explod: a framework for explaining recommendations based on the linked open data cloud

35
ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud Cataldo Musto, Fedelucio Narducci, Pasquale Lops, Marco de Gemmis, Giovanni Semeraro Sept 15-19, 2016 - Boston BPN

Upload: fedelucio-narducci

Post on 09-Feb-2017

132 views

Category:

Internet


1 download

TRANSCRIPT

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data CloudCataldo Musto, Fedelucio Narducci, Pasquale Lops, Marco de Gemmis, Giovanni Semeraro

Sept 15-19, 2016 - BostonBPN

PROBLEM DEFINTION

➤ The recommendation process is sometimes viewed from the user side as an obscure process

➤ The recommender usually implements a black box model with no knowledge on its internal mechanisms

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

PROBLEM DEFINITION

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

PROBLEM DEFINITION

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

PROBLEM DEFINITION

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

I recommend

you

PROBLEM DEFINITION

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

I recommend

you

PROBLEM DEFINITION

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

I recommend

you

PROBLEM AND A (POSSIBLE) SOLUTION

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

PROBLEM AND A (POSSIBLE) SOLUTION

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

I recommend you Cloud Atlas because you often like films with

Tom Hanks as Saving Private Ryan and Da Vinci Code. In addition, you sometimes like films directed by

The Wachowskis as The Matrix.

EXPLOD: THE INTUITION

➤ connecting user preferences and recommendations through a set of LOD-based properties

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

ITEM ATTRIBUTES: ENTITY LINKING

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

EXPLOD: THE FRAMEWORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Mapper Builder

Ranker Generator

Profile Recommendations (potentially any algorithm)

LOD-aware represent.

Top-N properties

Graph Data Model

EXPLOD: THE FRAMEWORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Mapper Builder

Ranker Generator

Profile Recommendations

LOD-aware represent.

Top-N properties

Graph Data Modelrepresents user profiles and recommendations as DBpedia nodes

http://dbpedia.org/page/Cloud_Atlas_(film)

http://dbpedia.org/page/Saving_Private_Ryan

EXPLOD: THE FRAMEWORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Builder

Ranker GeneratorTop-N

propertiesbuilds a graph-based data model that exploits the information in the LOD cloud

Mapper

Profile Recommendations

LOD-aware represent.

Graph Data Model

American Epic Films

Tom Hanks

Dystopian Films

The Wachowskis

Saving Private Ryan

Da Vinci Code

The Matrix

Cloud Atals

Items in the user profiles (Ip) LOD-cloud properties (P)

Recommended Items (Ir)

dcterms:subject

dcterms:subject

dcterms:subject

dcterms:subjectdbpedia-owl:starring

dbpedia-owl:starringdbpedia-owl:starring

dbpedia-owl:director

dbpe

dia-

owl:d

irect

or

EXPLOD: THE FRAMEWORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Mapper Builder

Ranker Generator

Profile Recommendations

LOD-aware represent.

Top-N properties

Graph Data Model

selects properties which explain the recommendation

c = property Ip = items in the user profile Ir = recommendations nc, Ip = # edges connecting c to items in Ip nc, Ir = # edges connecting c to items in Ir IDFc = adaptation of IDF for c

MapPro Recomme

LOD-aware

Graph

C1

C2

C3

C4

Ip1

Ip2

Ip3

Ir1

dcterms:

dcterms:

dcterms:

dcterms:dbpedia-

dbpedia- dbpedia-

dbpedia-

dbpe

dia-

EXPLOD: THE FRAMEWORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Mapper Builder

Ranker Generator

Profile Recommendations

LOD-aware represent.

Top-N properties

Graph Data Model

selects properties which explain the recommendation

c = property Ip = items in the user profile Ir = recommendations nc, Ip = # edges connecting c to items in Ip nc, Ir = # edges connecting c to items in Ir IDFc = adaptation of IDF for c

MapPro Recomme

LOD-aware

Graph

C1

C2

C3

C4

Ip1

Ip2

Ip3

Ir1

dcterms:

dcterms:

dcterms:

dcterms:dbpedia-

dbpedia- dbpedia-

dbpedia-

dbpe

dia-

EXPLOD: THE FRAMEWORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Mapper Builder

Ranker Generator

Profile Recommendations

LOD-aware represent.

Top-N properties

Graph Data Model

selects properties which explain the recommendation

c = property Ip = items in the user profile Ir = recommendations nc, Ip = # edges connecting c to items in Ip nc, Ir = # edges connecting c to items in Ir IDFc = adaptation of IDF for c

MapPro Recomme

LOD-aware

Graph

C1

C2

C3

C4

Ip1

Ip2

Ip3

Ir1

dcterms:

dcterms:

dcterms:

dcterms:dbpedia-

dbpedia- dbpedia-

dbpedia-

dbpe

dia-

EXPLOD: THE FRAMEWORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Mapper Builder

Ranker Generator

Profile Recommendations

LOD-aware represent.

Top-N properties

Graph Data Model

selects properties which explain the recommendation

c = property Ip = items in the user profile Ir = recommendations nc, Ip = # edges connecting c to items in Ip nc, Ir = # edges connecting c to items in Ir IDFc = adaptation of IDF for c

MapPro Recomme

LOD-aware

Graph

C1

C2

C3

C4

Ip1

Ip2

Ip3

Ir1

dcterms:

dcterms:

dcterms:

dcterms:dbpedia-

dbpedia- dbpedia-

dbpedia-

dbpe

dia-

EXPLOD: THE FRAMEWORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Mapper Builder

Ranker Generator

Profile Recommendations

LOD-aware represent.

Top-N properties

Graph Data Model

selects properties which explain the recommendation

c = property Ip = items in the user profile Ir = recommendations nc, Ip = # edges connecting c to items in Ip nc, Ir = # edges connecting c to items in Ir IDFc = adaptation of IDF for c

MapPro Recomme

LOD-aware

Graph

C1

C2

C3

C4

Ip1

Ip2

Ip3

Ir1

dcterms:

dcterms:

dcterms:

dcterms:dbpedia-

dbpedia- dbpedia-

dbpedia-

dbpe

dia-

EXPLOD: THE FRAMEWORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Mapper Builder

Ranker Generator

Profile Recommendations

LOD-aware represent.

Top-N properties

Graph Data Model

selects properties which explain the recommendation

c = property Ip = items in the user profile Ir = recommendations nc, Ip = # edges connecting c to items in Ip nc, Ir = # edges connecting c to items in Ir IDFc = adaptation of IDF for c

MapPro Recomme

LOD-aware

Graph

C1

C2

C3

C4

Ip1

Ip2

Ip3

Ir1

dcterms:

dcterms:

dcterms:

dcterms:dbpedia-

dbpedia- dbpedia-

dbpedia-

dbpe

dia-

EXPLOD: THE FRAMEWORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Mapper Builder

Ranker Generator

Profile Recommendations

LOD-aware represent.

Top-N properties

Graph Data Model

generates the natural language explanation

EXPLOD: THE FRAMEWORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Mapper Builder

Ranker Generator

Profile Recommendations

LOD-aware represent.

Top-N properties

Graph Data Model

generates the natural language explanation

I recommend you Cloud Atlas because…

EXPLOD@WORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

EXPLOD@WORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

I recommend you Cloud Atlas because

EXPLOD@WORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Tom Hanks

I recommend you Cloud Atlas because you often like films with TomHanks as Saving Private Ryan and Da Vinci Code.

dbpedia

-owl:star

ring

dbpedia-owl:starring dbpedia-owl:starring

EXPLOD@WORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Tom Hanks

The Wachowskis

I recommend you Cloud Atlas because you often like films with TomHanks as Saving Private Ryan and Da Vinci Code. In addition, yousometimes like films directed by The Wachowskis as The Matrix.

dbpedia

-owl:star

ring

dbpedia-owl:starring

dbpedia-owl:director

dbpedia-ow

l:director

dbpedia-owl:starring

EXPLOD@WORK

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

American Epic Films

Tom Hanks

Dystopian Films

The Wachowskis

I recommend you Cloud Atlas because you often like films with TomHanks as Saving Private Ryan and Da Vinci Code. In addition, yousometimes like films directed by The Wachowskis as The Matrix.

dbpedia

-owl:star

ring

dbpedia-owl:starring

dcterms:sub

ject!

dcterm

s:subje

ct!

dbpedia-owl:director

dbpedia-ow

l:director

dcterms:subject!dbpedia-owl:starring

dcterms:sub

ject!

EXPERIMENTAL EVALUATION

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

User study

308 subjects ➤ male = 70% ➤ medium interests in movies ➤ 90,318 movies from IMDB

EXPERIMENTAL EVALUATION

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

4 explanation styles ➤ popularity (baseline) ➤ LOD properties ➤ ExpLOD ➤ ExpLOD no IDF

EXPERIMENTAL EVALUATION

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Metrics ➤ Transparency

I understood why I received the recommendation ➤ Persuasion

The recommendation is more convincing ➤ Engagement

I discovered new information about the movie ➤ Trust

My trust in the recommender is increased ➤ Effectiveness

I like the recommendation

EXPERIMENTAL EVALUATION

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Research Question 1 Is ExpLOD able to lead to more transparent, persuasive, trustful, or engaging recommendations?

Research Question 2 Is there a correlation between the properties used to generate the explanation and the effectiveness of the explanation itself?

EXPERIMENTAL EVALUATION

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

user asked to ➤ provide basic demo data ➤ provide preferences ➤ read the explanation and

answer to 5 questions ➤ express a rating before

and after watching the movie trailer

EXPERIMENTAL EVALUATION

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Research Question 1 Is ExpLOD able to lead to more transparent, persuasive, trustful, or engaging recommendations? YES

transparency persuasion engagement trust effectiveness

popularity 3.01 2.59 2.31 2.67 0.93

LOD prop 3.04 2.84 3.28 2.81 0.66

ExpLOD-noIDF 4.00 3.39 3.48 3.39 0.72

ExpLOD 4.18 3.41 3.31 3.36 0.75

The best-performing configuration is in bold. Significant improvements over the baselines are in italics. Transparency, persuasion, engagement, and trust are evaluated by the average score collected through the questionnaires.Effectiveness is calculated as the normalized difference between pre- and post-trailer ratings

EXPERIMENTAL EVALUATION

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Research Question 2 Is there a correlation between the properties used to generate the explanation and the effectiveness of the explanation itself? YES

aim question positive negativetransparency I understood why I received the

recommendation topic, director distributor, music composer

persuasion The recommendation is more convincing awards, director location, producer

engagement I discovered new information about the movie writer, director producer, distributor

trust My trust in the recommender is increased awards, composer producer, topic

effectiveness I like the recommendation director, writer location, composer

QUESTIONS?

ExpLOD: a Framework for Explaining Recommendations based on the Linked Open Data Cloud

Thanks for your attention! [email protected]

BPN