My Mobile Music: An Adaptive Personalization System For Digital Audio Players

Thumbnail Image


umi-umd-4650.pdf (526.44 KB)
No. of downloads: 2822

Publication or External Link






This paper develops a music recommendation system that automates the downloading of songs into a mobile digital audio device. The system tailors the composition of the songs to the preferences of individuals based on past behaviors. By assuming that an individual will listen longer to a song that provides a higher utility, we describe and predict individual listening behavior using a lognormal hazard function. Our recommendation system is the first to accomplish this and there is no viable alternative. Yet, our proposed approach provides an improvement over naïve methods that could be used for product recommendations. Our system has a number of distinct features. First, we use of a Sequential Monte Carlo algorithm that enables the system to deal with massive historic datasets on listening behavior of individuals. Second, we apply a variable selection procedure that helps to reduce the dimensionality of the problem, because in many applications the collection of songs need to be described by a very large number of explanatory variables (in particular music genres variables). Third, our system recommends a batch of products rather than a single product, taking into account the predicted utility and the uncertainty in the parameter estimates, and applying experimental design methods.

The simulation section of this paper demonstrated that our model does achieve it objectives in handling massive data and improving predictions through model averaging. By using simulated data in the simulation, and thus knowing the true parameters, the Sequential Monte Carlo and variable selection procedures were shown to provide good estimates of an individual's preferences. Experimental results show that variable selection does simplify estimation and prediction as different individuals differ in the number of variables need to definite their listening behaviors. The results also show that for some individuals, model averaging does in fact help to improve predictions. The results of the experiment show that our model provides 23 - 35% improvement in recommendations. This improvement is achieved in a single wave and in a natural experimental setting in which the subjects have a choice or when, where and how they want to listen to the songs.