Implement uniform support for setting random seeds about lkpy HOT 1 CLOSED

Everything seems to be working pretty well for algorithms that randomize some part of their training - we can provide a seed, and it works.

What isn't quite there is support for algorithms that randomize in their recommendations or predictions. Right now, randomization is unpredictable, particularly in the case of parallel batch recommendation.

To make algorithms with random recommendations reproducible, I propose to leverage NumPy's flexible SeedSequence. If the user instantiates Random with rng='user', then it will:

• Spawn a child of the global seed sequence and save it for later use
• For each recommendation request, spawn a child of that sequence with the user ID as the next element of the spawn key, and use it to initialize an RNG to randomize the recommendations

This will achieve two things:

• Recommendations will be unpredictable per-user - by the time the user ID is mixed into the entropy, and the result used to seed a PRNG, the resulting bitstream is difficult to predict
• The same initial seed and the same user ID will produce the same recommendation list, regardless of parallelism or subprocess computations.

Experiment designs that require multiple recommendation lists per user may have a problem. If @kluver or @KimuraTian (or anyone else) has thoughts here, would love to hear them before doing this.

from lkpy.