content modelling for recommendation - consorcio mavir · content modelling for recommendation...
TRANSCRIPT
Content ModellingExperimental Environment: RepLab 2013
• Detect topics in tweets
• 61 Entities in 4 domains
• Annotated Training Set + Test Set to Classify– Clustering Task
• Reflect the Previous Knowledge about the Entities ¿?
• Set the number of topics ¿?
• Evaluation– Entity Level
– Reliability & Sensitivity
Content ModellingFCA-Based Content Modelling & Organization
• Unsupervised approach
• Avoid clustering problems
• Adaptation to new topics, but taking into account previous knowledge
FCA Overview [Wille, 1992]
Extent Intent
C1 {Doc1, Doc2, Doc3, Doc4} {∅}
C2 {Doc1, Doc2, Doc3} {P}
C3 {Doc1, Doc4} {J}
C4 {Doc1} {P, J, PJ}
C5 {Doc2} {P, NP}
C6 {Doc3} {P, AP}
C7 {∅} {P, NP, AP, J, PJ}
P NP AP J PY
Doc1 X X X
Doc2 X X
Doc3 X X
Doc4 X
Formal Context: 𝕂 ∶= (𝑮,𝑴, 𝑰)
• 𝑮: tweets
• 𝑴: terms (hashtag, url, word) in the tweets
• 𝑰 tweet 𝒈 has the term 𝒎
Term1Term2 Term3 Term4
Tweet1 X X X
Tweet2 X X
Formal Context Generation
Concept Lattice Generation
Term1Term2 Term3 Term4 Term5 Term6 Term7 Term8
Tweet1 X
Tweet2 X X X X
Tweet3 X X X X
Tweet4 X
Tweet5 X X X X
Tweet6 X X X X
Tweet7 X X X X
Topic Selection
Stability
𝜎𝑖 𝐴, 𝐵 =| 𝐶 ⊆ 𝐴 𝐶′ = 𝐵}|
2 𝐴
where 𝐴 is the number of objects in 𝐴 and 𝐶 is each subset of 𝐴 whoseconcept’s intent (𝐶′) is equal to the concept intent of 𝐴, that is, 𝐶′ = 𝐵.
FilteringManaging of Noisy Data
• Does it really affects? Sure, but how much?– Is it valuable to take into account?
• Comparison:– KLD-Filtering
• Initial bad-performing approach
– Best-performing case
Reliability Sensitivity F(R,S)
KLD-based Filtering [0,6735 - 0,8331] [0,1076 - 0,1092] [0,1548 - 0,1711]
Gold Standard Based Filtering [0,6184 - 0,6615] [0,1940 – 0,2469] [0,1730 - 0,2336]
Code Bug Fixing
Reliability Sensitivity F(R,S)
With Code Bug [0,6184 - 0,6615] [0,1940 – 0,2469] [0,1730 - 0,2336]
Without Code Bug [0,1678 - 0,3021] [0,3343 – 06678] [0,2242 - 0,2882]
Attribute SelectionPre-selection of the top-representative attributes
• Knowledge lost or noise reduction?
• 2 Parameters: Lower &Upper Threshold
LT UT Reliability Sensitivity F(R,S)
1 50 0,3021 0,3343 0,2882
1 25 0,3029 0,3324 0,2878
1 10 0,3039 0,3311 0,2877
5 50 0,1678 0,6778 0,2242
5 25 0,1680 0,6746 0,2235
5 10 0,1685 0,6715 0,2236
Attribute SelectionOnly lower threshold really affects. ¿Why?
LowerThreshold
UpperThreshold
# of GeneratedConcepts Average Concepts by Entity
1 50 29836 489
1 25 31384 514
1 10 32566 533
5 50 1100 18
5 25 1154 18
5 10 1258 20
1 E+0
1 E+1
1 E+2
1 E+3
1 E+4
1 E+5
1 E+6
Nu
mb
er o
f A
ttri
bu
tes
(lo
g)
Frequency of Occurrence
5%1%
Attribute SelectionThe lower the threshold, the better the results ¿?
LT Reliability Sensitivity F(R,S)
5 0,1678 0,6778 0,2242
1 0,3021 0,3343 0,2882
0.5 0,3836 0,2412 0,2710
0.1 0,4075 0,2204 0,2671
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
5 1 0.5 0.1
F-measure
Reliability
Sensitivity
Topic Selection - StabilitySelect the most-suitable concepts
Stability Value Reliability Sensitivity F(R,S)
0.2 0,4090 0,3163 0,3258
0.4 0,4090 0,3163 0,3258
0.5 0,3455 0,3228 0,3041
0.7 0,3407 0,3236 0,3027
0.9 0,3029 0,3324 0,2878
BEST_REPLAB_APPROACH 0,4624 0,3246 0,3252
Topic AdaptationSame performance for seen & unseen topics¿?
Topics Reliability Sensitivity F(R,S)
All 0,4090 0,3163 0,3258
Seen 0,5764 0,3018 0,3730
Unseen 0,4504 0,3302 0,3447
FCA for Recommendation: SoftACFRS
• Entry Levels
• Association Rules
• Fuzzy FCA
CBRS
• Keyword-Based
• Item Modelling
Toy ExamplesNot really representatives
Don’t take advantage of the CBRS potential(features, keywords)
FCA for Recommendation: Basics
CFRS vs. CBRS = Model Users vs. Model Contents
FCA can infer concepts from contents but…
Can be inferred concepts from User preferences?
• FCA vs. Recommendation Exploratory vs. Predictive task
FCA for Recommendation: Basics
Intuition
• Content-Based
• Collaborative Filtering– Solution: Association rules ¿?
{milk, nappies} {beer}
FCA Recommendation ApproachesCollaborative Filtering: User vs. Items
• Take similar users
• Recommend new items related to these users
• Algorithm already developed: Based on the lattice structure
Content-Based: Content vs. Features
Hybrid Filtering: User vs. Features
Test Bed: FCA-based CFRSMovieLens• State of the Art Dataset
LDosCoMoDa• Context Information:
– Time, Daytype, Season, Location, Weather, Social, Mood, Physical,Decision and Interaction
• 122 Users, 1233 Items, 2300 ratings
Movie Tweetings• 100k Tweets rating IMDB movies
“I rated The Matrix 9/10 http://www.imdb.com/title/tt0133093/ #IMDb”
Test Bed: FCA-based CBRSPlista: 80GB Dataset
• Features: Contextual and Content-related
• User Interactions
DBBook
• Features: LD Information about the books
• Interactions: Users and a set of books
RERmovie
• Features: LD Information about the movie
• Interactions:– User-item ratings
– User-attributes ratings
Explanations in RecommendationChristian Scheel, Angel Castellanos, Thebin Lee, Ernesto WilliamDe Luca. The Reason Why: A Survey of Explanations forRecommender Systems
Attribute-Based ExplanationsYou like Mafia movies I recommend you “The Godfather”
SoftA
• Based on item ratings Infer wrong preferencesI like The GodfatherI like Marlon Brando but I don’t like Robert Duvall
Proposal
• Based on item attributes ratings
• Dataset: RERmovie
Dataset - RERmovieUser Study: Items ratings attribute by attribute
Some numbers:• 53 Users
• 650 ratings in 299 different movies
• 6597 reasons for movie ratings
ProposalExplanation Retrieval• Attribute rating = quality measure
– Based on the scores of the users.
• ∀𝒊: Item Model 𝒎𝒊 → 𝑨𝒊: Attribute set to explain 𝒊– 𝑨𝒊
+: 𝒂 ∈ 𝑨𝒊 where 𝒓𝒂 > 𝒕𝟏– 𝑨𝒊
−: 𝒂 ∈ 𝑨𝒊 where 𝒓𝒂 < 𝒕𝟐
• Explanations– ∀𝒊 : Pro (Pr) and contra (Cr) reasons
Evaluation• Pr and Cr are compared to the user feedback
– Measure: Precision and Recall and F1
Work LoadConcept Modelling
• FCA posed as a more-than-suitable approach
• Extensive data analysis
FCA 4 Recommendation
• FCA Algorithm almost developed
Explanations 4 Recommendation
• Novel approach presented
Reduction Algorithm (1)
S
G
AUX
ddddddddP
DDDDDDDDDDD
87654321
10987654321
,,,,,,,
,,,,,,,,,
76521
2
76521
8765431
109843
,,,,
,,,,
,,,,,,
,,,,
DDDDDG
dS
DDDDDAUX
dddddddP
DDDDDD
87654321
32
843
876541
109
,,,,,,,
,
,,
,,,,,
,
DDDDDDDDG
ddS
DDDAUX
ddddddP
DDD
10987654321
432
109
87651
,,,,,,,,,
,,
,
,,,,
DDDDDDDDDDG
dddS
DDAUX
dddddP
D
Reduction Algorithm (2)
10987654321
432
109
87651
,,,,,,,,,
,,
,
,,,,
DDDDDDDDDDG
dddS
DDAUX
dddddP
D
RERmovie Data AnalysisMost prominent Attributes
• Positives: Genres (35%), Actors (31%), Director (6%)
• Negatives: Actors (26%), Genres (18%)
Most ratings are positives…
255
217
103
48
27