Closed tomerk closed 9 years ago
So we don't actually want the default model to be the average user weight. This falls into the warm-start problem (what's the optimal thing to do for users we don't know anything about). One reasonable approach is to take the average label across all users for each training data point and use that to train the "warm-start" model. This should still be done during bulk retrain, so your point still stands.
Average user weight should be recomputed during a bulk retrain. Could probably just be per-partition if we have a reasonable partitioning strategy.