Käyttäjä- ja käyttäytymistietojen
hyödyntäminen suositusjärjestelmissä
User and behavioural data in
recommendation systems
Asiakasseminaari 24.9.2013
Asta Bäck, VTT
26.9.2013

2

Growth - Driven by Data
Linked data

Recommendations
Open data
Content
Own websites
and mobile apps

Gathering

Analysing
Learning

Acting
26.9.2013

Trust

3

People find
more than they
expected
Service more
relevant and
easier to use

Higher
ARPU
Reduced
churn

Longer
visits
Willing to
return
26.9.2013

What do consumers think
of recommendations?

4
26.9.2013

5

Which Recommendation Methods to Use?

User 1
User 2
User 3
User 4
User 5
User 6
User 7
…

Item 1
5
4
1

Item 2
4
2
3
1
4

Item 3

4
1

Item 4
1

…

Wellbeing
Health

3
2

Collaborative filtering

Content based recommendations
Knowledge based recommendations

Contextual recommendations

Social recommendations

Hybrid recommendations
26.9.2013

6

Technology for content and knowledge based recommendations
An example of TV-program synopsis:

(1960) Lyhytelokuvassa esitellään suomalaista taideteollisuutta.
Mukana ovat mm. Dora Jung, Kaj Franck ja Timo Sarpaneva.
taideteollisuus
Timo Sarpaneva
Inferred meanings of keywords:
http://www.yso.fi/onto/koko/p34662

Keyword extraction

Semantic analysis of keywords

http://dbpedia.org/resource/Timo_Sarpaneva
Labels in different languages:
konstindustri (sv), industrial art
(en)
Related concepts: käsi- ja
taideteollisuus, lasitaide,
muotoilu, taidekäsityö,
sisustustaide…
Broader concept: teollisuus

Semantic expansion
Entities: Person, Finnish
industrial designers…

Semantic recommendations
26.9.2013

7

Semantic profiles and recommendations in SP3/Mediatutka

New content and
semantic content
enrichment
Location sensitive
proactive notifications
Recommendations

Semantic user profiles and
semantic enrichment of
user metadata
26.9.2013

8

Recommendations Based on Collaborative Filtering
A generic self organizing method (UPCV), based on the
behavior of individual users and user groups:
no metadata; applicable to any item in any service
easily converted to a hybrid recommender
simple to understand and use; contains no secret sauce

UPCV is able to provide:
user-to-item recommendations (items for a user)
item-to-item recommendations (items similar to an item)
item-to-user recommendations (interested users for an item)
user-to-user recommendations (users with similar behavior)

Patents pending (FI, EPO, US)
26.9.2013

9

Open Recommendation Initiative based on UPCV
Each party (service, user) owns all recommendation data related to it
Recommendation is based on all user activities, in any service
Customership belongs to services - not e.g. to Google!
Uncompromised privacy
UPCV recommendation engine:
Distributed architecture
Simple HTTP REST API
Small footprint
Easy deployment
26.9.2013

VTT creates business from technology

10

User and behavioural data in recommendation system - Asta Bäck

  • 1.
    Käyttäjä- ja käyttäytymistietojen hyödyntäminensuositusjärjestelmissä User and behavioural data in recommendation systems Asiakasseminaari 24.9.2013 Asta Bäck, VTT
  • 2.
    26.9.2013 2 Growth - Drivenby Data Linked data Recommendations Open data Content Own websites and mobile apps Gathering Analysing Learning Acting
  • 3.
    26.9.2013 Trust 3 People find more thanthey expected Service more relevant and easier to use Higher ARPU Reduced churn Longer visits Willing to return
  • 4.
    26.9.2013 What do consumersthink of recommendations? 4
  • 5.
    26.9.2013 5 Which Recommendation Methodsto Use? User 1 User 2 User 3 User 4 User 5 User 6 User 7 … Item 1 5 4 1 Item 2 4 2 3 1 4 Item 3 4 1 Item 4 1 … Wellbeing Health 3 2 Collaborative filtering Content based recommendations Knowledge based recommendations Contextual recommendations Social recommendations Hybrid recommendations
  • 6.
    26.9.2013 6 Technology for contentand knowledge based recommendations An example of TV-program synopsis: (1960) Lyhytelokuvassa esitellään suomalaista taideteollisuutta. Mukana ovat mm. Dora Jung, Kaj Franck ja Timo Sarpaneva. taideteollisuus Timo Sarpaneva Inferred meanings of keywords: http://www.yso.fi/onto/koko/p34662 Keyword extraction Semantic analysis of keywords http://dbpedia.org/resource/Timo_Sarpaneva Labels in different languages: konstindustri (sv), industrial art (en) Related concepts: käsi- ja taideteollisuus, lasitaide, muotoilu, taidekäsityö, sisustustaide… Broader concept: teollisuus Semantic expansion Entities: Person, Finnish industrial designers… Semantic recommendations
  • 7.
    26.9.2013 7 Semantic profiles andrecommendations in SP3/Mediatutka New content and semantic content enrichment Location sensitive proactive notifications Recommendations Semantic user profiles and semantic enrichment of user metadata
  • 8.
    26.9.2013 8 Recommendations Based onCollaborative Filtering A generic self organizing method (UPCV), based on the behavior of individual users and user groups: no metadata; applicable to any item in any service easily converted to a hybrid recommender simple to understand and use; contains no secret sauce UPCV is able to provide: user-to-item recommendations (items for a user) item-to-item recommendations (items similar to an item) item-to-user recommendations (interested users for an item) user-to-user recommendations (users with similar behavior) Patents pending (FI, EPO, US)
  • 9.
    26.9.2013 9 Open Recommendation Initiativebased on UPCV Each party (service, user) owns all recommendation data related to it Recommendation is based on all user activities, in any service Customership belongs to services - not e.g. to Google! Uncompromised privacy UPCV recommendation engine: Distributed architecture Simple HTTP REST API Small footprint Easy deployment
  • 10.