachers that are guiding some activity for the
first time can have a big support if they receive
comments from teachers (peers) that have
already conducted that activity .
6. What is being recommended?
Educational Activity (materials, durations, … MAX 3 students
link to comments, status) per group…
Recommendations: comments & … ppt too long,
suggestions made by other teachers might skip some
that already guided the same activity slides…
8. Recommendation Model (2/3)
• Peers’ similarity considers (u1)
Variable ID Values
Low: 1
School Socio-economical Mean Low: 2
Level NS Mean: 3
Mean High: 4
High: 5
School Quality of TIC CT Sum of three values
Infrastructure
Lower than 10.000 habitants: 1
Size of School Locality TL Between 10.000 and 100.000 habitants: 2
Greater than 100.000 bahitants: 3
Mean Number of Students N Mean number of stufents by classroom for
each teacher
per Classroom
9. Recommendation Model (3/3)
• Items assessed based on creator’s value (u2):
Variable ID Values
Number of Number of didacticaldesignsappliedbyteachert
(sendingthesuggestionss), normalizedtothe [0,1]
didacticaldesignsappliedbyteache interval
rt (sendingthesuggestions)
Verywell: 0
Activityevaluationfromtheteacher Well: 1
sendingthesuggestions Bad: 1
VeryBad: 0
MessageNature Activitycomment: 0
Suggestion: 1
Number of rating of Number of rating of allmessagesfromteachert
(sendingthesuggestions), normalizedtothe [0,1]
allmessagesfromteachert interval
(sendingthesuggestions)
Number of ratingsof suggestions Number of rating of suggestions, normalizedtothe
[0,1] interval
10. Evaluation(1/2) nDCG
• Evaluation done with nDCG (normalized discounted cummulative gain)
• Phase 1: 29 teachers provided comments and suggestions to their
educational activities inside the Kelluwen system
• Phase 2: # recommendations generated between Oct/Nov 2011: 314
• Phase 2: # recommendations that were provided feedback: 96
• Here there’s a plot with nDCG of 39 recommendation sets (6 people).
11. User satisfaction survey
Used the
70% recommendations in
their educational
activities
Thought that the
recommendations were
useful during the
68%
activity execution
¿Did I consider the Totally agree ¿Were the recommendations useful
recommendation while to develop my activities?
executing the educational Disagree
activities?
Agree
Totally agree
12. Conclusions
• We introduced a RecSys that recommends
novel items (comments of teachers to
educational activities) in a novel domain (a
community portal for students, teachers and
researchers)
• The RecSys had a good level of acceptance:
68% of the teachers that considered the
recommendations found them useful in their
educational activities.
13. Future work
• Incorporate user engagement techniques to
obtain more feedback of the utility of
recommendations.
• To Develop a “recommendation history”: Users
would be able to see which recommendations
they have received and how they evaluated
them.
• Include new metrics to evaluate the
recommender system.
16. Sensibility Analysis
• Best combination of k1 (0.8) and k2 (0.2) values to
combine u1 (peer similarity) and u2 (item and item-
creator value) ~ best nDCG
17. Resultados
• Encuesta de Satisfacción de Usuarios
El sistema adaptativo de
recomendación de pares es un
mecanismo de apoyo a sus
70%
actividades docentes
¿Aprecio las recomendaciones de mis
pares como un mecanismo de apoyo
para el desarrollo de las actividades?
Notas do Editor
The “vulnerable” is in terms of isolation: They are from rural areas or poor suburbs of southern Chile.Students and teachers have access to internet mainly at the Schools, but some of them ( that live in the cities) also have access from home.
Details of u1 and u2 can be seen in the paper. Next 2 slides explain user profile and item profile dimensions
The PRESTIGE of the teacher who created the item is a way to measure how good is the item.