Presentation is loading. Please wait.

Presentation is loading. Please wait.

Xinxi Wang, David Rosenblum, Ye Wang School of Computing, National University of Singapore 1.

Similar presentations


Presentation on theme: "Xinxi Wang, David Rosenblum, Ye Wang School of Computing, National University of Singapore 1."— Presentation transcript:

1

2 Xinxi Wang, David Rosenblum, Ye Wang School of Computing, National University of Singapore 1

3 Do you prefer the same or different music when running or sleeping? 2

4 Users short-term music information needs are influenced by users activities [A. C. North 2004] Traditional recommender systems model user long-term preferences only Running Sleeping Girlfriend This song is generally good, but Im going to sleep. Its too noisy! Girlfriend Sleepsong Collaborative filtering (CF)... 3

5 How to recommend songs to this new user? new user problem How to recommend this new song? new song problem Collaborative Filtering (CF) cannot handle both Content-based filtering cannot solve the new user problem 4

6 Running Sleeping Audio content analysis Sensor based activity detection Personalization and adaptation … … …… Our system detects users daily activities in real-time and recommends suitable music automatically 5

7 Music Database Music Database Running Walking Sleeping Working Studying Shopping Binary classifiers (Adaboost) Binary classifiers (Adaboost) Music audio feature extraction Sensor signal features feature extraction ACACF Probabilistic Graphical Model ACACF Probabilistic Graphical Model User feedback Classification results Recommendation Backend Frontend 6

8 Activities Ranked songs list Playback Controls automatic mode Skipped Listened completely manual mode 7

9 Given the sensor feature f, a song s is scored as: Sensor-Context Model Music-Context Model 8

10 Different people have agreement on suitable music for an activity. Initialization. Prior beta(a, b) is initialized from music content analysis results. 9

11 Approximate prior update: User preference update: 10

12 Six models are compared based on three criteria: (1) Energy consumption (2) Accuracy; (3) Incremental learning. Energy consumption and incremental learning 11

13 Accuracy of different models: 12

14 Naïve Bayes: Training and incremental training by MLE: 13

15 Retrieval performance 14

16 15

17 16

18 With existing technologies, their short-term needs cannot be satisfied well. 17

19 18

20 The first context-aware mobile music recommendation system for daily activities It satisfies users short-term needs better A solution the cold-start problem Unified probabilistic model 19

21 Let more people use it Exploration/exploitation tradeoff using reinforcement learning Incorporate collaborative filtering into the system 20

22 21

23 A. C. North, D. J. Hargreaves, and J. J. Hargreaves, Uses of Music in Everyday Life, Music Perception: An Interdisciplinary Journal, vol. 22, no. 1, A. I. Schein, A. Popescul, L. H. Ungar, and D. M. Pennock, Methods and metrics for cold- start recommendations, in SIGIR,


Download ppt "Xinxi Wang, David Rosenblum, Ye Wang School of Computing, National University of Singapore 1."

Similar presentations


Ads by Google