What is better for starting a recommender system, a cold start or a dataset shifted towards the highest rating?

Question

Machine learning

Hello, I use the SVD algorithm and to start work there is a choice:
1) At the very beginning, turn off the PC and collect user ratings
2) Turn on the PC and collect user ratings to recalculate matrices
3) Either I have a dataset, but there are only 4 and 5 ratings five-point system - there are no negative ratings ..
What is the right thing to do in this situation?

Answer 1

First - you need to run a full cycle of Full-Mesh binary comparison once and, thus, adjust the starting weights for each position.
Since you already have something: scores 4 and 5: use half division (by the lower bound on all indicators: you have a 4) to create binary reviews: 4 => bad / no (-1 ) and 5=> good/yes (1).
Then - iteratively adjust as the values ​​of specific indicators change.