A Method for Dealing with Data Sparsity and Cold-Start Limitations in Service Recommendation Using Personalized Preferences

Data sparsity and cold-start remains to be the main limitations and weaknesses in recommendation systems that employ collaborative filtering (CF). These limitations cause lack of convergence in CF recommendation algorithms which ultimately affect the overall accuracy of the recommendation system. Efforts to alleviate these limitations typically require additional user or item information such as social context of users and features of items, besides ratings that are usually available. The additional information could possibly provide useful information about the underlying model and complement the rating data. However, existing choices of additional information may not improve recommendation accuracy. In some cases, some choices of additional information may diminish the accuracy of recommendation systems, although they seemingly alleviate data sparsity and cold-start limitations. In this article, a method to resolve data sparsity and cold-start limitations using users' personalized preferences on non-functional attributes, as additional information, is proposed. The proposed method focuses on improving similarity functions used in the recommendation process. We show that incorporating users' personalized preference on non-functional attributes in a certain fashion, alleviates these limitations and improves recommendation accuracy. We also conduct thorough experiments on real-world services that demonstrates the effectiveness of our method. The experiments also highlights some benefits of using good choice of additional information in dealing with both data sparsity and cold-start limitations.