Presentation on theme: "Context-Aware Mobile Music Recommendation for Daily Activities"— Presentation transcript:
1Context-Aware Mobile Music Recommendation for Daily Activities Xinxi Wang, David Rosenblum, Ye WangSchool of Computing, National University of Singapore
2A simple questionBefore we start, I would like to ask you a simple question.Do you listen to the same or different music when running or sleeping?After they raise up their hands:Thanks.Do you prefer the same or different music when running or sleeping?
3Motivation - short-term needs Users’ short-term music information needs are influenced by users’ activities [A. C. North 2004]Traditional recommender systems model user long-term preferences onlyRunningSleepingGirlfriendSleepsongGirlfriendMost of you prefer different music for different activities. This has been studied by some music psychologist.From this study, we know users’ short-term music information needs are influenced by users’ activities.However, most traditional recommender systems are based on collaborative filtering. And they model user long-term preferences only. They don’t know what a user is doing. So they can make mistakes.For example, if I’am this user and I’m going to sleep. According to CF, Girlfriend will be recommended to me. This is obviously a mistake, and I don’t think I can fall asleep with this song.This song is generally good, but I’m going to sleep. It’s too noisy!Collaborative filtering (CF)...
4Motivation - the cold-start problem How to recommendthis new song?new song problemHow to recommendsongs to this new user?new user problemCollaborative Filtering (CF) cannot handle bothContent-based filtering cannot solve the new user problem
5The Main IdeaOur system detects users’ daily activities in real-time and recommends suitable music automaticallyRunningSleepingThe main idea of our system is like this:Our system detects users’ daily activities in real-time and recommends suitable music automaticallySpecifically, we first predict every song’s suitable activities by examine the audio features of very song. For example, we predict this song is suitable for running, but this song is not suitable for running. We do the same prediction for other activities.On the other hand, we predict a user’s current activity by examine the low level sensor data collected from the mobile phone. Then we use the predicted activity to lookup this table and find out the best song and then recommend it to the user.Since this process is identical to every user, to make the recommendation adaptive and personalized, we also utilize users’ feedback.Let’s do a short demo of our system.…………Audio content analysisSensor based activity detectionPersonalization and adaptation
6System Architecture Backend Frontend Music audio feature extraction Binary classifiers(Adaboost)RunningWalkingSleepingMusicDatabaseWorkingStudyingShoppingClassification resultsRecommendationACACFProbabilistic Graphical ModelLet’s see how our system works in detail.Our system is based heavily on machine learning. It’s consist of two parts: the backend and the frontend. The backend is built in a powerful cluster. And the frontend is all inside a mobile phone.In the backend, we have a large music database. We first extract music audio features for every song, and then we use a set of binary classifiers to predict whether a song is suitable for an activities or not. The classification results will then be transferred to the frontend via wireless network. On the other hand, we collect features extracted from the low level sensor data, and also the user feedback. Then we use a probabilistic graphical model to integrate the three kinds of information together to provide accurate recommendation.FrontendUser feedbackSensor signal featuresfeature extraction
22ReferencesA. C. North, D. J. Hargreaves, and J. J. Hargreaves, “Uses of Music in Everyday Life,” Music Perception: An Interdisciplinary Journal, vol. 22, no. 1, 2004.A. I. Schein, A. Popescul, L. H. Ungar, and D. M. Pennock, “Methods and metrics for cold-start recommendations,” in SIGIR, 2002.