Improving Prediction Accuracy by Distinguishing the Importance of Different User-Item Pairs

Nowadays, the network brings a lot of conveniences along with problems. One of the challenges is how to efficiently obtain useful information from mass data. The rating prediction is an essential tool to handle this. However, the lack of rating data makes recommendation algorithms not accurate enough. We find that the existing algorithms treat each rating record's contribution to prediction results as equivalent. In fact, there is an implicit condition that different records have different importance for prediction. In this article, we consider this and propose two methods to distinguish the importance of rating data, one assigns weights during modeling and another during model learning. The experimental result demonstrates our algorithms are effective and outperforms several state-of-the-art methods.