Flexible Adjustment of Feature Vector Correlations for Personalized Recommendations

Personalized recommendation vectors are often correlated with the global feature vector, such as a popularity vector. Prior works tried to adjust the correlation using parameters. However, there are no parameter setting guidelines, and it is unclear whether the desired correlation is achieved. We propose CoCoA to generate the recommendation vector with a linear combination of the given source and global vectors. It ensures the cosine similarity between the recommendation and global vectors is a user-input value. A case study with a movie dataset observed CoCoA can flexibly control the effect of popular movies. Evaluations on three real-world graph datasets showed that only CoCoA achieved negative cosine similarity, where the globally important nodes are suppressed.