Enhancing recommendation systems performance using highly-effective similarity measures

摘要

In Recommendation Systems (RS) and Collaborative Filtering (CF), the similarity measures have been the operating component upon which CF performance is essentially reliant. A dozen of similarity measures have been proposed to reach the desired performance particularly under the circumstances of data sparsity (the cold-start problem). Nevertheless, these measures still suffer the cold-start problem, and have a complex design. Moreover, a comprehensive experimental work to study the impact of the cold-start problem on CF performance is still missing. To these ends, therefore, this paper introduces three simply-designed similarity measures, namely, difference-based similarity measure (SMD), hybrid difference-based similarity measure (HSMD), and, triangle-based cosine measure (TA). Along with proposing these measures, a comprehensive experimental guide for CF measures using the K-fold cross validation is also presented. In contrary to all previous CF studies, the evaluation process is split into two sub-processes: the estimation process and recommendation process to accurately obtain the desired appropriateness in the evaluation. In addition, a new formula to calculate the dynamic recommendation count is developed depending on both the dataset and rating vectors. To draw a comprehensive experimental analysis, a dozen state-of-the-art similarity measures (30 similarity measures) including the proposed and the most widely-used traditional measures are comparatively tested. The experimental study has critically been made on three datasets with five-fold cross-validation grounded on the K nearest neighbor algorithm (KNN). The obtained results on both estimation and recommendation processes prove unquestionably that SMD and TA are preeminent measures with the lowest computational complexity outperforming all state-of-the-art CF measures.