Symbolic Data Analysis Tools Symbolic Data Analysis Tools for - - PowerPoint PPT Presentation

▶
symbolic data analysis tools symbolic data analysis tools
SMART_READER_LITE
LIVE PREVIEW

Symbolic Data Analysis Tools Symbolic Data Analysis Tools for - - PowerPoint PPT Presentation

Feb 25, 2024 357 likes •714 views

Symbolic Data Analysis Tools Symbolic Data Analysis Tools for Recommendation Systems for Recommendation Systems Byron Leite Dantas Bezerra Byron Leite Dantas Bezerra bldb@dsc.upe.br bldb@dsc.upe.br Francisco Assis Tenrio Carvalho

slide-1
SLIDE 1

Byron Leite Dantas Bezerra Byron Leite Dantas Bezerra

bldb@dsc.upe.br bldb@dsc.upe.br

Francisco Assis Tenório Carvalho Francisco Assis Tenório Carvalho

fatc@cin.ufpe.br fatc@cin.ufpe.br Centro de Informática Centro de Informática – – CIn CIn/UFPE /UFPE Recife Recife – – Pernambuco Pernambuco – – Brasil Brasil

Symbolic Data Analysis Tools Symbolic Data Analysis Tools for Recommendation Systems for Recommendation Systems

slide-2
SLIDE 2
  • !"#$

!"#$

slide-3
SLIDE 3
  • %

% %

!"#$ !"#$

Summary Summary

  • Problem Definition
  • Personalization based on Symbolic Data

Analysis

– CMBF – SMCF – HMBF

  • Experimental Evaluation
slide-4
SLIDE 4
  • &

& &

!"#$ !"#$

Problem Problem Definition Definition

  • Key issues

– Which kind of information should be added in the user profile? – How to acquire information about the user preference? – How to represent the user profile in computer memory? – How to recommend items to the user based on his profile? – How much information we need about the user in order to delivery good recommendations? – How Symbolic Data Analysis can be a powerful tool to the Recommendation Systems field?

slide-5
SLIDE 5
  • '

' '

!"#$ !"#$

Case Case Study Study

.: .: Movie Movie Recommendation Recommendation :. :.

  • Items attributes
  • Rates matrix

G3 A1,A2,A7,A8 D2 F5 G1 A3,A5,A6,A7 D3 F4 G3 A2,A3,A7,A8 D7 F3 G2 A4,A6,A8,A9 D5 F2 G1 A1,A3,A4,A5 D3 F1 Genre Cast Director Movie

5 ∅ ∅ ∅ ∅ 4 ∅ ∅ ∅ ∅ 4 Vanessa 5 3 5 4 1 Elaine 5 ∅ ∅ ∅ ∅ 2 2 3 Bryan ∅ ∅ ∅ ∅ 5 2 ∅ ∅ ∅ ∅ 5 Brícia

F5 F4 F3 F2 F1

slide-6
SLIDE 6
  • !"#$

!"#$

Personalization Personalization based based on

  • n

Modal Modal Symbolic Symbolic Profiles Profiles

C Content

  • ntent M

Modal

  • dal

B Based ased F Filtering System iltering System

S Social

  • cial M

Modal

  • dal

C Collaborative

  • llaborative F

Filtering System iltering System

H Hybrid ybrid M Modal

  • dal

B Based ased F Filtering System iltering System

slide-7
SLIDE 7
  • (

( (

!"#$ !"#$

.: CMBF :. .: CMBF :. Content Modal Content Modal

Based Filtering System Based Filtering System

slide-8
SLIDE 8
  • )

) )

!"#$ !"#$

.: CMBF :. .: CMBF :.

  • Steps:
  • 1. Build the user profile
  • 1. Pre-processing
  • 2. Generalization
  • 2. Compare the modal symbolic user profile with

the symbolic description of each item in the target repository

  • 3. Build a personalized list to the user based on

the similarity scores obtained in the previous step

slide-9
SLIDE 9
  • !"#$

!"#$

.: CMBF :. .: CMBF :.

( (Step Step 1.1 1.1 -

  • preprocessing

preprocessing) )

G3 A1,A2,A7,A8 D2 F5 G1 A3,A5,A6,A7 D3 F4 G3 A2,A3,A7,A8 D7 F3 G2 A4,A6,A8,A9 D5 F2 G1 A1,A3,A4,A5 D3 F1 Gênero Elenco Diretor Filme ({G3},(1.0)) ({A1,A2,A7,A8},(¼, ¼, ¼, ¼)) ({D2},(1.0)) F5 ({G1},(1.0)) ({A3,A5,A6,A7},(¼, ¼, ¼, ¼)) ({D3},(1.0)) F4 ({G3},(1.0)) ({A2,A3,A7,A8},(¼, ¼, ¼, ¼)) ({D7},(1.0)) F3 ({G2},(1.0)) ({A4,A6,A8,A9},(¼, ¼, ¼, ¼)) ({D5},(1.0)) F2 ({G1},(1.0)) ({A1,A3,A4,A5},(¼, ¼, ¼, ¼)) ({D3},(1.0)) F1 Filme

Director Fi

X ~

Cast Fi

X ~

Genre Fi

X ~

slide-10
SLIDE 10
  • *

* *

!"#$ !"#$

.: CMBF :. .: CMBF :.

( (Step Step 1.2 1.2 -

  • generalization

generalization) )

  • For each level of user preference available in

the system, we associate a set of modal set of modal symbolic variables symbolic variables, where each variable concerns with some attribute in the movie domain

1 2 3 4 5

slide-11
SLIDE 11
  • **

** **

!"#$ !"#$

.: CMBF :. .: CMBF :.

( (Step Step 1.2 1.2 -

  • generalization

generalization) )

({G3},(1.0)) ({A1,A2,A7,A8},(¼, ¼, ¼, ¼)) ({D2},(1.0))

∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅

({G1},(1.0)) ({A1,A3,A4,A5},(¼, ¼, ¼, ¼)) ({D3},(1.0)) ({G2,G3},(½,½)) ({A2,A3,A4,A6,A7,A8,A9}, (1/8,1/8,1/8,1/8,1/8,1/4,1/8)) ({D5,D7},(½,½))

∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅ ∅

Genre Cast Director Movie

1

Bryan

y

2

Bryan

y

3

Bryan

y

4

Bryan

y

5

Bryan

y

5 ∅ ∅ ∅ ∅ 2 2 3 Bryan

F5 F4 F3 F2 F1

slide-12
SLIDE 12
  • *

* *

!"#$ !"#$

.: CMBF :. .: CMBF :.

Comparing Comparing the the user user profile profile with with some item some item ( (Step Step 2) 2)

  • Preprocess each target item i

i in the repository, building the modal symbolic descriptions of each one

  • The following function measures the similarity between the

user profile u u and the item i i

where

slide-13
SLIDE 13
  • *%

*% *%

!"#$ !"#$

.: CMBF :. .: CMBF :.

Comparing Comparing the the user user profile profile with with some item some item ( (Step Step 2) 2)

  • The dissimilarity function φ takes into account the

differences in the support support and the associated weight distributions weight distributions.

[ ]

  • =

+ =

p j j g j cd j g j cf i u

i q u q i S u S p x y

g

1

)) ( ), ( ( )) ( ), ( ( 2 1 1 ) ~ , ( φ φ φ

slide-14
SLIDE 14
  • *&

*& *&

!"#$ !"#$

.: CMBF :. .: CMBF :.

Suggesting Items (Step 3) Suggesting Items (Step 3)

  • Sort the items of the repository according to

their respective scores produced in the previous step.

slide-15
SLIDE 15
  • *'

*' *'

!"#$ !"#$

.: HMBF :. .: HMBF :. Hybrid Modal Hybrid Modal

Based Filtering System Based Filtering System

slide-16
SLIDE 16
  • *

* *

!"#$ !"#$

.: HMBF :. .: HMBF :.

  • Steps:
  • 1. Build the user profile
  • 2. Compute the similarity between the active

user and other users in the community

  • 3. Select the h nearest neighbors
  • 4. Build the personalized list for the active

user based on the suggestions of his neighbors

slide-17
SLIDE 17
  • *(

*( *(

!"#$ !"#$

.: HMBF :. .: HMBF :.

Comparing Comparing User User Profiles Profiles ( (Step Step 2) 2)

  • The comparison between u e v is

accomplished through the similarity function:

  • Where:
slide-18
SLIDE 18
  • *)

*) *)

!"#$ !"#$

1 2 3 4 5

  • Ψ

Ψ

φ φ φ φ φ After execution of step 2 for all users, we go forward with step 3 by selecting the h best users according Ψ Ψ Ψ Ψ.

.: HMBF :. .: HMBF :.

Comparing Comparing User User Profiles Profiles ( (Step Step 2) 2)

slide-19
SLIDE 19
  • *

* *

!"#$ !"#$

.: HMBF :. .: HMBF :.

Suggesting Items Suggesting Items ( (Step Step 4) 4)

  • After computing the neighborhood of the active user, we

select the most valued items for this users and recommend them for the active user, taking into account the level of similarity between the active user and a given neighbor.

slide-20
SLIDE 20
  • !"#$

!"#$

.: SMCF :. .: SMCF :. Social Modal Social Modal

Collaborative Filtering System Collaborative Filtering System

slide-21
SLIDE 21
  • *

* *

!"#$ !"#$

.: SMCF :. .: SMCF :.

  • Steps:
  • 1. Build the user profile
  • 2. Compute the similarity between the

active user and other users in the community

  • 3. Select the h nearest neighbors
  • 4. Build the personalized list for the active

user based on the suggestions of his neighbors

HMBF HMBF ≠ ≠ ≠ ≠ ≠ ≠ ≠ ≠ SMCF SMCF HMBF HMBF = = SMCF SMCF

slide-22
SLIDE 22
  • !"#$

!"#$

.: SMCF :. .: SMCF :.

5 ∅ ∅ ∅ ∅ 4 ∅ ∅ ∅ ∅ 4 Vanessa 5 3 5 4 1 Elaine 5 ∅ ∅ ∅ ∅ 2 2 3 Bryan ∅ ∅ ∅ ∅ 5 2 ∅ ∅ ∅ ∅ 5 Brícia

F5 F4 F3 F2 F1

0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% 1 2 3 4 5

slide-23
SLIDE 23
  • %

% %

!"#$ !"#$

.: SMCF :. .: SMCF :.

( (Step Step 1.1 1.1 – – preprocessing preprocessing) )

5 ∅ ∅ ∅ ∅ 4 ∅ ∅ ∅ ∅ 4 Vanessa 5 3 5 4 1 Elaine 5 ∅ ∅ ∅ ∅ 2 2 3 Bryan ∅ ∅ ∅ ∅ 5 2 ∅ ∅ ∅ ∅ 5 Brícia

F5 F4 F3 F2 F1

slide-24
SLIDE 24
  • &

& &

!"#$ !"#$

.: SMCF :. .: SMCF :.

( (Step Step 1.1 1.1 – – preprocessing preprocessing) )

5 ∅ ∅ ∅ ∅ 4 ∅ ∅ ∅ ∅ 4 Vanessa 5 3 5 4 1 Elaine 5 ∅ ∅ ∅ ∅ 2 2 3 Bryan ∅ ∅ ∅ ∅ 5 2 ∅ ∅ ∅ ∅ 5 Brícia

F5 F4 F3 F2 F1

slide-25
SLIDE 25
  • '

' '

!"#$ !"#$

.: SMCF :. .: SMCF :.

( (Step Step 1.2 1.2 -

  • generalization

generalization) )

5 ∅ ∅ ∅ ∅ 2 2 3 Bryan

F5 F4 F3 F2 F1

slide-26
SLIDE 26
  • !"#$

!"#$

1 2 3 4 5

  • Ψ

Ψ

φ φ φ φ φ After execution of step 2 for all users, we go forward with step 3 by selecting the h best users according Ψ Ψ Ψ Ψ.

.: SMCF :. .: SMCF :.

Comparing Comparing User User Profiles Profiles ( (Step Step 2) 2)

slide-27
SLIDE 27
  • (

( (

!"#$ !"#$

.: SMCF :. .: SMCF :.

Suggesting Suggesting Items Items ( (Step Step 4) 4)

  • Similar to step 4 of the HMBF!
slide-28
SLIDE 28
  • )

) )

!"#$ !"#$

Experimental Experimental Evaluation Evaluation

slide-29
SLIDE 29
  • !"#$

!"#$

Experimental Evaluation Experimental Evaluation

  • Objective: investigate the quality of personalized

recommendations taking into account the following issues

– Size of the user community – Number of items in the user profile – User preferences aquiring process

  • Domain: movie recommendations
  • Database: EachMovie + IMDB
  • Methods: CMBF, SMCF, HMBF, kNN-CF, kNN-CB
  • Metrics: half-life, precision, recall, f-measure, speed
  • Methodology

– inverted 10-fold cross validation X holdout

slide-30
SLIDE 30
  • %

% %

!"#$ !"#$

Experimental Evaluation Experimental Evaluation

slide-31
SLIDE 31

Byron Leite Dantas Bezerra Byron Leite Dantas Bezerra

bldb@dsc.upe.br bldb@dsc.upe.br

Francisco Assis Tenório Carvalho Francisco Assis Tenório Carvalho

fatc@cin.ufpe.br fatc@cin.ufpe.br Centro de Informática Centro de Informática – – CIn CIn/UFPE /UFPE Recife Recife – – Pernambuco Pernambuco – – Brasil Brasil

Symbolic Data Analysis Tools for Symbolic Data Analysis Tools for Recommendation Systems Recommendation Systems

slide-32
SLIDE 32
  • %

% %

!"#$ !"#$

Avaliação Experimental Avaliação Experimental

slide-33
SLIDE 33
  • %%

%% %%

!"#$ !"#$

Avaliação Experimental Avaliação Experimental

slide-34
SLIDE 34
  • %&

%& %&

!"#$ !"#$

Avaliação Experimental Avaliação Experimental