Active Preference Learning for Ordering Items In- and Out-of-sample

Learning an ordering of items based on pairwise comparisons is useful when items are difficult to rate consistently on an absolute scale, for example, when annotators have to make subjective assessments. When exhaustive comparison is infeasible, actively sampling item pairs can reduce the number of annotations necessary for learning an accurate ordering. However, many algorithms ignore shared structure between items, limiting their sample efficiency and precluding generalization to new items. It is also common to disregard how noise in comparisons varies between item pairs, despite it being informative of item similarity. In this work, we study active preference learning for ordering items with contextual attributes, both in- and out-of-sample. We give an upper bound on the expected ordering error of a logistic preference model as a function of which items have been compared. Next, we propose an active learning strategy that samples items to minimize this bound by accounting for aleatoric and epistemic uncertainty in comparisons. We evaluate the resulting algorithm, and a variant aimed at reducing model misspecification, in multiple realistic ordering tasks with comparisons made by human annotators. Our results demonstrate superior sample efficiency and generalization compared to non-contextual ranking approaches and active preference learning baselines.

Paper

References (86)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer tuDy5/10 · confidence 3/52024-07-02

Summary

This paper proposes an active learning algorithm for selecting pairs of items for comparison in order to learn a ranking of the items. The ranking error is measured by the (normalized) number of swapped pairs compared to the true ordering (Kendall’s Tau distance). The algorithm chooses pairs of items based on an upper bound on this ranking error. Experiments are conducted on one synthetic and 3 real-world benchmarks.

Strengths

* Active learning for pairwise comparisons is an interesting problem to study * The approach is justified by theoretical analysis * The empirical evaluation looks promising

Weaknesses

* The paper uses strong assumptions on the pairwise comparisons. Specifically, the response is assumed to be: $P(C_{ij}=1) = \sigma(\theta_*\cdot(x_i-x_j))$, where $\sigma$ is the sigmoid function (eq 2). These assumptions are used for fitting MLE parameters. These assumptions are common in the literature. * The optimization problem for choosing a pair (eq 5 and 7) seems hard. The cost is quadratic in the number of items since all pairs are considered. An approximation that depends linearly on the number of items would make the approach more practical.

Questions

* “We restrict algorithms to only query pairs for which an annotation exists and remove the pair from the pool once queried.” – What is the number of available annotations for each dataset in figure 2 in the experiments? Only the number of items n and the dimension d are specified. Also, there is no notion of annotation noise here, since the labels are set, right? * What is $\Delta_*$? It seems from line 159 that: $\Delta_*=\min_{i,j} \Delta_{ij}/|i-j|$, but it would be better to define it explicitly. Does it really depend on the index difference $|i-j|$? What is the reasoning behind this? Minor: * Line 74: “Ordering algorithms based only on preference feedback cannot solve this problem since observed comparisons are uninformative of new items.” This statement is not clear, as long as items are represented as attributes/features $x_i$ then comparisons may be informative for new items.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

I suggest adding the assumption on response (eq 2) and the $O(n^2)$ complexity to the list of limitations.

Reviewer 6ajg5/10 · confidence 3/52024-07-11

Summary

This paper considers the ranking problem based on pairwise comparisons. The goal is to get the best sampling strategy for the best ordering from a limited number of comparisons. Under a logistic model on the difference between scores, the authors provide the analysis for the upper bound on the ordering error, which provides insights on sample selections. Following the idea of minimizing the bound, this paper proposes the GURO algorithm for pair selections. The proposed method is evaluated in four image ordering tasks with either synthetic labels or real world labels.

Strengths

The proposed algorithm is well motivated by the theoretical result on the ordering error. It has very strong theoretical guarantees on the performance. The theory presented in the paper looks good to me. And it helps the reader to understand the algorithm better with some justification from Bayesian analysis. I find the paper very well written. The way the authors presented the results is very clear and easy to follow.

Weaknesses

Although the theory presented in the paper looks good to me, I find it very similar to the result presented in the original Logistic bandit paper [1]. By treating the input space as the difference between features, the problem is simplified to a standard logistic bandit problem. And The result in Lemma 1 and some analysis before Theorem 1 are very similar to Lemma 2 and 3 in [1]. While I understand Theorem 1 is specifically for the ranking error, I think it is still straightforward to get Theorem 1 from existing lemmas. And for the empirical study, I see the proposed approach actually does not always perform better than baselines, especially BALD. The good result is on synthetic data, there is no significant advance on real data. Therefore, I am not convinced with the claimed statement in the paper.

Questions

Please provide more discussion on how difference the proposed method is comparing to logistic bandits.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes, the authors touch upon the limitation in the conclusion section.

Reviewer nNXZ4/10 · confidence 3/52024-07-13

Summary

Active preference learning is different from deriving a complete ordering from preferences. It focuses on “If we collect comparisons D_T, how good is the resulting model’s predicted ordering in the item set”. The paper proposes a sampling method in the active learning scenario. Theoretical analysis is also provided.

Strengths

--Preference learning is critical to many downstream tasks. --The proposed method is somewhat novel. --The theoretical analysis is provided. --Experimental results are shown to verify its effectiveness.

Weaknesses

--The assumption 1 and 2 are not so intuitive. It is better to illustrate an example. --Baseline methods are weak. Though many related studies are mentioned in the related work section, performances of baselines are not shown in the experiments. --The number of comparisons is not reduced tremendously on the ImageClarity in Table1. --Performances should be emphasized in terms of the prediction ordering quality.

Questions

State-of-the-art baselines should be added for performance comparison.

Rating

4

Confidence

3

Soundness

2

Presentation

3

Contribution

2

Limitations

Limitation should be added.

Reviewer LY2M8/10 · confidence 3/52024-07-13

Summary

This paper considers the setting of learning an ordering between items according to some scoring rule. The assumption is that this ordering is determined by a contextual scoring rule, determined from the features of each item. Such contextual structure can aid in more rapidly learning an ordering, and generalizing to out of sample items. This ordering is learned from asking pairwise preferences to an oracle, reducing uncertainty about the total order. Since there are a large number of possible comparisons to be asked, active learning is deployed to only query labels for a subset of comparisons. A theoretical argument is made about the optimal balance between aleatoric and epistemic uncertainty to target in adaptive selecting queries, motivating the GURO adaptive sampling strategy and variants. The performance of GURO is demonstrated in several empirical experiments on simulated data and data collected (offline) from real humans.

Strengths

I believe this paper is excellent - it is *clearly* written, has a great flow, and theoretical and empirical arguments are tied together nicely to motivate the problem setting, establish the problem fundamentals, convey mathematical intuition about uncertainty reduction, and justify active selection. There is also a robust set of experiments demonstrating GURO (and variants) in practice against baselines, along with explanatory discussion and implementation details. To my knowledge the analysis and algorithmic ideas here are *original*, and this is a *high quality* submission. Although not the centerpiece of this work, in an age where RLHF and efficiently learning from human preferences is paramount in training large models and ranking queries, work in active preference learning and the contributions made here are *significant*. There is also a robust and thorough appendix providing details and theoretical proofs (disclaimer: I have read the main paper in careful detail, but only skimmed the appendix). Overall, this is elegant, interesting, and impactful work (both theoretically and empirically).

Weaknesses

I do not have any explicit weaknesses to list. Instead, I have a list of comments and questions below that I would like the authors to address. However I feel confident that these can be addressed during the rebuttal phase. One comment is that there is no discussion (unless I missed it) about computational complexity stating and comparing the big-O computational complexity of GURO, its variants, and other selection methods. This would be an interesting and strengthening addition to this work.

Questions

- I think the statement "Moreover, the set we want to order is often larger than the set of items observed during training—we may want to rank new X-rays in relation to previous ones. This cannot be solved using per-item scores alone." should be clarified. If one knew absolute scores for all items (regardless if they are observed in training), isn't it trivial to compute an ordering? Or did the authors mean that pairwise responses collected during training could not generalize outside of training, without example features to predict from? [Edit: this does seem to be clarified in Section 2, but should be made more clear in the introduction] - I find the sentence "However, as we show in Section 4, learning this map to recover a complete ordering is distinct from the general preference learning setting, and existing algorithms lack theoretical justification for this application" to be vague and should be clarified. What exactly is the "general preference learning setting", versus learning a map to recover complete orderings? What does "this application" refer to? Which of these two settings are you concerned with here? - I'm confused by line 159. Why can one lower bound $\Delta_{ij}$ by a factor depending on the index difference $\lvert i -j \rvert$? Aren't the indices arbitrary, and agnostic to the underlying geometry of the feature space? Does this mean that a simple index permutation would drastically change this $\Delta$ quantity? - in line 166, should the dependence not be on $\theta_T$ rather than $\theta_*$?. See line 157 which uses $\theta_T$. line 169 also jumps back to $\theta_T$ - line 189 is missing an important point: by definition the entirety of $\mathcal{I}$ is unavailable, only $\mathcal{I}_D$ is available to select from. This should be commented on. - I think line 198 is too vague: "As θt converges to θ∗, this pair becomes representative of the maximizer of (4) provided there is no major systematic discrepancy between ID and I." Can you comment more on what constitutes acceptable vs unacceptable discrepancies between I and I_D? This does bring up a problematic point: what if $I_D$ is not sufficiently representative of $I$? Line 215 starts to hint at this discussion but I think it needs to be elaborated on, ideally more formally - in GURO Hybrid, how are these $\zeta_i$ parameters actually learned? Is it just a joint MLE on $\theta$ and $\zeta_i$? In this case, what prevents the model from learning an arbitrary $\theta$ (i.e., $\theta = 0$) and just using the full expressivity of $\zeta_i$? Is there some sort of regularization in practice? - for completeness, can you include a figure in the Appendix showing the experiment in Fig 1b, but just plotting $R_{I_D}$ instead of the difference? It would be good to know how each algorithm does in an absolute sense on $I_D$ - figure 2 would benefit from a log y scale - it is very difficult to discern between methods Minor: - line 147 uses the notation $\widehat{H}$ instead of $\widetilde{H}$. I assume this is a typo - missing left parentheses in line 5 of GURO algorithm - the word BayesGURO in Algorithm 1 should also probably be colored green to show the association to (7) and (11) - be careful with red and green as distinguishing colors for readers with color vision deficiency - fix quotes on line 258

Rating

8

Confidence

3

Soundness

4

Presentation

4

Contribution

4

Limitations

Yes

Reviewer LY2M2024-08-12

Thank you for your response. I am satisfied with these points and leave my review unchanged (at an 8). Also, please double check the definition of $\Delta_*$ in line 125. It is missing the quotient you defined above. It just says $\Delta_* = \min_{ij} \Delta_{ij}$

Reviewer tuDy2024-08-13

Thank you for the clarifications.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC