Impact of Training Epoch Size on the Accuracy of Collaborative Filtering Models in GraphChi Utilizing a Multi-Cyclic Training Regimen

Abstract

More and more outlets are utilizing collaborative filtering techniques to make sense of the sea of data generated by our hyper-connected world. How a collaborative filtering model is generated can be the difference between accurate or flawed predictions. This study is to determine the impact of a cyclical training regimen on the algorithms presented in the Collaborative Filtering Toolkit for GraphChi. Initial testing shows that some of the algorithms benefited from a multi-cyclic approach, a result that is reinforced by repeating the experiment on a separate dataset. Additional testing focuses on the effectiveness of dynamic versus fixed training cycle sizes. However, there is no additional benefit to adopting this more complex training scheme. While the results are far from universal, half of the algorithms saw a significant increase in accuracy when subjected to a multi-cyclic training regimen.