Adaptive music recommendation system

Abstract

While sources of digital music are getting more abundant and music players are becoming increasingly feature-rich, we still struggle to find new music that we may like. This thesis explores the design and implementation of the MusicPlanner - a music recommendation application that utilizes a goal-oriented framework to recommend and play music. Goal-oriented programming approaches problems by modeling them using Goals, Techniques, and a Planner. The Goals are representations of a user's intent, while the Techniques are the methods that can be used to satisfy the Goals. The Planner connects the Goals and Techniques in a user-defined way to find solutions to user's requests. In the MusicPlanner, the Planner defines the top level Goal of recommending music, which can be satisfied by a set of recommendation Techniques. Each of the recommendation Techniques then declares the sub-Goal of playing music, which can be satisfied by a set of play Techniques. The Planner evaluates each of the Techniques and iterates through the results to choose the best set of Techniques to satisfy the top-level goal of music recommendation. The MusicPlanner allows the user to create personal music stations and for each station, constructs a model of user's music taste based on queries and feedback to the songs played. The extensible design of the architecture and the ease of implementing the MusicPlanner show how goal-oriented framework can simplify the work for programmers. In evaluating the performance of the MusicPlanner, we demonstrate that the Planner in the goal-oriented framework outperforms each individual recommendation Technique.