hcdm / banditlib Goto Github PK

Hello from what I understood this repo runs algorithms on simulated data.
What are the places I should change to feed it anything like movielens dataset for example?
What should be the data format, implicit (0,1) or explicit ratings?

So the simulation's training iteration is 0? So how does the model be trained? And why does the plot show that regret is decreasing?

I have read the code, and I got a question that described as the title. There was an variable called coTheta which is used to calculated the reward. Anyone who know why?
Any help is appreciated.

hi, I'm interested in your paper that was published in AAAI 17,
and am trying to get the simulation running, but am receiving the following error.

$ python Simulation.py --alg factorUCB
...
File "Simulation.py", line 508, in
algorithms['LinUCB'] = N_LinUCBAlgorithm(dimension = context_dimension, alpha = alpha, lambda_ = lambda_, n = n_users)
TypeError: init() got an unexpected keyword argument 'n'

This occurs because the constructor for the N_LinUCBAlgorithm class doesn't expect an "n" arg, but instead an "init" one which does not seem related to the "n" in question.

class N_LinUCBAlgorithm:
def init(self, dimension, alpha, lambda_, init="zero"): # n is number of users
self.users = {}

Should I just remove " n = n_users' from the call on line 508 in Simulation.py,
as you sometimes call N_LinUCBA without it from within that same file ( see line 495) ?

algorithms['LinUCB'] = N_LinUCBAlgorithm(dimension = context_dimension, alpha = alpha, lambda_ = lambda_)

Thank you

Hello, I'm working on mab algorithms and I'm having a problem with data preprocess. Is there a place where you can share the preprocessing of the lastfm or Delicious datasets?

Hi,

I've been running some evaluation to compare our algorithm with your implemented factorUCB in the evaluation of LastFM dataset, and I encountered a "MemoryError" for the initializing an identity matrix "identity(n=self.d * userNum)". In my experiment, the user number in this dataset is 1892. Could you please tell me the memory capacity of your experimental machine, so I could move my experiment to another bigger memory one. Or if you have some other versions of implementation, please give me some tips about how to optimize the current version codes.

Looking forward to your reply.

Thanks.

In paper "A Contextual-Bandit Approach to Personalized News Article Recommendation", algorithm LinUCB does not consider the differences between users.
Users' information are only contained in contextFeatureVector, which is referred to "x_t,a" in Algorithm 1. Also, variables "A", "b", "theta" varied from arms(articles) rather than users.
Therefore, I wonder if there are mistakes in class LinUCBUserStruct in your code. I do not think it needs structure for each user and the implementation of "A", "b", "theta" may be wrong.
Could you help me figure out this confusion, thanks!