Research map / Point clouds / Point cloud classification
2D-OOB: Attributing Data Contribution Through Joint Valuation Framework
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes 2D-OOB, an out-of-bag subset-bagging framework that assigns a value to each cell (data point × feature) to jointly localize detrimental samples and cells, and shows it is much faster than 2D-KNN on outlier detection, cell fixation, and backdoor trigger localization.
Problem
Existing data valuation methods assign one scalar score per data point, so they cannot tell which cells within a point are noisy or poisoned; prior joint (cell-level) valuation via 2D-Shapley is computationally infeasible at scale.
Approach
Extends Data-OOB to a subset bagging model with random feature subsets per weak learner; each cell (i,j) is scored by the average out-of-bag accuracy of weak learners that are OOB for point i and include feature j. This is shown to equal a conditional expectation of Data-OOB over feature subsets, giving a single-model, model-agnostic joint valuation.
Key results
- On 12 tabular datasets, 2D-OOB achieves average cell-outlier detection AUC 0.83 vs 0.67 for 2D-KNN, with average run-time 6.78 s vs 1636.80 s.
- For backdoor trigger localization, 2D-OOB averages detection AUC 0.95 across datasets and attack types versus 0.83 for 2D-KNN.
- Marginalized 2D-OOB-data attains mislabeled-detection AUCPR average 0.71, close to Data-OOB at 0.73 and well above other data valuation baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use 2D-OOB instead of 2D-KNN for cell-level outlier detection | Joint valuation · Cell-level outlier detection (12 binary datasets) | AUC: 0.67 -> 0.83 | Helped |
| Use 2D-OOB instead of 2D-KNN for cell-level outlier detection | Joint valuation · Cell-level outlier detection (12 binary datasets) | Run-time (seconds): 1636.80 -> 6.78 | Helped |
| Use 2D-OOB instead of 2D-KNN to localize backdoor triggers | Joint valuation · Backdoor trigger detection | Detection AUC: 0.83 -> 0.95 | Helped |
| Use logistic regression weak learners instead of decision tree weak learners (ablation) | 2D-OOB · Cell-level outlier detection (12 binary datasets) | AUC: 0.83 -> 0.80 | Hurt |
| Use single-layer MLP weak learners instead of decision tree weak learners (ablation) | 2D-OOB · Cell-level outlier detection (12 binary datasets) | AUC: 0.83 -> 0.82 | Hurt |
| Use two-layer MLP weak learners instead of decision tree weak learners (ablation) | 2D-OOB · Cell-level outlier detection (12 binary datasets) | AUC: 0.83 -> 0.83 | No clear effect |
| Increase the number of base learners from 500 to 1000 (ablation) | 2D-OOB · Cell-level outlier detection (12 binary datasets) | AUC: 0.81 -> 0.83 | Helped |
| Increase the number of base learners from 1000 to 3000 (ablation) | 2D-OOB · Cell-level outlier detection (12 binary datasets) | AUC: 0.83 -> 0.83 | No clear effect |
| Use feature subset ratio K/d = 0.25 instead of 0.50 (ablation) | 2D-OOB · Cell-level outlier detection (12 binary datasets) | AUC: 0.83 -> 0.81 | Hurt |
| Use feature subset ratio K/d = 0.75 instead of 0.50 (ablation) | 2D-OOB · Cell-level outlier detection (12 binary datasets) | AUC: 0.83 -> 0.82 | Hurt |
| Use 2D-OOB-data instead of 2D-KNN-data for marginalized point valuation | Marginalized joint valuation · Point-level mislabeled data detection (12 datasets) | AUCPR: 0.34 -> 0.71 | Helped |
Surprising findings
- Two-stage attribution (Data-OOB followed by TreeSHAP) performs much worse than joint 2D-OOB (average AUC 0.72 vs 0.83).
Limitations
- Evaluated mainly with random forest / decision tree weak learners on tabular data and simple image data; neural networks not explored.
- Joint valuation quality depends on the choice of weak learner; some learners perform slightly worse.
- Image experiments require super-pixel reduction to 256-dimensional vectors to make baselines tractable.
Experimental setup
Datasets: OpenML tabular binary datasets (12, e.g., lawschool, electricity, musk) CIFAR-10 class pairs with BadNets and Trojan triggers UCI multi-class datasets (Covertype, Dry Bean, Wine Quality)
Benchmarks: OpenML 12 binary tabular datasets CIFAR-10 backdoor pairs UCI multi-class datasets
Metrics: AUC run-time (seconds) detection rate test accuracy under cell fixation AUCPR for mislabeled detection
Baselines: 2D-KNN Data-OOB DataShapley KNNShapley DataBanzhaf LAVA two-stage attribution (TreeSHAP)
Builds on: Data-OOB 2D-Shapley DataShapley subset bagging
Closest papers by meaning
- Robust Conformal Outlier Detection under Contaminated Reference Data (ICML 2025 · similarity 0.82)
- ALTBI: Constructing Improved Outlier Detection Models via Optimization of Inlier-Memorization Effect (AAAI 2025 · similarity 0.82)
- DOS: Diverse Outlier Sampling for Out-of-Distribution Detection (ICLR 2024 · similarity 0.82)
- Beyond Clean Training Data: A Versatile and Model-Agnostic Framework for Out-of-Distribution Detection with Contaminated Training Data (CVPR 2025 · similarity 0.82)
- Distributionally Robust Data Valuation (ICML 2024 · similarity 0.82)
- ConjNorm: Tractable Density Estimation for Out-of-Distribution Detection (ICLR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.