Recommendation system MOPSI project KAROL WAGA 23.04.2013.

Recom

men

datio

n

syst

em

MOPSI p

roje

ct

http://

cs.u

ef.fi

/mop

si

KAROL WAGA

23.04.2

013

CONTENT

• CONCEPT OF RECOMMENDATION SYSTEM

• CURRENT SOURCE OF INFORMATION• CONTEXT OF RELEVANCE• SYSTEM ARCHITECTURE• SCORING SYSTEM• EXAMPLE

• PROPOSED SYSTEM IMPROVEMENTS • USER ACTIVITY• PHOTOGRAPH CONTENT ANALYSIS

23.04.201

3

2

3

CONCEPT – RECOMMENDATION SYSTEM

• is a subclass of information filtering system that seek to predict the 'rating' or 'preference' that user would give to an item (such as music, books, or movies) or social element (e.g. people or groups) they had not yet considered, using a model built from the characteristics of an item (content-based approaches) or the user's social environment (collaborative filtering approaches).

BENEFITS OF THE RECOMMENDATION SYSTEM:

1. finding items relevant to user among many items

2. personalized based on real activity

3. allow discovering things similar to what one already liked

23.04.201

3

4

CONCEPT – RECOMMENDATION SYSTEM

23.04.201

3

CURRENT SOURCE OF INFORMATION

SERVICES

523.04.201

3

CURRENT SOURCE OF INFORMATION

PHOTOS

623.04.201

3

CURRENT SOURCE OF INFORMATION

ROUTES

723.04.201

3

CONTEXTS OF RELEVANCE

P - Position (what is nearby)

I - Information (filter relevant information)

N - Network (what others have looked for and found useful)

T - Time (what is useful now)

823.04.201

3

CONTEXTS OF RELEVANCE

P – if user is in Science Park lunch restaurants in Käpykangas are not relevant

I – if user does not like sports then nearby gyms, jogging tracks, skiing tracks are not important for him

N – restaurant rated well by users should be recommended even if it's further than restaurants without user rating

T – in summer time skiing tracks and skating rinks are not relevant

923.04.201

3

CONTEXTS OF RELEVANCE

POSITION

1023.04.201

3

CONTEXTS OF RELEVANCE

INFORMATION

1123.04.201

3

CONTEXTS OF RELEVANCE

NETWORK

1223.04.201

3

CONTEXTS OF RELEVANCE

TIME

1323.04.201

3

SYSTEM ARCHITECTURE

1423.04.201

3

THE SCORING SYSTEM

• Items for scoring are selected based on distance from user’s location

• Services are scored based on position, search history and rating. As ”high quality” source services are promoted by adding 1 to their score (instead of time scoring that is applied to photos and routes)

• Photos are scored based on position, search history and rating and time

1523.04.201

3

THE SCORING SYSTEM

• Routes are scored based on position, attractivity (number of services and pictures in the end point and along the route) and time

• Scores are normalized to <0,1> and the results from the three groups are merged into one list sorted decreasingly

• Top items are shown as recommendation to user

1623.04.201

3

EXAMPLE

1723.04.201

3

EXAMPLE

18

Utra church (262 m)

Total score: 3.93

L: 0.97 H: 1.0 R: 0.0

- the nearest service- popular keyword

Utra swimming place (575 m)

Total score: 3.0

L: 0.90 H: 0.33 R: 0.0 T: 0.88

- nearby photo- taken in the same season of the year

Utrantie 88 – Kalevankatu 29 (34 m)

Total score: 3.1

L: 1.0 A: 1.0 T: 0.1

- the nearest item in database- leads to attractive destination with many pictures

23.04.201

3

PROPOSED SYSTEM IMPROVEMENTS

• USER ACTIVITY• USER PROFILE• DETECTING USER ACTIVITY• RECORDING USER ACTIVITY• CREATING ACTIVITY MODEL

• PHOTOGRAPH CONTENT ANALYSIS

1923.04.201

3

USER PROFILE

• is the computer representation of a user model.

The main goal of user modeling is customization and adaptation of systems to the user's specific needs.

Gathering information about user is done by recording user activity on website and in mobile application, detecting user activities in the real world and analysing user's collection.

2023.04.201

3

RECORDING USER ACTIVITY

1) Storing activities on client side in web browser (Javascript) and on mobile devices

2) Sending data to server (JSON)

3) Parsing data and saving to database (PHP and MySQL)

All the stages are based on activity model.

2123.04.201

3

DETECTING USER ACTIVITY (http://www.cs.uef.fi/paikka/karol/doktorat/events%202.4.swf)

2223.04.201

3

CONTENT of user profile

List of favorite keywords based on rating (services and photos) and visits (services) to recommend items with these keywords with higher probability – involved keyword clustering

List of services and photos rated bad to avoid recommending these items

Movement type statistics to recommend favorite type of routes

Similarity matrix with other users based on similarity of favorite keywords, route types and number of common friends (Facebook), detected meeting number

2323.04.201

3

PHOTOGRAPH CONTENT ANALYSIS

INPUT: a MOPSI photo

• retrieve pictures from Flickr in the same location

• use open source project for image similarity

• use perceptual hash to sort output based on similarity

• get tags from Flickr of the most similar images

OUTPUT: set of keywords describing the MOPSI photo 2423.04.201

3

Thank you for attention…

Any questions?

2523.04.201

3