Two questions

[greeness]

Hi, thanks for sharing the code.

Q1: IIUC, at
https://github.com/CGCL-codes/HCB-pHCB/blob/main/algs/hcb.py#L80, we will never initialize a base_linucb for depth (L-1), which contains the leaf nodes (because we already quit the while loop at line 75 when it is a leaf).

On the other hand, we did not add the root into the path. So when we update the parameters at
https://github.com/CGCL-codes/HCB-pHCB/blob/main/algs/hcb.py#L104, we actually update the parameters for depth 0 to L-2. Is this expected?

Q2: in pHCB, at line
https://github.com/CGCL-codes/HCB-pHCB/blob/main/algs/phcb.py#L95, the reward is multiplied by the depth, I did not see this in the paper, is this intended?

Thanks!

[yusonghust]

Thanks for your question!
Q1: Yes, it's expected, because another bandit algorithm will select the item at

HCB-pHCB/algs/simulator_multi_process.py

Line 128 in cadd262

item = alg.users[user.uid].linucb.decide(user.uid,arms)

Q2: Yes. I am in charge of doing the experiment, and a co-author is in charge of writing the paper, so we may have missed some details. Thank you for pointing out such an issue, we will modify it in a new version.

[greeness]

Thanks for the quick response. Sounds good on Q2.

Regarding Q1, I thought at level L-1, a level-specific base_linucb is still beneficial because this decides which leaf node to choose. If I understand correctly, the linucb at

HCB-pHCB/algs/hcb.py

Line 50 in cadd262

self.linucb = NaiveLinUCB(cfg.linucb_para)

is used to choose the item after the leaf node is chosen.

Any comment? Thanks!

[yusonghust]

Yes, your understanding is correct! I'm sorry for the less rigorous description in the paper. Actually, if the hierarchy tree structure is set to {1, 100, 10000}, which means there are 1, 100, 10000 non-leaf nodes at each level from top to bottom, and the leaf nodes are mounted to a non-leaf node, which is not considered as a certain level.

[greeness]

Thanks for the clarification.