Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Summary: This paper studies the problem of regret minimization for low-rank linear contextual bandits. In the setting considered, there are k arms, each with an unknown d dimensional vector where k and d are assumed to be large. However, the k by d matrix of arms’ feature vectors is assumed to be low rank with rank r, making some amount of information sharing for estimating different arms possible. At each time step, the algorithm proceeds as follows. First, a context is drawn iid from a distribution, similar to existing work. The algorithm uses an estimate of the k by d matrix (referred to as B) to select the optimal arm to pull and observes a reward which is the inner product of the selected arm and the context with some additive, sub-Gaussian noise. The objective of the algorithm is to minimize cumulative regret. It achieves this via a two phase algorithm, where some initial random exploration builds an estimate of B used to toss away definitely suboptimal arms, and then an exploitation phase for the remainder of the arms. The paper achieves regret scaling like r^2(k+d)log(T), better that the sqrt(kd) log(T) regret necessary without a low rank assumption. Strengths: The paper is generally well written, and the first three sections are well laid out. The flow is logical and shows how to construct tighter estimates of individual arms, and then just plugs this result straight in. Furthermore, the result is a strong one and it can easily be seen the effect of the rank both on the algorithm itself and the resulting bound. Additionally, the row-wise estimation bounds are of separate interest and could have impact in other problems outside the bandit setting. All in all, a fun and interesting paper to read. Questions and comments: - The authors mention the issue of having different numbers of samples for different rows due to the adaptive nature of the algorithm, and note that this breaks down guarantees for one of the estimates of B. This seems like a delicate and important issue, but I found the description of how it is solved to be lacking. - Lines 70-74 are redundant. Perhaps one is an older draft that should have been deleted? - For readers coming from the (non-contextual) linear bandit setting, achieving log(T) regret as opposed to sqrt(T) regret may be a bit surprising. Perhaps you can add a sentence in the related work section as to why this is expected and reasonable beyond directly citing previous work. - In the contributions, you list an implementation detail, noting that the estimation subroutine need not be run to completion, but only until a certain property is satisfied (eq. 2, presumably?). To what extent do you think this is a theoretical artifact versus an actual prescription for someone wishing to implement this? - To what extent is assuming knowledge of a gap common and reasonable? In other bandit setting this can have strong impact on the guarantees. I guess in this case, assuming a smaller gap, would simply lead to more exploration early on. - As a matter of interest, to what extent do you think the regret bound is tight? In the matrix completion literature, the sample complexity can scale like O(r(k +d)). Where does the extra factor of r come from in your bound and is it tight? - NeurIPS seems like a kind of a strange setting for a paper of this length compared to COLT, for instance. I feel like by the end of the main body, we really only have the algorithm set up, and then the analysis follows a blitzkrieg of assumptions and a quick theorem statement which makes the reading strange. _________________update___________________ Thank you for taking the time to address my concerns and answer the questions. I appreciated the simulation given in the author response and think it would make a valuable addition to the paper. I maintain that the structure could do with some rearrangement if possible to address how compressed section 4 is, and hope to see a longer form of this released somewhere as well.
This submission studies the MAB in high dimension case with the low-rank structure and proposes an algorithm with a scale of regret r^2*(k+d). Comments: -add related references (1-3) whose cluster size is equivalent to the rank and sharing the same spirit which should be discussed at least, and (1) is also like a way in alternating minimization. References: (1) Collaborative Filtering Bandits, (2) On Context-Dependent Clustering of Bandits, (3) Distributed Clustering of Linear Bandits in Peer to Peer Networks -add a section to describe to your main assumptions and motivate them -add a section to articular what's the main theoretical difference with (6,15) as well as suggested advances -section 3 needs to significantly shrink to save some spaces for the experimental part which is also crucial to the success of this draft -section 4 seems incomplete, add a proof sketch would be sufficient for readers to get the main view -change your name from REAL to RELR, etc, that better represents proposed algorithm -move two-phase bandit algorithm from appendix to main text -add an experimental section to verify the proposed claims -add an explanation regarding lemma 5 and 6 and their relationship Overall, I went through the whole submission, and found that your theoretical contribution is expected, though needs a better organization and polish, whereas currently your empirical contribution is none, which is hard to go through at this venue, this you are highly encouraged to provide, and there are some jobs need to be done to improve the quality of this draft, in short, my recommendation is this submission with current content is slightly below the acceptance bar, but I am open to change my score when I see their response. %%%updated comments I read all comments and the response from the authors, my concern has been addressed, and my overall recommendation is an acceptance.
it seems that REAL-estimator is an post-processing after standard low-rank optimization. If so I didn't quite get what's the significance in the technical improvement. The paper is clearly written but not compact. I have lists a few confusing/unclear points in the improvement section.