Summary and Contributions: The paper takes the approaches commonly which commonly utilizes black box predictors to estimate label shift and then proposes a common framework through which one can analyze them. The paper, then proceeds to prove the conditions of consistency for these estimators, given estimation error analysis of the two methods (BBSE and MLLS) and shows why BBSE methods under perform as compared MLLS method.
Strengths: The paper casts the MLLS with a calibrated predictor as a distribution matching problem and notices that BBSE method can also be casted in the same way, thereby giving a unified view. The theoretical analysis of the consistency of these estimators looks sound. The paper proposes an explanation of why BBSE methods might be doing worse than MLLS and then proceeds to give empirical justification for the same. Prior work like RLLS give finite sample error bounds on the estimate. However, this work additionally shows that, it is not just the finite sample error that is in play, there is a role of how accurate the calibration is, which also plays a role in the final estimation error, which makes sense.
Weaknesses: The work currently restricts the analysis to black box predictors. However, it might have been possible to analyze even white box predictors, by assuming some structure on f.
Correctness: The claims made in the paper look correct, the proof of the theoretical analysis looks correct. The experimentations done are mainly to support some of the theoretical claims made and they look correct too.
Clarity: The title feels a little broad compared to the work. The work is focused on label shift estimation using black box predictors. It would be good to modify the title to reflect this scoped problem. Otherwise, the paper is well written and easy to understand.
Relation to Prior Work: There are no prior works that theoretically analyze MLLS and also provide a common frame work for both MLLS and BBSE estimators. The paper does, however, give prior works like RLLS which is a BBSE estimator and shows why MLLS methods may do better than such estimators.
Additional Feedback: In the code, there are some hardcoded paths - which may not hold true on all machines. The README does not give instructions on how this paths need to be updated. Execution of the code takes a long time with no print statement, making it unclear as to what is happening. Some idea of what kind of machines were used to execute this code and what is the expected time to run the code would have been useful. ---- Post Author Response: Thanks to the authors for their response. I have read through the other reviews and also the responses of the authors. After going through them, I am still feeling quite positive about this work and will retain my earlier evaluation of acceptance of the paper.
Summary and Contributions: In this paper, authors provided a unifying framework which summarizes two major approaches in label shift estimation: BBSE and MLLS. The consistency and estimation error are proved for MLLS. The result implies that BBSE's inefficiency is probably due to a loss of information when calibrating the confusion matrix.
Strengths: The first strength of this paper is it summarizes MLLS and BBSE using a distribution matching framework. To my best knowledge, this is the first time such a framework has been proposed, although it is not a hugly surprising result given BBSE and MLLS are both based on label shift assumption (p_sX|y = p_tX|y). However, this derivation allows authors later to derive the consistency for MLLS using classic maximum likelihood results. The second contribution is that the estimation error of MLLS of can be decomposed into finite sample approximation error and miscalibration error. However I am not sure how the calibration error can be determined in practice which would be really useful. I imagine this is something like an oracle inequality that cannot be quantified? It is also important that authors have pointed out that the calibratedness of the predictor f as a sufficient and necessary condition to the consistency of MLLS which aligns with previous empirical observations. The unified framework gives rise to MLLS-CS which has a similar performance to BBSE. This is a very interesting result as it shows the differnece in terms of perforamnce is due to the aggregation of confusion matrix. This is a very important mesage and is not knwon to the community to my best knowledge.
Weaknesses: I do not find major flaws in this paper. A few minor points: 1. Should line 165 be E_t [log p_s(z)] not E_t [p_s(z)] ? 2. Figure 2 is hard to read: which lines correspond to the left scale and which lines correspond to the right scale? 3. Line 323, Figure (6) -> Figure 1?
Correctness: The method and claims seems sound.
Clarity: Thisa paper is well-written and I do not have troubles following the main ideas of this paper.
Relation to Prior Work: This paper is not directly related to previous works and other related works are properly referenced.
Additional Feedback: Thanks authors for replying our comments. I do not have further questions.
Summary and Contributions: This paper studies the label estimation problem and unifies a previously proposed perspective--maximum likelihood and calibration-- with the recent method using black-box predictors. The main takeaway is that these two perspectives can be regarded as the same framework and the calibration is a necessary step to achieve better performance. In the analysis, the weight estimation error is analyzed by decomposing it into an estimation error due to finite samples and an calibration error due to label shift. The empirical evaluation demonstrates that the combined method (MLLS with confusion matrix) outperforms using only black-box predictors. ======= Thanks for the author response. I keep my score after reading it. Some of my concerns are resolved. The following to remains: The first is about the "miscalibration error", which I actually mean the gap between target error (the RHS in lemma4 part 1) and the calibrated source error (the RHS in lemma4 part 2). Intuitively, the larger the source and target difference are, the larger the gap is. Eventually, the bound is presented using calibrated source error, which can be very loose. The second is about weight estimation and downstream tasks. It is true that most previous analysis combines weight estimation analysis with a standard IW learning algorithm analysis and obtain similar guarantees as lemma 5. However, it is not validated empirically in this paper that the proposed method would benefit the downstream tasks. Also, even in previous work, there are practical cases when this reflection in the improvement of learning is not obvious. I encourage the authors to further improve the paper.
Strengths: The label shift estimation problem is very relevant to the community. The perspective of the unification of two perspectives in label shift estimation is novel and interesting. The derivation and analysis in this paper is sound and clear.
Weaknesses: The proposed method is not very novel, since it builds on the previous two perspectives. So if looking at the specific steps for the algorithm, they are not new to the community. However, given that, I still think the paper would benefit from having an algorithm to better summarize the proposed method. The analysis of the miscalibration error is not very informative. The miscalibration error can be quite large if the source and target data differ a lot. So it seems this analysis does not support the argument that the calibration is the main reason that MLLS outperforms methods using black-box predictor. A good label shift estimation does not mean it would perform well in shift correction or other down-stream tasks. This paper only focuses on half of the problem while in reality, it is the end goal (adaptation/learning with the weights) that matters.
Correctness: The technical content is mostly correct. Regarding the calibration and the performance I have the following concerns: This paper uses empirical results and a synthetic example to show that the performance gain of MLLS is due to the calibration, and sometimes it is necessary. This is a reasonable claim but there is still a gap in this question, which is when and why. Since the calibration in reality is not perfect, then the questions are when it is necessary and why in practice, it seems to be always better. Does it have anything to do with the label shift magnitude? Can we do a careful ablation study to see when calibration using source data would fail? A related point is also that BCTS-calibrated method is dominating in the experiments, which is not the focus of discussion in the paper. This mismatch between experiments and theoretical analysis make me feel that the claims and insights from the analysis is not validated in the experiments.
Clarity: The paper is generally well-written.
Relation to Prior Work: This paper surveys related work extensively and discusses connections and differences with them clearly.
Summary and Contributions: The authors explore a number of prior arts in supervised learning under label shift with a focus on Black Box Shift Estimation (BBSE) and Maximum Likelihood Label Shift (MLLS). They show how calibration is an essential step to obtain good properties for MLLS. They show that both BBSE and MLLS can be regarded as two instances of a more general class of label shift correction mechanisms that rely on (i) designing a latent space for calibration and (ii) use a distribution matching technique in this latent space. Specifically for MLSS, the authors also provide error bounds to take into account the finiteness of the sample which induces an estimation error of the parameters of the latent space distribution family and in the calibration procedure.
Strengths: The paper is theoretically well grounded and addresses an interesting practical ML problem. The claims are supported by empirical evidence. References and benchmarked methods are up to date and the novelty of the contribution with respect to them is clearly stated.
Weaknesses: In my opinion, the main weakness of the paper is that the addressed problem is one of the simplest form of data generative distribution mismatch between train and test times. The label shift assumption is a strong assumption which in practice may not hold. The authors do not discuss how the label shift correction framework they established could be plugged into more general domain adaptation techniques.
Correctness: I have checked some of the proofs in details and found them correct. The empirical methodology is adapted to the tackled issue.
Clarity: The paper reads pretty well. In my opinion, there is only one odd wording: I suggest to replace "absent of .." by "in the absence of .."
Relation to Prior Work: The relation to prior art is clear as the paper is meant to bind some of these arts under a common framework and draw new conclusions on their respective abilities to address label shift correction.
Additional Feedback: Even if I would have liked to see some perspectives on more general forms of domain adaption, I believe the paper is insightful to the ML community and contains interesting take-home-messages. ---- Post Author Response: the authors have provided a fair feedback to some of my remarks. After reading their answers and the other reviews, I have the feeling the paper qualities outweigh its shortcomings. I therefore maintain my score.