improvise
DESCRIPTION
Improvise presentation at the CBRecSys Workshop in San Mateo, CA.TRANSCRIPT
A User-centeredMusic Recommendation Approach
for Daily Activities
Ricardo Dias, Ricardo Cunha, Manuel J. FonsecaINESC-ID/Instituto Superior Técnico, Universidade de Lisboa
Music listening on daily activities
Mark Stevens@Flickr
Listening experience is ubiquitous
clemsonunivlibrary@flickr
> 20 Million Songs
Hard to choose songs
Orin Zebest@flickr
Different users, Different tastes
Chloe Muro@flickr
How to recommend songs?
Jacob Bøtter@flickr
Neglecting User Control
Improvise
davejdoe@flickr
User-centered approach
mammal@flickr
RELATED WORK
Different Approaches
Demographic
Collaborative Filtering
Content
Context
Hybrid
[Celma][Celma2010]
Collaborative Filtering
Song 1 Song 2 Song 3 Song 4 Song 5
Alice 5 3 4 4 ?
User 1 3 1 2 3 3
User 2 4 3 4 3 5
User 3 3 3 1 5 4
User 4 1 5 5 2 1
Content
Moyan Brenn@flickr
Context
Peter Merholz@flickr
Where are the users?
• Text
Goals
1. Generic recommendation technique that takes user input for personalization
2. Recommend songs to be listened performing different activities
IMPROVISE
Recommend songs for Activities
abethne15@flickr
User-centered
the rik pics@flickr
Personalization
Approach
Song1Song1
Song1Song1
Song1Song1
Song1Song5
Song2Song2
Song2Song3
Activity 1
Activity 2
Activity 3
giuvanpelt@flickr
Content for Similarity
ASSOCIATING SONGS WITH ACTIVITIES
• Text
Web Application
• Text
Selection Flow
15 Genres
30 Artists of the selected genres
100 songs from those artists
• Text
Data Collected
98 participants
251 answers to the activities
8,518 songs associated to the activities
• Text
Song Characterization
Accousticness Energy Loudness Tempo
4D vector of content features
Ourania@flickr
Analyze to detect similarities
One feature, one interval
Tempo
Max
Min
[Min; Max]
Hyper-Rectangles
Running
Relaxing
Shopping
One Hyper-Rectangle, 4 Features
Tempo Accoustiness Energy Loudness
• Text
One to One
One Activity
One Hyper-rectangle
• Text
Main Challenge
Defining the hyper-rectangles
Establishing the min/max values for each one the four intervals
HYPER-RECTANGLES
Nick Wheeler@flickr
How to choose the best method?
• Text
Classification Task
Activity 2
Activity 3
Activity 2
Song3
• Text
Testing Procedure
Two datasets for training classifier: top-100 and the top-20
Hyper-rectangles determined using the different methods
• Text
First Trial
M1 - AVG +/- STD for the min/max of each interval
M2 – 10%/90% percentiles for min/max
Running
Relaxing
Shopping
Intervals overlapped
• Text
Second Trial
M3 - % of the STD (15% to 30%) + AVG
M4 - % of the STD (15% to 30%) + Median
Increments of 5% from 15% to 30%
Top-100 songs dataset
Median +/- 20% of the STD
Best Approach
GENERIC RECOMMENDER
Cold-start Approach
Top-100 songs selected by the 98 users for each activity
4 hyper-rectangles generated using the Median +/- 20% of the STD
Generic Model
PERSONALIZATION
Greg Peverill-Conti@flickr
How about individual tastes?
Greg Peverill-Conti@flickr
Method
Same as in Generic Model
Top 100 songs
Median +/- 20% of STD
Greg Peverill-Conti@flickr
Add new songs
List of SongsTop
Bottom
New Song
Greg Peverill-Conti@flickr
First In, First Out
Top
Bottom
Top
BottomIn In
Out Out
Song
Song
Time
Greg Peverill-Conti@flickr
Benefits
1. New songs remain more time in the list to determine the hyper-rectangles
2. Constantly adapts to users current preferences and tastes
EVALUATION
1. Recommend songs to different activities
2. Personalize the recommendations to users with different tastes
Goals
2 F / 8 M
Aged between 22 and 29 years old
Students (graduate and undergraduate from computer science)
Participants
1. Generic model: 10 users
2. Personalized model (two iterations)1. First iteration: 10 users2. Second iteration: 5 users
Two Experiments
PROCEDURE
Characterize users:
•Gender
•Age
• Listening Habits (frequency)
• Etc.
Demographic Questionnaire
Song Selection
Selecting appropriate songs from a list of 50 songs (one list per activity)
Evaluation Task
Satisfaction Questionnaire
Nr. of songs
Metric
RESULTS
Lambros Roussodimos@flickr
Happy users
yourte contemporaine@flickr
Personalized better than Generic
Generic Model
12
38
0 5 10 15 20 25 30 35 40
Selected
NotSelected
Number of Songs
Generic Model
12
38
0 5 10 15 20 25 30 35 40
Selected
NotSelected
Number of Songs
24%
Personalized Model (1º)
12
38
15
35
0 5 10 15 20 25 30 35 40
Selected
NotSelected
Personalized Generic
Personalized Model
12
38
15
35
0 5 10 15 20 25 30 35 40
Selected
NotSelected
Personalized Generic
25%
Personalized Model (2º)
10% increase
Satisfaction Results
Special cases
User 1
User 2
User 3User 4
User 5
Discussion
For some users more interaction is required to improve recommendation
Might improve by using more songs than just the top 100
Discussion
For some users more interaction is required to improve recommendation
Might improve by using more songs than just the top 100
Number of songs steadily increased
LIMITATIONS
Lack of comparison
No comparison between the
Generic Recommender and other solutions
Elvin@flickr
Few Users
Peter Merholz@flickr
Activities were simulated
CONCLUSIONS AND FUTURE WORK
Recommend songs for Daily Activities
Silvia Viñuales@flickr
User-centered Approach
ca_heckler@flickr
Experimental Evaluation
Kat@flickr
More evaluation required
Explore new methods
Use more songs
> 8.000
A User-centeredMusic Recommendation Approach
for Daily Activities
Ricardo Dias, Phd Student and Researcher
http://web.tecnico.ulisboa.pt/ricardo.dias
The End