Improving capabilities of recommender systems using swarm intelligence
Supiya Ujjin · OpenGrey (Institut de l'Information Scientifique et Technique) · 2004
This thesis investigates the use of novel genetic and swarm intelligence algorithms to increase performance and capabilities of recommender systems. Through the use of these algorithms, the work introduces the concept of recommender systems adapting to the user recommendation preferences to provide more accurate and useful suggestions. An extensive literature review evaluates all related areas of research and reveals that such systems do not exist. The development of two generic recommender systems, one using a genetic algorithm (GA) and the other using particle swarm optimisation (PSO), are then presented. The thesis describes a number of significant advances. Firstly, a new data representation of a user profile is introduced, which takes into account multiple features of the data. Secondly, similarity measures are investigated in order for user profiles to be compared accurately. It is vital that an appropriate measure is used as the success of any recommender system based on a collaborative filtering approach depends significantly on this. Thirdly, a new fitness function is devised, which uses the quality of recommendations to guide evolution, by reformulating the problem of making recommendations into a supervised learning task. Additionally, the thesis describes significant advances in the field of swarm intelligence during the development of the second system. Two novel swarming algorithms were created. The ClusterPSO algorithm is the first application of swarm intelligence to the problem of adaptive recommender systems. The ClusterWeight then builds on the ClusterPSO algorithm and is the first to simultaneously cluster and search for solutions. The idea of dynamic swarming is introduced, which allows users themselves to form or join a group of similar users, enabling real-time open-ended adaptation to continuously changing data. The prediction accuracy, speed and usability (in terms of relevance of recommendations and user scalability) of the systems are assessed by performing comparisons with existing algorithms used within recommender systems and through a pilot study. The results of these experiments show that capabilities of recommender systems can be improved by the use of evolutionary algorithms and swarm intelligence.