phd defense

54
Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions Recommender Systems for Social Tagging Systems Leandro Balby Marinho Machine Learning Lab University of Hildesheim PhD Defense Leandro Balby Marinho 1 / 32 Machine Learning Lab, University of Hildesheim

Upload: leandro-marinho

Post on 17-Jul-2015

91 views

Category:

Education


4 download

TRANSCRIPT

Page 1: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Recommender Systems for Social Tagging Systems

Leandro Balby Marinho

Machine Learning LabUniversity of Hildesheim

PhD Defense

Leandro Balby Marinho 1 / 32 Machine Learning Lab, University of Hildesheim

Page 2: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Outline

1. Motivation

2. Problems and Contributions

3. Tag Recommender Systems

4. Nearest Neighbor-based Tag Recommendation

5. Cross-Tagging

6. Tag Enrichment

7. Conclusions and Future Work

Leandro Balby Marinho 2 / 32 Machine Learning Lab, University of Hildesheim

Page 3: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

I Web 2.0 sites more used than e-mail! [Nielsen Online (2009)]

I In Web 2.0, the user plays the main role!

Leandro Balby Marinho 2 / 32 Machine Learning Lab, University of Hildesheim

Page 4: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

I Tags help users to organize and retrieve content.

Leandro Balby Marinho 3 / 32 Machine Learning Lab, University of Hildesheim

Page 5: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

I Tags also help other users to organize and retrieve their content.

Leandro Balby Marinho 4 / 32 Machine Learning Lab, University of Hildesheim

Page 6: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Folksonomy

I A folksonomy is a structure F := (U,R,T ,Y )

I U ... users

I R ... resources

I T ... tags

I Y ⊆ U × R × T ... tag assignments

I X := {(u, r) | ∃t ∈ T : (u, r , t) ∈ Y } ... set of posts

Leandro Balby Marinho 5 / 32 Machine Learning Lab, University of Hildesheim

Page 7: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Outline

1. Motivation

2. Problems and Contributions

3. Tag Recommender Systems

4. Nearest Neighbor-based Tag Recommendation

5. Cross-Tagging

6. Tag Enrichment

7. Conclusions and Future Work

Leandro Balby Marinho 6 / 32 Machine Learning Lab, University of Hildesheim

Page 8: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Problems and Contributions

I Tag Sparsity: Users are lazy to tag!

I 1− |Y ||U|×|R|×|T | ≈ 0.99 in all datasets used!

I Solution: Tag Recommendation

I Social Network Divide: Compatible social systems are disconnected.

I Tag Idiosyncrasy: Tags bearing unclear semantics.

I Solution: Tag Enrichment.

Leandro Balby Marinho 6 / 32 Machine Learning Lab, University of Hildesheim

Page 9: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Problems and Contributions

I Tag Sparsity: Users are lazy to tag!

I 1− |Y ||U|×|R|×|T | ≈ 0.99 in all datasets used!

I Solution: Tag Recommendation

I Social Network Divide: Compatible social systems are disconnected.

I Tag Idiosyncrasy: Tags bearing unclear semantics.

I Solution: Tag Enrichment.

Leandro Balby Marinho 6 / 32 Machine Learning Lab, University of Hildesheim

Page 10: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Problems and Contributions

I Tag Sparsity: Users are lazy to tag!

I 1− |Y ||U|×|R|×|T | ≈ 0.99 in all datasets used!

I Solution: Tag Recommendation

I Social Network Divide: Compatible social systems are disconnected.

I Tag Idiosyncrasy: Tags bearing unclear semantics.

I Solution: Tag Enrichment.

Leandro Balby Marinho 6 / 32 Machine Learning Lab, University of Hildesheim

Page 11: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Outline

1. Motivation

2. Problems and Contributions

3. Tag Recommender Systems

4. Nearest Neighbor-based Tag Recommendation

5. Cross-Tagging

6. Tag Enrichment

7. Conclusions and Future Work

Leandro Balby Marinho 7 / 32 Machine Learning Lab, University of Hildesheim

Page 12: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Tag Recommender Systems

I ...change the process from creation to recognition!

I Personalized methods take the user preferences for tags intoconsideration.

I Value for the industry, e.g., youtube, flickr, last.fm, amazon.

Leandro Balby Marinho 7 / 32 Machine Learning Lab, University of Hildesheim

Page 13: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Evaluation and Metric

I Xtrain∪Xtest = X ... train/test splits based on posts

I For each user, randomly pick one post for test.

I Task: For (u, r) ∈ Xtest compute T (u, r)

I Metric: Recall((u, r) ∈ Xtest, n) := |T (u,r)∩T (u,r)||T (u,r)|

Leandro Balby Marinho 8 / 32 Machine Learning Lab, University of Hildesheim

Page 14: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Formalization

I Given (u, r) ∈ Xtest, a tag recommender system first computes:

Utility : {u} × {r} × T → R (1)

I And then presents the tags in descending order of their utility:

T (u, r) :=n

argmaxt∈T

Utility(u, r , t) (2)

Leandro Balby Marinho 9 / 32 Machine Learning Lab, University of Hildesheim

Page 15: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Outline

1. Motivation

2. Problems and Contributions

3. Tag Recommender Systems

4. Nearest Neighbor-based Tag Recommendation

5. Cross-Tagging

6. Tag Enrichment

7. Conclusions and Future Work

Leandro Balby Marinho 10 / 32 Machine Learning Lab, University of Hildesheim

Page 16: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Nearest Neighbor-based (NN) Tag Recommenders

I Collaborative Filtering (CF): Similar users tend to like similar things.

I Here: Similar users tend to tag alike.

I Traditional CF cannot be directly applied to folksonomies unless:

resources

tagsresources

user

s

user

s

user

s

tags

Y

πUTYπURY

Leandro Balby Marinho 10 / 32 Machine Learning Lab, University of Hildesheim

Page 17: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Collaborative Filtering for Tag Recommendation

I Neighborhood Formation: Nku :=

kargmaxv∈Ur\{u}

sim(~mu, ~mv )

I Recommendation:

T (u, r) :=n

argmaxt∈T

∑v∈Nk

u

sim(~mu, ~mv )δ(v , r , t)

where δ(v , r , t) := 1 if (v , r , t) ∈ Y and 0 else.

Leandro Balby Marinho 11 / 32 Machine Learning Lab, University of Hildesheim

Page 18: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Ensembles of CF

I Projections’ Ensemble:

I Similarities’ Ensemble:

T (u, r) =n

argmaxt∈T

∑v∈Nu

(λsim(~mu, ~mv ) + (1− λ)sim(~zu,~zv ))δ(v , r , t)

where ~mu and ~mv are rows of πUTY , and ~zu and ~zv rows of πURY .

Leandro Balby Marinho 12 / 32 Machine Learning Lab, University of Hildesheim

Page 19: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

A Graph-Based Tag Recommender based on Posts

We represent X as a homogeneous, undirected graph G := (X ,E ) overthe post set. Posts are related to each other if they share the same user:

Ruser := {(x , x ′) ∈ X × X | user(x) = user(x ′)}

the same resource:

Rres := {(x , x ′) ∈ X × X |res(x) = res(x ′)}

or either share the same user or resource:

Rresuser := Ruser ∪Rres

where user(x) and res(x) are the user and resource associated with the

post x respectively.

Leandro Balby Marinho 13 / 32 Machine Learning Lab, University of Hildesheim

Page 20: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Relational Graph based on Posts

Leandro Balby Marinho 14 / 32 Machine Learning Lab, University of Hildesheim

Page 21: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Weighting Schemes

For x ∈ Xtest and (x , x ′) ∈ E :

1. User-Tag Profile:

φuser-tag := (|Y ∩ ({user(x)} × R × {t})|)t∈T

2. Resource-Tag Profile:

φres-tag := (|Y ∩ (U × {res(x)} × {t})|)t∈T

Weight:

w(x , x ′) :=〈φ(x), φ(x ′)〉‖φ(x)‖‖φ(x ′)‖

Leandro Balby Marinho 15 / 32 Machine Learning Lab, University of Hildesheim

Page 22: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Relational Classification

Weighted Average (WA) [Marinho et al. (2009)]:

P(t|x) :=

∑x′∈Nx |t∈T (x′) w(x , x ′)∑

x′∈Nxw(x , x ′)

where:

Nx := {x ′ ∈ X | (x , x ′) ∈ R, T (x) 6= ∅}

Runtime: O (|T ||Nx |))

Leandro Balby Marinho 16 / 32 Machine Learning Lab, University of Hildesheim

Page 23: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Evaluation

Datasets:

dataset |U| |R| |T | Triples |Y | Posts |X |BibSonomy 116 361 412 10,148 2,522Last.fm 2,917 1,853 2,045 219,702 75,565Delicious 37,399 74,874 22,170 7,487,319 3,055,436

Evaluated methods:

I Baselines: (Locally) Constant Models (GCT,LCR, LCU).

I Ensemble of Locally Constant Models (LCE) [Jaschke et al. 2008].

I TopicRank, FolkRank [Jaschke et al. 2007]

I RTF [Rendle et al. 2009]

I PITF [Rendle et al. 2010]

I Our NN-based Recommenders

Leandro Balby Marinho 17 / 32 Machine Learning Lab, University of Hildesheim

Page 24: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Results: NN Methods

0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10

Rec

all

Number of recommended tags

Top-10 Tag Recommendations in Delicious

WACF UTCF UR

matrixExtsimEns

LCRGCT

Leandro Balby Marinho 18 / 32 Machine Learning Lab, University of Hildesheim

Page 25: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Results: WA vs. State-of-the-Art

0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10

Rec

all

Number of recommended tags

Top-10 Tag Recommendations in BibSonomy

WARTFPITF

FolkRankLCE

TopicRank

Leandro Balby Marinho 19 / 32 Machine Learning Lab, University of Hildesheim

Page 26: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Results: WA vs. State-of-the-Art

0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10

Rec

all

Number of recommended tags

Top-10 Tag Recommendations in Last.fm

PITFWA

RTFFolkRank

LCETopicRank

Leandro Balby Marinho 20 / 32 Machine Learning Lab, University of Hildesheim

Page 27: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Results: WA vs. State-of-the-Art

0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10

Rec

all

Number of recommended tags

Top-10 Tag Recommendations in Delicious

PITFWA

FolkRankLCE

TopicRank

Leandro Balby Marinho 21 / 32 Machine Learning Lab, University of Hildesheim

Page 28: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Runtime: WA vs. PITF

BibSonomy Last.fm DeliciousMethod Runtime Runtime Runtime

WA < 1 second < 1 minute ≈ 3 minutes

PITF ≈ 5 minutes ≈ 7 hours ≈ 33 days

Leandro Balby Marinho 22 / 32 Machine Learning Lab, University of Hildesheim

Page 29: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

ECML/Discovery Challenge 2009

2nd Place ECML/PKDD Discovery Challenge 2009!

Rank Method Top-5 F1

1 PITF [Rendle et al. (2009)] 0.355942 Relational Ensemble [Marinho et al. (2009)]1 0.33185– WA (not submitted) 0.325193 Content-based [Lipczak et al. (2009)] 0.32461

1With Christine PreisachLeandro Balby Marinho 23 / 32 Machine Learning Lab, University of Hildesheim

Page 30: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Outline

1. Motivation

2. Problems and Contributions

3. Tag Recommender Systems

4. Nearest Neighbor-based Tag Recommendation

5. Cross-Tagging

6. Tag Enrichment

7. Conclusions and Future Work

Leandro Balby Marinho 24 / 32 Machine Learning Lab, University of Hildesheim

Page 31: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Problem

Use resources overlap to cross tags between systems.

Leandro Balby Marinho 24 / 32 Machine Learning Lab, University of Hildesheim

Page 32: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Tag Recommendation for Cross-Tagging

I Cross-Tagging Approaches:

I LCR (locally constant per resource).I Collaborative Filtering.

Leandro Balby Marinho 25 / 32 Machine Learning Lab, University of Hildesheim

Page 33: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Evaluation

Tag-Aware-based Evaluation

I The better the tags the better a tag-aware recommender that usesthose tags.

I Tag-Aware based on HOSVD [Symeonidis et al. (2008)]

Datasets

Blogger.com Last.fm Annotated Blog

|U| 6,620 44,143 3,827|R| 17,372 17,372 1,323|T | 0 4,903 422|Y | 0 254,388 32,900

Leandro Balby Marinho 26 / 32 Machine Learning Lab, University of Hildesheim

Page 34: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Recall on the top-5 resources of HOSVD

I n - Number of tags used to annotate the test posts of Blogger.com.

Leandro Balby Marinho 27 / 32 Machine Learning Lab, University of Hildesheim

Page 35: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Outline

1. Motivation

2. Problems and Contributions

3. Tag Recommender Systems

4. Nearest Neighbor-based Tag Recommendation

5. Cross-Tagging

6. Tag Enrichment

7. Conclusions and Future Work

Leandro Balby Marinho 28 / 32 Machine Learning Lab, University of Hildesheim

Page 36: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Problems

I First we map tags from a folksonomy to concepts C of an ontology

H : T → C

I Then we learn an ontology P such that:

CP := T ∪ C

I The better the ontology the better a ontology-aware recommenderthat uses this ontology.

I Taxonomy driven CF [Ziegler et al. (2004)]

Datasets:

dataset |U| |T | |R| |Y |Last.fm 3,532 7,081 982 130,899musicmoz - 555 982 -

Leandro Balby Marinho 28 / 32 Machine Learning Lab, University of Hildesheim

Page 37: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Results

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

Trivial Ontology Domain Expert Ontology Learned Ontology

Rec

all

Leandro Balby Marinho 29 / 32 Machine Learning Lab, University of Hildesheim

Page 38: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Outline

1. Motivation

2. Problems and Contributions

3. Tag Recommender Systems

4. Nearest Neighbor-based Tag Recommendation

5. Cross-Tagging

6. Tag Enrichment

7. Conclusions and Future Work

Leandro Balby Marinho 30 / 32 Machine Learning Lab, University of Hildesheim

Page 39: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Conclusions

I Tag Sparsity: Nearest Neighbor Method that

I Performs competitively to more sophisticated methods.I Require modest computational effort.

I Social Network Divide:

I Cross-tagging as a tag recommendation problem.I Personalized cross-tagging better than non-personalized

cross-tagging.

I Tag idiosyncrasy: Tag enrichment

I Well agreed concepts that match the semantic intention ofusers.

I Learned ontology better than trivial or domain expert ontology.

I New recommender systems-based evaluation protocols.

Leandro Balby Marinho 30 / 32 Machine Learning Lab, University of Hildesheim

Page 40: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Future Work

I Optimzed weight learning for WA.

I Bidirectional Cross-Tagging.

I Optimized Cross-Tagging/Ontology learning.

Leandro Balby Marinho 31 / 32 Machine Learning Lab, University of Hildesheim

Page 41: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Results NN vs. Baselines

0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10

Rec

all

Number of recommended tags

Top-10 Tag Recommendations in BibSonomy

WACF UTCF UR

matrixExtsimEns

LCRGCT

0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10R

ecal

l

Number of recommended tags

Top-10 Tag Recommendations in Last.fm

WACF UTCF UR

matrixExtsimEns

LCRGCT

Leandro Balby Marinho 32 / 32 Machine Learning Lab, University of Hildesheim

Page 42: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

PageRank for Folksonomias

I Based on PageRank [Hotho et al. 2006]

I Each hyperedge is broken into three undirected edges:

I Now PageRank can be applied:

~wt+1 ← λAT~wt + (1− λ)~p

I Rank will be dominated by popular nodes (Skewd distribution of tagassignments)

Leandro Balby Marinho 32 / 32 Machine Learning Lab, University of Hildesheim

Page 43: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

FolkRank

1. First compute vector ~w (0) with ~p = 1.

2. Next compute vector ~w (1) with ~p[u] := 1 + |U|, ~p[r ] := 1 + |R|, and~p[v ] := 1 for v 6= u, r .

3. Finally compute ~w := ~w (1) − ~w (0).

4. Recommendation list T (u, r) is the top-n nodes in the rankrestricted to tags.

Leandro Balby Marinho 32 / 32 Machine Learning Lab, University of Hildesheim

Page 44: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

RTF: Ranking with Tensor Factorization

I Tag Recommendation as a tensor completion problem.

I Positive tags have higher rank than negative ones [Rendle et al. 2009].

yu,r,t1 > yu,r,t2 ⇔ (u, r , t1) ∈ T+u,r ∧ (u, r , t2) ∈ T−u,r

T+u,r := {t | (u, r) ∈ Xtreino ∧ (u, r , t) ∈ Y }, T−u,r := {t | (u, r) ∈ Xtreino ∧ (u, r , t) 6∈ Y }

Leandro Balby Marinho 32 / 32 Machine Learning Lab, University of Hildesheim

Page 45: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Tucker Decomposition Model

Y := C ×u U ×r R ×t T

or equivalently:

yu,r ,t =∑

u

∑r

∑t

cu,r ,t · uu,u · rr ,r · tt,t

where the model parameters are:

C ∈ RkU×kR×kT , U ∈ R|U|×kU , R ∈ R|R|×kR , T ∈ R|T |×kT

Leandro Balby Marinho 32 / 32 Machine Learning Lab, University of Hildesheim

Page 46: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

PITF: Pairwise Interaction Tensor Factorization

PITF only models the two-way interactions between user and tags as wellas between resources and tags:

au,r ,t =k∑f

uu,f · tUt,f +

k∑f

rr ,f · tRt,f

where U ∈ R|U|×k , R ∈ R|R|×k , TU ∈ R|T |×k and TR ∈ R|T |×k

Leandro Balby Marinho 32 / 32 Machine Learning Lab, University of Hildesheim

Page 47: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Complexity

Learning Runtime Complexity

Method RuntimeWA O(1)FolkRank O(1)RTF O

(iter · |Xtrain||T |2 · kU · kR · kT

)PITF O(iter · |Xtrain||T |2 · 2k)

Prediction Runtime Complexity

Method RuntimeWA O (|T ||Nx |+ |T | log(n)))Folkrank O(iter · (|Y |+ |U|+ |R|+ |T |) + |T |+ |T | log(n))RTF O(|T | · kU + kR · kT · kT )PITF O(|T |2k + |T | log(n))

Leandro Balby Marinho 32 / 32 Machine Learning Lab, University of Hildesheim

Page 48: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Relation Rewarding

We can reward the best relation by a factor c ∈ R

Leandro Balby Marinho 32 / 32 Machine Learning Lab, University of Hildesheim

Page 49: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Results Cross-Tagging

Leandro Balby Marinho 32 / 32 Machine Learning Lab, University of Hildesheim

Page 50: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Tag Enrichment Approach

I Semantic mapping as an ontology matching problem.

I P(A,B) ≈ |JAK∩JBK||R| [Doan et al. (2004)]

I Jaccard coefficient:

JS(A,B) := P(A ∩ B)/P(A ∪ B) :=P(A,B)

P(A,B) + P(A, B) + P(A,B)

Leandro Balby Marinho 32 / 32 Machine Learning Lab, University of Hildesheim

Page 51: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Ontology learning

I Frequent itemset mining for ontology learning [Marinho et al. 2008]2.

2Algorithm proposed by Krisztian Buza co-author of [Marinho et al. 2008]Leandro Balby Marinho 32 / 32 Machine Learning Lab, University of Hildesheim

Page 52: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

Semantic mapping

tags mapped concepts

electro electronica

hip hop hip hop

chillout rock

old skool dance house

anything else but death heavy metal

post-hardcore emo

california punk

political punk

urban hip hop

60s stuff country-rock

relaxing folk rock

explorer experimental rock

rock en espanol latin pop

Leandro Balby Marinho 32 / 32 Machine Learning Lab, University of Hildesheim

Page 53: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

An Extract of Domain Expert Ontology

heavy_metal

death_metal

doom_metal

black_metal

thrash rap-metal

hair_metal

speed_metal

grindcore

metal

Pajek

Leandro Balby Marinho 32 / 32 Machine Learning Lab, University of Hildesheim

Page 54: PhD Defense

Motivation Problems Tag Recommender Systems NN Tag Recommendation Cross-Tagging Tag Enrichment Conclusions

An Extract of Learned Ontology

maynard james keenan

powerful

technical death metal

brit-rock

metalcore

nu metal

doom metal

new age

finnishprogressive metal

alternative metal

melodic death metal

melancholic

black metal

progressive

ethereal

swedish metal

gothenburg metal

progressive death metal

german

bands i have seen live

speed metal

nwobhm

heavy

power metal

symphonic metal

guitargasm

death-doom metal

gothic metal

famous frontman

art rock

viking metalgroove metal

melodic metal

violent

aggressive alternative - at work music

moody

faves

a-o-t-w

slipknot

grindcore

great lyrics

gothenburg

dark

g00ds

70s progressive rock

depressing

cold

doom

art-rock

prog

trash metal

depression

brutal death metal

usloud

sad

korn

soad

mezmerize

fall out boy

rap-metal

seen them live

nu-metal

cello metal

melodic black metal

folk metal

guitar music

symphonic prog

british metal

awesome

zeuhl

female fronted metal

love metal

aggressive

finland

epic

nellis1

symphonic black metal

new metal

ominous

buen metal

bands ive seen live

classic thrash

bands i have seen

prog metal

classic metal

prog rock metal gods

my band inspiration

metal of some persuasionfavorite shitnice music

grooving metal

fav artistsblizzards main tags

symphonic death

grind

melodic power metal

everything

speed

favs

melodic death

heavy_metal

death

progressive_rock

doom_metal

death_metal

thrash

metal

speed_metal

periods

Pajek

Leandro Balby Marinho 32 / 32 Machine Learning Lab, University of Hildesheim