Tensor Decomposition Method Applied to Recommendation Systems
With the growth of network web services, web service recommendations based on the quality of service (QoS) attribute have become a research interest topic. Service recommendation technologies can help users discover new web services and make their online experience better. Further, by providing users with recommendations for high-quality web services, these technologies ultimately benefit both users and service providers. In this thesis, we study the tensor decomposition in web service recommendations. In particular, we propose new tensor computational methods and algorithms for QoS attribute prediction to improve the recommendation accuracy. Our methods follow the machine learning techniques. First, to remedy the shortage of low prediction accuracy rates caused by the lack of initial data samples, a traversal-tensor method (TTM) is proposed to enhance the sampling scheme. The new method integrates the feature factor matrices to construct more data samples for tensor decomposition. We analyze and validate the new algorithm in comparison with the traditional tensor decomposition applied to service recommendations. Empirical studies with multiple datasets show that the TTM effectively improves the prediction performance. Second, a modified regularization term is designed and applied with the TTM to overcome the overfitting problem. This is done by using a linear combination of two commonly applied regularization models. It is shown that the updated term can increase the accuracy rate of predicting QoS attributes and better support the TTM method. Third, a two-step strategy approach involving a K-means clustering with TTM is introduced to deal with the initial unorganized data. The pre-clustered data are used as input to the TTM to complete the QoS attribute prediction. This process is evaluated between our methods and the clustering method. The thesis describes a framework of tensor-based web service recommendation by synthesizing the above methods. This framework is centered on TTM, with a modified regularization term to support TTM and a method to handle the initial unorganized data.
URI for this recordhttp://hdl.handle.net/1974/29446
Request an alternative formatIf you require this document in an alternate, accessible format, please contact the Queen's Adaptive Technology Centre
The following license files are associated with this item: