Most Influential Subset Selection: Challenges, Promises, and Beyond

How can we attribute the behaviors of machine learning models to their training data? While the classic influence function sheds light on the impact of individual samples, it often fails to capture the more complex and pronounced collective influence of a set of samples. To tackle this challenge, we study the Most Influential Subset Selection (MISS) problem, which aims to identify a subset of training samples with the greatest collective influence. We conduct a comprehensive analysis of the prevailing approaches in MISS, elucidating their strengths and weaknesses. Our findings reveal that influence-based greedy heuristics, a dominant class of algorithms in MISS, can provably fail even in linear regression. We delineate the failure modes, including the errors of influence function and the non-additive structure of the collective influence. Conversely, we demonstrate that an adaptive version of these heuristics which applies them iteratively, can effectively capture the interactions among samples and thus partially address the issues. Experiments on real-world datasets corroborate these theoretical findings and further demonstrate that the merit of adaptivity can extend to more complex scenarios such as classification tasks and non-linear neural networks. We conclude our analysis by emphasizing the inherent trade-off between performance and computational efficiency, questioning the use of additive metrics such as the Linear Datamodeling Score, and offering a range of discussions.

Paper

References (81)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Uyvf7/10 · confidence 4/52024-07-03

Summary

The paper presents a theoretical analysis of why existing greedy additive methods fail to solve the most influential subset selection (MISS) problem, which aims to find the subset of training data with the largest collective influence. Greedy additive methods (first assigning individual influence scores, ordering and taking the top-k group for example) assume linearity of collective influence and fail to account for the non-linearity due to interactions between samples of the group. Building on the analysis, the paper proposes an adaptive greedy algorithm where the individual influence scores are dynamically updated to capture interactions among samples and ensure that samples of the subset have consistent influence (sign + order of influence are preserved). This algorithm is demonstrated using synthetic and real-world (MNIST) experiments.

Strengths

- The paper is very well written. - The paper contributes a thorough theoretical analysis of failure modes in existing greedy approaches to solve the most influential subset selection (MISS) problem and therefore shows why these approaches often fail to find meaningful subsets. - The problem of subset influence is interesting and increasingly relevant given the increasing scale of datasets. The paper is therefore not only relevant for the area of data influence but also data-centric AI overall.

Weaknesses

This paper is mainly a theory paper, focusing on the failure modes of influence-based greedy heuristics for most influential subset selection (MISS). The choice of datasets and experiments to run (synthetic data, MNIST on MLPs) fits the scope of the paper. Yet, if I understand correctly, the proposed alternative approach to solve MISS dynamically with the adaptive greedy algorithm is shown for a subset size of 2 and does not scale to larger subset sizes. In Remark 4.3 (line 288), the authors hypothesize that the method would scale to subsets larger than 2 if the found subsets are truly already belonging to the most influential subset. Given the fragility of influence scores (Basu et al., 2021), I doubt that the scenario in the authors' hypothesis is realistic. I am unsure though if I understood the claim in this remark correctly.

Questions

The paper was very clear, and I only have a few questions to gain a better understanding of the potential impact: - See weaknesses. - You mention the second-order group influence functions by Basu et al. (2020) in your discussion as an alternative approach to compute group influence. How does the adaptive greedy approach you suggest compare to the second-order approach by Basu et al. (2020) in terms of computational efficiency (when considering group sizes >2, too)? Is it a faster alternative? - While I acknowledge that this is mainly a theory paper, I would like to understand the potential impact better. What would be potential application scenarios for finding the **most** influential subset?

Rating

7

Confidence

4

Soundness

4

Presentation

4

Contribution

2

Limitations

Yes, limitations are discussed in detail in the discussion section.

Reviewer 3kwV2/10 · confidence 3/52024-07-12

Summary

The authors investigate the Most Influential Subset Selection (MISS) problem, which aims to identify a subset of training samples with the greatest collective influence on machine learning model predictions. They discuss limitations of prevailing approaches in MISS and highlight issues with influence-based greedy heuristics. The paper proposes a new method to greedily select the k most influential samples, utilizing adaptive iterative updates to the importance weight of remaining samples that has not been selected by the greedy procedure.

Strengths

- The concept of adaptively updating the influence weight per sample is reasonable. - The notion of influence is sensible. - The submodularity approach to studying data subset selection is reasonable. - The authors discuss the limitations of non-adaptive greedy heuristics for MISS. - The paper includes an in-depth example and visualizations of the shortcomings of regular greedy algorithms for MISS problems, clearly demonstrating the problems. - The implementation is publicly available. - The technical appendix appears to follow a sensible proof strategy; however, I haven't had the chance to formally check the appendix (disclaimer).

Weaknesses

- By their nature, being greedy procedures that iteratively re-evaluate the influence of each individual sample, the proposed algorithms are too inefficient to be applicable to even moderately sized datasets or datasets where subset selection matters most. Consequently, the authors mostly consider small datasets in their experiments. - As reported in D.5, k-MISS-MLP-MNIST took 28 hours, which is quite wasteful. - The paper does not formally discuss potential problems and limitations with the applicability of k-MISS in the context of non-linear models; therefore, it remains unclear whether the influence function is truly applicable to non-linear models where samples may have unknown non-linear future influence. - The influence of randomness from stochastic gradients and initializations (of the MLP) is not discussed or modeled. - The paper does not convincingly demonstrate that the proposed solution has practical relevance. For example, there is no case study that demonstrates the practical usefulness of the reported results. - The paper does not convincingly demonstrate that the proposed solution advances our theoretical understanding of MISS. - The work lacks contextualization regarding highly related research areas, including Data Subset-Selection, Active Learning, Data Shapley, Importance Sampling, Landmark Selection, and Core-set Selection. - The paper does not compare the proposed solution to established baselines and highly related data, such as SELCO [http://proceedings.mlr.press/v139/s21a/s21a.pdf], which employs submodularity-based approximation strategies to jointly estimate coefficients and the most relevant data subset. - The paper does not recognize highly influential prior work on Data Subset Selection [http://proceedings.mlr.press/v37/wei15.html]. - The dataset selection lacks representativeness. - The experimental assessment is based on too few datasets (Concrete Compressive Strength, Waveform Database Generator, and MNIST). - The test set sizes are very small. - It remains unclear how robust the proposed algorithm can deal with variations in hyperparameters. - It remains unclear how robust the notion of influence can handle random effects from initialization and gradient batching. - The experiment does not report any robustness checks, such as sensitivity analysis or cross-validation. - The experiments do not include a comparison with state-of-the-art methods from data subset selection or Data Shapley, which makes it difficult to assess its relative performance.

Questions

./.

Rating

2

Confidence

3

Soundness

3

Presentation

2

Contribution

1

Limitations

- The authors haven't discussed the implications data sample selection from senstive data: If the data is user-generated, non-private, and sensitive, this procedure might select and flag data records of certain individuals. - The experiments are limited and do not showcase practical limitations under different circumstances (see above)

Reviewer 2zFQ7/10 · confidence 4/52024-07-15

Summary

The paper explores the challenge of understanding the collective influence of subsets of training data on machine learning models, referred to as the Most Influential Subset Selection (MISS) problem. Traditional influence functions, which focus on individual data points, often miss the more complex interactions within subsets. The authors analyze existing influence-based greedy heuristics, revealing their potential failures, particularly in linear regression, due to errors in the influence function and the non-additive nature of collective influence. They propose an adaptive version of these heuristics, which iteratively updates sample scores to better capture interactions among data points. Their experiments on both synthetic and real-world datasets validate the theoretical findings, demonstrating that the adaptive approach can extend to more complex scenarios like classification tasks and non-linear models.

Strengths

- Addresses a timely important problem of how to leverage pointwise influence estimates to remove datapoints (dataset selection is a very important topic imo, and a related work of [1] recently won a best paper at ICLR). Even if some of the results were known in the literature (see weaknesses), the paper provides a clear story and analysis. - Exceptional exposition for the most part; very lucid and clear narrative, and addresses both high-level and subtler points (for example, section 6 is a nice discussion of natural follow-up questions)

Weaknesses

- My main concern is that the results presented in Section 3 are sort of a "folklore" in the IF literature (see, e.g., [1] and the references therein). The under-estimation of group effects is also analyzed in earlier work [2], for example. [1] https://arxiv.org/abs/2309.14563 [2] https://arxiv.org/abs/1905.13289

Questions

- I understand that re-training was modified for the MLP experiments due to computational constraints, but I'm very curious how doing the full correct version would further improve the results in Figure 4. In particular, given the nature of NNs to overfit, I would suspect that partial re-training would not be enough to fully "erase" the influences of deleted points. - Though the paper focuses on MISS, it would be nice to comment a bit on related problems of dataset selection and coreset selection (which is more of an LISS) - Figures 1 through 3 should have axes labels. Also, would give more intuition of the examples in the original feature space are also shown.

Rating

7

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

Satisfactory

Reviewer btdJ7/10 · confidence 3/52024-07-21

Summary

The paper explores the problem of most influential subset selection, that is, selecting a subset of training data points whose removal would change a machine learning model the most. The paper develops further on previous works (most notably of Chatterjee and Hadi, and Kuschnig et al) and shows why the greedy heuristics do not work well in practice. Greedy heuristics operate by computing the influence (e.g., following Koh and Liang) and simply selecting the data points with highest individual influence. The paper shows how indirect interactions can lead to under- and over-estimation of influence. The paper then shows why a specific adaptive method works well in practice. The analysis of the paper is limited to OLS. While the paper does not propose a new influential subset selection algorithm, the analysis conducted in the paper adds important theoretical insights that are likely to be helpful to the research community.

Strengths

1. The paper does a very good job of setting up the problem and explaining the main results. The formalism is clean and easy to understand and the results are accompanied by intuitive explanations, e.g., in Figure 1. 2. The key limitation of greedy heuristic methods (considering the full training dataset) is explained quite well in Section 4. 3. While the paper does not quite add an algorithmic improvement, the theoretical analysis provided is an important building block in developing our understanding of the underlying mechanisms.

Weaknesses

1. Some of the analysis is limited to simpler models like OLS and takes assumptions like the inevitability of the $N=X^TX$ matrix (which may not be the case in real world datasets). This per se is not a big weakness. However, some discussion here on how we expect these results to behave on real world datasets where $N$ is non-invertible, or on models beyond OLS would be greatly helpful. 2. The result of theorem 4.2 seems a bit limited and too tailored to the adaptive greedy algorithm. Specifically, given the steps described in line 251, (remove most influential sample, retrain the model, repeat), and the proposition 4.1 (order preservation between two points), it is not immediately clear if/how much the Theorem 4.2 would extend to 3-MISS and beyond. 3. The title of the paper seems to be quite general and a bit mismatched with the content. While the content shows the weaknesses of existing influence estimation approaches (that too with a OLS model, in a 2-MISS settings and a very specific adaptive algorithm), reading the title feels as if the paper would address a much broader range of models and algorithms.

Questions

1. Section 3.1: To what extent is the finding in Section 3.1 limited to the specific target function used here, that is, the prediction on $x_{test}$? After all, reading the legend in Figure 1, there is not a huge difference on the target function value between points 1 and 8 (0.120 vs. 0.117). More broadly, do we expect the effect of high leverage points to be this large when we take “less noisy” functions like the sign of the prediction on $x_{test}$ or accuracy on the whole test set? 2. Line 171: About correcting the influence with its corresponding leverage score, how does that process look like precisely? 3. If the reviewer understood correctly, $\bar{A_{-S}}$ in Figure 3 shows the actual influence of the two approaches considered. Why not baseline this comparison with the “ground truth influence”, that is, the influence of the actual most influential data points? Clearly, this will ground truth will be very difficult to compute as $k$ increases, but some understanding over smaller values of $k$ might be very helpful for the reader. 4. The results in Figure 4 are mostly as one would expect. However, in the case of MLP, the gap in $\bar{A_{-S}}$ and winning rate starts shrinking around $k=50$. Any reason why that might be happening? How robust are the results to different choices of training data random seed?

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

The paper is generally quite open about the limitations. However, please see the point 3 in weaknesses about the generality of the title.

Reviewer Uyvf2024-08-08

Thank you for the detailed response. I appreciate the author's suggestion to clarify that their presented method addresses the 2-MISS problem and offers a starting point for the solution for the more general k-MISS problem and believe it will be useful for readers to assess the scope of the paper. My main concern was the potential application impact of this work, where the author's response convinced me of practical scenarios. Hence, I raise my score from 6 -> 7.

Reviewer btdJ2024-08-14

Raising my score

Thank you for the detailed response. Most of my concerns were addressed and I am raising my score to accept. It would be great to add the additional discussion here on target functions and robustness to the final version of the paper.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC