2D-OOB: Attributing Data Contribution Through Joint Valuation Framework

Data valuation has emerged as a powerful framework for quantifying each datum's contribution to the training of a machine learning model. However, it is crucial to recognize that the quality of cells within a single data point can vary greatly in practice. For example, even in the case of an abnormal data point, not all cells are necessarily noisy. The single scalar score assigned by existing data valuation methods blurs the distinction between noisy and clean cells of a data point, making it challenging to interpret the data values. In this paper, we propose 2D-OOB, an out-of-bag estimation framework for jointly determining helpful (or detrimental) samples as well as the particular cells that drive them. Our comprehensive experiments demonstrate that 2D-OOB achieves state-of-the-art performance across multiple use cases while being exponentially faster. Specifically, 2D-OOB shows promising results in detecting and rectifying fine-grained outliers at the cell level, and localizing backdoor triggers in data poisoning attacks.

Paper

Similar papers

Peer review

Reviewer ECZV6/10 · confidence 4/52024-07-12

Summary

This paper studies data attribution. Unlike the previous works that only quantifies the contribution of each datum, this work also considers the attribution of the cells within every datum. The proposed the joint valuation framework, 2D-OOB, can accurately identify cell outliers, as well as the poisoning attacks. The experiments verify the authors claims and show the efficiency of the proposed method.

Strengths

1. The paper is well-written and easy to follow. 2. The proposed method has been theoretically generalized and connected to the existing works. 3. The experiments are comprehensive and show the efficiency of the proposed method.

Weaknesses

1. In section 4.3, the experiments did not mention that the classification label is altered or not after poisoning (e.g, from correct to incorrect). 2. The paper does not state clearly about the differences of the application on image, comparing with the other model attribution methods? They both assign a value to each feature/pixel/cell in the input data. Why would the proposed method be better in this scenario? 3. How about the scalability of the proposed method? The experiments are either conducted on tabular data or small images. The large images or high-dimensional data are not mentioned. 4. What is the distance regularization term mentioned in L168 in the utility function $T$? 5. The size of CIFAR-10 is 32x32, but the cell valuation displayed in Figure 5 is 16x16. Why is that?

Questions

See weaknesses

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Reviewer nD8c5/10 · confidence 3/52024-07-14

Summary

The paper proposes 2D-OOB, an out-of-bag estimation framework for jointly detecting both samples as well as their cells that are outliers. In other words, 2D-OOB allows attribution of the value of a data sample to its individual features. Evaluation shows that the framework is also able to localize backdoor triggers in data poisoning attacks. In comparison to the 2D-Shapley model, which also performs joint valuation of a sample and its features, the 2D-OOB is a more efficient and model-agnostic formulation and builds on the subset bagging model, where each weak learner is trained on a randomly selected subset of features. In the experiments, decision trees as used as the weak learners and trained using randomly selected features. The framework measures (using Eqn. (4)), the average accuracy of the bag of predictors for a selected sample and its features.

Strengths

* The paper is well written. * Given the background in moving fro DataShapley to the Data-OOB formulation, the transition from 2D-Shapley to 2D-OOB was intuitive and easy to follow. * The proposed method can identify data poisoning attack triggers (square patches placed in the original images to elicit misclassification) which could be very useful.

Weaknesses

* The paper is low on novelty. Given, OOB-Data formulation and the 2D-Shapley for joint valuation, the 2D-OOB formulation seems like a natural extension -- interesting, but not high on novelty. * In Fig. 4, panel B, for BadNets poisoning attack, the detection performance of 2D-OOB and 2D-KNN (proxy for 2D-Shapley) are very similar. It is difficult to assess how 2D-OOB will perform for more stronger poisoning attacks. A more detailed study on this aspect as well as analysis showing that 2D-OOB is more likely to identify such triggers, even when attack takes such a detection mechanism into account, will be a good indicator of the strength of this framework. * Table 1, shows the average run time is much improved from 2D-KNN, however, most of the datasets used (Appendix A, Table 3) have small input dimensions (only three have >= 100 but have small sample size) and hence it is difficult to assess how the technique would scale.

Questions

* One of the arguments mentioned in the description of Fig. 2 is that 2D-OOB detects a majority of outlier cells by examining a small fraction of the total cells. It would be great if the authors could further explain this statement and which part of the formulation in (4) motivates this argument. * In Fig. 3 (showing 2D-OOBs ability to identify and prioritize the outliers to fix), the test accuracy for the Jannis dataset is relatively low compared to the others and the performance is also similar to 2D-KNN. Any insights as to why this is happening? * In Section 4.4, were the experiments also performed by varying B using a 2-layer MLP? What is the overhead of training an out-of-bag estimator versus a single model (maybe an autoencoder) and using that for outlier detection. * Are there any other stronger baselines against which comparison can be made?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes.

Reviewer 4K7L6/10 · confidence 5/52024-07-15

Summary

This paper proposes a joint valuation framework for not only obtaining data value but also attributing a data points value to its individual features (cells). The framework is build on top of data-OOB. The authors compared the proposed 2D-OOB with 2D-KDD on several tasks, including cell-level outlier detection and backdoor trigger detection in data poisoning attacks.

Strengths

(+) This paper targets an important problem that is useful for improving the interpretability of data valuation. The joint valuation provides more fine-grained information. Identifying outliers at the cell level is informative for determining which cell to fix in the next step when improving data quality, especially useful when data acquisition is expensive. (+) The proposed method naturally integrates the idea of random forest - random feature selection in the construction of weak learners - with data-OOB, offering a new solution for joint evaluation. This approach addresses a relatively underexplored problem, with few existing works in the area. The proposed 2D-OOB method demonstrates greater computational efficiency compared to 2D-KDD. (+) The improvement in backdoor trigger detection is promising, as cell contamination in this context is targeted and deliberate, unlike artificial random outliers. (+) The authors provide an ablation study to assess the impact of the selection and number of weak learners, as well as the feature subset ratio, which helps in understanding the stability of the proposed method.

Weaknesses

(-) Despite the strengths mentioned above, the entire framework is built on top of the OOB framework. This brings a main limitation, as the OOB framework requires the training algorithm to be a bagging method with weak learners. In many applications, bagging may not be the best-performing method, and data valuation under this method may not be of interest. (-) Typically, the number of randomly selected features for each weak learner is an important hyperparameter to fine-tune. What is the impact when the number of random features is small? While an ablation study was provided in Appendix, certain variation was still observed. Should we correspondingly update the bagging size? Some practical guidance would be helpful.

Questions

1. What does fitting a logistic model in line 187 on page 6 mean? Why change from bagging decision trees to a logistic model in this specific cell fixation experiment? 2. It would be helpful to further explain the difference between Trojan and BadNets trigger methods, as it seems that 2D-OOB is much more effective than 2D-KNN under the Trojan square. This would help us understand in which scenario 2D-OOB works better. 3. Regarding the ablation study: I am also curious about the original prediction performance for each task, not just the outlier detection performance. 4. What is the implication of the connection to data-OOB in Proposition 3.1? Are there any insights we can gain after establishing the connection with data-OOB?

Rating

6

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

None

Reviewer 6nYm5/10 · confidence 2/52024-07-15

Summary

This paper studies data valuation in a fine-grained fashion: determine helpful samples as well as the particular cells that drive them. The proposed method is tested with 12 datasets and has application in backdoor detection.

Strengths

- It makes sense to use cells rather than single data samples as the smallest unit for data valuation. - The proposed framework is clearly stated. - The backdoor trigger detection is an interesting and useful application of the proposed framework. - The proposed method is significantly more efficient than 2D-KNN.

Weaknesses

- The proposed method seems to be a combination of existing algorithms 2D-KNN and Data-OOB. The novelty and contribution may be relatively marginal.

Questions

Could you please discuss what are the *unique* challenges of 2D-OOB that do not arise for 2D-KNN and Data-OOB?

Rating

5

Confidence

2

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes.

Area Chair idkm2024-08-08

Dear reviewers, All the reviews and rebuttals are now available. Please go over all such information and ask any clarifying questions to the authors.

Reviewer 4K7L2024-08-11

Thanks for the response

I appreciate the authors' response. Overall I think this is a solid paper, although its impact may be limited by its applicability to bagging methods only.

Authorsrebuttal2024-08-14

Thank you

We appreciate your thoughtful review again and are pleased you found the paper solid!

Reviewer ECZV2024-08-13

Thanks for the authors' response, which has addressed most of my concerns. The authors should include the key points of the discussion into their revised version.

Area Chair idkm2024-08-13

Dear authors, Are you planning to revise the paper accordingly?

Authorsrebuttal2024-08-13

Thank you!

Dear AC and Reviewer ECZV, We are grateful for your feedback and suggestions. We will make sure to include necessary clarifications, discussion on our method and image feature attribution, as well as scalability results in the revision.

Reviewer nD8c2024-08-14

Thank you for the detailed response. I am satisfied with the scalability experiments as well as the results on the attacks. I am increasing my score to 5.

Authorsrebuttal2024-08-14

Thank you

We greatly appreciate your constructive review and are glad our new results addressed your concerns. We will incorporate all these new results and discussions in the revision. Thank you for re-evaluating our paper.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC