Bigger data set, better personalized recommendation performance?

In recommendation systems, the relationship between information size and recommendation performance is an important research point. Here, we study this relationship based on a new method, variable precision, and design a new algorithm. We demonstrate that recommendation systems perform better with higher data precision, however which should be controlled within a threshold. We collect movie rating records from MovieLens data sets. Original dataset is classified into three different datasets by using variable precision method. We calculate the Cosine similarities and find the change rules of user similarities. Then, top-N neighbours of each user are used to predict movie rating and choose both MAE and F₁-mean as metrics. In our experiment, we get the optimum value of variable precision, meanwhile the value of MAE and F₁-mean are average improved by 16.9%, 2.1% respectively. We also illustrate some important rules and conclusions which can be useful when being applied to other platforms.