35
© author(s) of these slides including research results from the KOM research network and TU Darmstadt; otherwise it is specified at the respective slide 21-Sep-12 Prof. Dr.-Ing. Ralf Steinmetz KOM - Multimedia Communications Lab ECTEL__Sem_Info_rec_learning_resources_v6.0_20120921_MA.pptx Exploiting Semantic Information for Graph-based Recommendations of Learning Resources Mojisola Anjorin Thomas Rodenhausen Renato Domínguez García Christoph Rensing EC-TEL 2012, Saarbrücken Research Talk Ranking Algorithms Slideshare Tags Resources Users Prepare Talk Read-Up on Basics Activities Find Related Work Friends Friends Friends Blue Group

Exploiting Semantic Information for Graph-based Recommendations of Learning Resources

Embed Size (px)

Citation preview

© author(s) of these slides including research results from the KOM research network and TU Darmstadt; otherwise it is specified at the respective slide 21-Sep-12

Prof. Dr.-Ing. Ralf Steinmetz KOM - Multimedia Communications Lab

ECTEL__Sem_Info_rec_learning_resources_v6.0_20120921_MA.pptx

Exploiting Semantic Information for Graph-based Recommendations of Learning Resources

Mojisola Anjorin Thomas Rodenhausen Renato Domínguez García Christoph Rensing

EC-TEL 2012, Saarbrücken

Research Talk

Ranking

Algorithms

Slideshare

Tags ResourcesUsers

Prepare Talk

Read-Up on Basics

Activities

Find Related Work

Friends

Friends

FriendsBlue Group

KOM – Multimedia Communications Lab 2

Resource-Based Learning

KOM – Multimedia Communications Lab 3

Application Scenario: CROKODIL

CROKODIL is a platform offering support for resource-based learning § Semantic Tag Types § Activities § Learner Groups & Friendships § Recommendations

[Anjorin et al, 2011]

http://demo.crokodil.de

KOM – Multimedia Communications Lab 4

§ Motivation: Resource-based Learning § Application Scenario: CROKODIL § CROKODIL’s Extended Folksonomy Model § Ascore and AInheritScore § Evaluation Methodology, Metrics and Results § Conclusion & Future Work

Overview

KOM – Multimedia Communications Lab 5

A folksonomy is a quadruple F:= (U, T, R, Y), where U – Users T – Tags R – Resources Y ⊆ U × T × R - tag assignment

Folksonomy Model

Research Talk

Ranking

Algorithms

Slideshare

Tags ResourcesUsers

[Hotho et al. 2006]

KOM – Multimedia Communications Lab 6

CROKODIL Extends the Folksonomy Model …

Research Talk

Ranking

Algorithms

Slideshare

Tags ResourcesUsers

KOM – Multimedia Communications Lab 7

… with Semantic Tag Types

[Böhnstedt et al. 2009]

Research Talk

Ranking

Algorithms

Slideshare

Tags ResourcesUsers

Genre

Event

Person

Location

Other

Topic

KOM – Multimedia Communications Lab 8

… with Activities

Research Talk

Ranking

Algorithms

Slideshare

Tags ResourcesUsers

Prepare Talk

Read-Up on Basics

Activities

Find Related Work

KOM – Multimedia Communications Lab 9

… with Learner Groups and Friendships

Research Talk

Ranking

Algorithms

Slideshare

Tags ResourcesUsers

Prepare Talk

Read-Up on Basics

Activities

Find Related Work

Friends

Friends

FriendsBlue Group

KOM – Multimedia Communications Lab 10

CROKODIL‘s Extended Folksonomy

FC:= (U, TTyped, R, YT, (A, <), YA, YU, G, friends) where U – users TTyped – typed tags R – learning resources YT ⊆ U × TTyped × R – tag assignment (A, <) – activities with sub-activities YA ⊆ U × A × R – activity assignment YU ⊆ U × A – activity membership

assignment G ⊆ P(U) – groups of learners friends ⊆ U × U – friendship relation

Research Talk

Ranking

Algorithms

Slideshare

Tags ResourcesUsers

Prepare Talk

Read-Up on Basics

Activities

Find Related Work

Friends

Friends

FriendsBlue Group

KOM – Multimedia Communications Lab 11

Resource Recommendations for CROKODIL

http://demo.crokodil.de

KOM – Multimedia Communications Lab 12

Graph-based recommender techniques can be classified as neighbourhood-based collaborative filtering approaches

Graph-based Resource Recommendations

Graph-based Ranking

Algorithm

Resource Score r1 0.9 r2 0.7 r3 0.5 r4 0.2

1 1

2 1

P1

P2

P4

P3

3

4

2

1

2

Folksonomy Graph e.g. FolkRank based on “Random Walk” of PageRank

Recommendation List (ranked resources)

[Desrosiers et al. 2011]

KOM – Multimedia Communications Lab 13

§ Motivation: Resource-based Learning § Application Scenario: CROKODIL § CROKODIL’s Extended Folksonomy Model § Ascore and AInheritScore § Evaluation Methodology, Metrics and Results § Conclusion & Future Work

Overview

KOM – Multimedia Communications Lab 14

1.  Add activity nodes Vc = VF ∪ A 2.  Add edges: § activity assignments (u, r, a) § assignments of a user to an

activity (u, a) § activity hierarchies (asub , asuper)

4.  Assign weights to edges: § w(r,a) = w(r,u) = w(u,a)

= max(|Ut,r|) § w(u, a) = max(|Ru,t|) § w(asub,asuper) = max(|Ut,r|, |Ru,t|)

5.  Run graph-based ranking algorithm e.g. FolkRank

AScore

[Abel et al, 2011] Inspired by GFolkRank

Extend the Folksonomy Graph F = (V, E) with Activities

Research Talk

Ranking

Algorithms

Slideshare

Tags ResourcesUsers

Prepare Talk

Read-Up on Basics

Activities

Find Related Work

KOM – Multimedia Communications Lab 15

§ Depending on the tags of a user,

scores are “inherited” over the activity hierarchy

§ Resources and users assigned to activities influence the scores as well

§ Scores are attenuated depending on activity distance § Activity distance between two

activities: the number of hops from one activity to the other

AInheritScore

[Abel et al, 2011] Inspired by GRank

Leveraging Activity Hierarchies to Calculate Scores Research Talk Ranking

Algorithms

Research Talk Prepare Talk

Read-Up on Basics

Find Related Work

...

... ...

KOM – Multimedia Communications Lab 16

§ Motivation: Resource-based Learning § Application Scenario: CROKODIL § CROKODIL’s Extended Folksonomy Model § Ascore and AInheritScore § Evaluation Methodology, Metrics and Results § Conclusion & Future Work

Overview

KOM – Multimedia Communications Lab 17

GroupMe! dataset

Evaluation Corpus and Evaluation Metrics

[Abel et al, GroupMe!]

Elements Count Users 649 Tags 2580 Resources 1789 Groups of Resources

1143

Posts 1865 Tag assignments 4366

The mean of the Average Precision over several queries Q

Mean Normalized Precision: The mean of the Precision@k over several queries Q

MAP(Q) =1

|Q|

|Q|�

j=1

1

mj

mj�

k=1

Precision(Rjk)

Mean Average Precision:

MNP(Q,k) =1

|Q|

|Q|�

j=1

Precisionj(k)

Precisionmax,j(k)

[Manning et al 2008]

KOM – Multimedia Communications Lab 18

Tango

Buenos

Aires

DancingFestival

Tango

Buenos

Aires

Dancing Festival

A post is a Pu,r= {(u,r,t)|(u,r,t) ∈ Y} For LeavePostOut, the recommendation task with user as input is harder as with tag as input

Evaluation Methodology: LeavePostOut

[Jäschke et al. 2007]

KOM – Multimedia Communications Lab 19

RTr,t= {(u,r,t)|(u,r,t) ∈ Y} For LeaveRTOut, the recommendation task with tag as input is harder as with user as input

Evaluation Methodology: LeaveRTOut

Tango

Buenos

Aires

DancingFestival

Tango

Buenos

Aires

Dancing Festival

KOM – Multimedia Communications Lab 20

A violin plot is a combination of a box plot and a density trace

Visualization of Results with Violin Plots

[Hintze et al. 1998]

KOM – Multimedia Communications Lab 21

A violin plot is a combination of a box plot and a density trace

Visualization of Results with Violin Plots

Median

3rd Quartile

1st Quartile [Hintze et al. 1998]

KOM – Multimedia Communications Lab 22

Evaluation results with user as input

Evaluation Results for LeavePostOut

KOM – Multimedia Communications Lab 23

Evaluation results with user as input

Evaluation Results for LeavePostOut

KOM – Multimedia Communications Lab 24

Evaluation results with user as input

Evaluation Results for LeavePostOut

KOM – Multimedia Communications Lab 25

Evaluation results with user as input

Evaluation Results for LeavePostOut

KOM – Multimedia Communications Lab 26

Evaluation results with user as input

Evaluation Results for LeavePostOut

KOM – Multimedia Communications Lab 27

Evaluation results with user as input

Evaluation Results for LeavePostOut

KOM – Multimedia Communications Lab 28

Evaluation Results for LeavePostOut

Approaches MAP GFolkRank 0.70 AScore 0.70 AInheritscore 0.47 GRank 0.38 FolkRank 0.19 Popularity 0.00

KOM – Multimedia Communications Lab 29

Evaluation Results for LeaveRTOut

Evaluation results with user as input

KOM – Multimedia Communications Lab 30

Evaluation Results for LeaveRTOut

Approaches MAP AScore 0.20 GFolkRank 0.20 FolkRank 0.18 GRank 0.14 AInheritscore 0.11 Popularity 0.02

KOM – Multimedia Communications Lab 31

Exploiting hierarchical activity structures as found in CROKODIL can improve the ranking of resources for the purpose of recommending learning resources § AScore § AInheritscore

Future Work § Evaluation using a data set from CROKODIL § User Study § Hybrid approaches

Conclusion and Future Work

www.crokodil.de

KOM – Multimedia Communications Lab 32

Questions & Contact

KOM – Multimedia Communications Lab 33

Statistical Significance Tests – LeavePostOut

More effective than à

Popularity Folk Rank

GFolkRank

AScore GRank AInheritScore

Poularity FolkRank X GFolkRank X X X X X AScore X X X X GRank X X AInheritScore X X X

Significance matrix of pair-wise comparisons of LeavePostOut results Based on Average Precision with a significance level of p = 0.05

KOM – Multimedia Communications Lab 34

Statistical Significance Tests – LeaveRTOut

More effective than à

Popularity Folk Rank

GFolkRank

AScore GRank AInheritScore

Poularity FolkRank X X X GFolkRank X X X X AScore X X X X X GRank X X AInheritScore X

Significance matrix of pair-wise comparisons of LeaveRTOut results Based on Average Precision with a significance level of p = 0.05

KOM – Multimedia Communications Lab 35

Adapted PageRank

!

!

!"

"

"

"

"

# #

$%&'()*+,& Tango0

Buenos

Aires0

Buenos

Aires0

Dancing

Festival0

1

"-.

#-.

#-.

"-.

PageRank‘s intelligent surfer model The ranking of a node is determined by how often the surfer visits the node Adjoining edges are followed with a certain probability – determined by the edge weights The query node acts as the starting point and focus i.e. the surfer returns to this node with a certain probability – determined by the node weights

[Hotho et al. 2006]