Yes we all know: ratings from users are very noisy and not consistent. If you ask users to re-rate items they will do it differently in most cases. This was pointed out by a number of scientists. Knowing and removing noise could lead to better prediction performance. One simple approach to reduce noise is achieved by asking people to re-rate all objects they have rated so far. This was also pointed out by technocalifornia.
However, I doubt that users are very happy to re-rate everything. So, how can we learn and remove noise? Currently I try to model user ratings, taking into account inconsistency and peer influence. I hope the model will give some insights, when compared to real world data generated by recommender systems.
Recent Comments