In order to improve customer experience companies can invest in recommendation systems. Recommendation systems attempt to profile user preferences and provide users with good personalized recommendations. We focus on the neighborhood-based methods for Collaborative Filtering. Collaborative Filtering only relies on past user behavior. Neighborhood-based methods select the most similar items or users (neighbors) and make predictions by determining interpolation weights for these neighbors. We implement the methods proposed by Bell and Koren in [3]. Bell and Koren proposed an improved neighborhood-based Collaborative Filtering method, which addresses some of the issues of previous neighborhood-based methods. This improved method consists of three main components, namely data normalization, the selection of neighbors and the determination of interpolation weights. This method was evaluated on the Netflix prize data [4]. The inclusion of global effects alone gives us a RMSE of 0.9658. If we perform no data normalization the root mean squared error (RMSE) for Bell and Koren’s neighborhood interpolation method is worse, around 0.975. For Double Centering, the RMSE improves with approximately 0.05. We find the best RMSE by applying full data normalization, which resulted in a RMSE of 0.9194. We extend the methods by Bell and Koren by performing further data normalization. We remove the Genre effect and some temporal effects, but this does not improve the RMSE much. Only the Genre effect lowers the RMSE for the inclusion of global effects to 0.9655.

Additional Metadata
Thesis Advisor Burg, G. van den
Persistent URL
Series Econometrie
Mathol, E.L.J. (2016, July 6). Neighborhood-based Collaborative Filtering: Providing the best recommendations. Econometrie. Retrieved from