Source-linked AI summary
Automated Data Slicing for Model Validation:A Big data - AI Integration Approach
Yeounoh Chung, Tim Kraska, Neoklis Polyzotis, Ki Hyun Tae, Steven Euijong Whang
TL;DR
Machine-learning practitioners need to find validation subsets where a model performs poorly because overall metrics can miss failures on smaller slices. Slice Finder searches for large, interpretable slices using statistical tests and interactive visualization. The paper reports that its approaches identify problematic slices effectively and can achieve 0.88 relative accuracy at a 1/128 sample fraction, about two orders of magnitude faster.
Problem
Overall model performance can hide failures on smaller validation slices, while clustering and manual approaches may not produce large, interpretable problematic subsets.
Method
Slice Finder combines automated data slicing, hypothesis testing with effect-size checks, false discovery control, and interactive visualization to find large interpretable problematic slices.
Results
LS and DT outperform clustering on real datasets for average slice size and effect size, while both retain 0.88 relative accuracy at a 1/128 sample fraction, about two orders of magnitude faster.
Takeaways & Limitations
Slice Finder supports model validation and applications such as fairness and fraud detection where human interpretation of problematic behavior is important.
Takeaways & Limitations
Slice Finder can still search exponentially many slices on large high-dimensional datasets, and parallelization adds memory and communication overhead.
Abstract
from arXiv · showhide
As machine learning systems become democratized, it becomes increasingly important to help users easily debug their models. However, current data tools are still primitive when it comes to helping users trace model performance problems all the way to the data. We focus on the particular problem of slicing data to identify subsets of the validation data where the model performs poorly. This is an important problem in model validation because the overall model performance can fail to reflect that of the smaller subsets, and slicing allows users to analyze the model performance on a more granular-level. Unlike general techniques (e.g., clustering) that can find arbitrary slices, our goal is to find interpretable slices (which are easier to take action compared to arbitrary subsets) that are problematic and large. We propose Slice Finder, which is an interactive framework for identifying such slices using statistical techniques. Applications include diagnosing model fairness and fraud detection, where identifying slices that are interpretable to humans is crucial. This research is part of a larger trend of Big data and Artificial Intelligence (AI) integration and opens many opportunities for new research.
1 INTRODUCTION
Slice Finder addresses the difficulty of detecting model failures in smaller validation subsets by automatically finding large, interpretable, statistically supported problematic slices. It applies data-management techniques to model validation and supports applications including fairness and fraud detection.
- Motivation: Overall model metrics can hide poor performance on smaller slices, motivating more granular validation than domain-expert-defined subsets alone.Practitioners may not know all important underperforming slices in advance.
- Approach: Slice Finder efficiently discovers large, interpretable, possibly overlapping slices whose model-performance differences are unlikely to be chance findings.It treats candidate slices as hypotheses and addresses false positives from many statistical tests.
- Design goals: A useful slicing technique must balance interpretability, slice size, meaningful performance differences, and protection against false discoveries across an exponentially large search space.These requirements make exhaustive manual inspection infeasible.
- Applications: Interpretable slices can support model-fairness analysis and applications such as fraud detection by exposing underperformance on human-meaningful subpopulations.The paper also discusses broader data-slicing uses based on scoring functions.
- Contributions: The paper applies data-management techniques to model validation and provides automated slicing approaches, fairness use cases, and evaluations on real and synthetic datasets.The contributions include hypothesis testing, false discovery control, a clustering baseline, and extensive experiments.
2 DATA SLICING PROBLEM
The paper formalizes data slicing as finding feature-defined subsets where a model has significantly higher loss than its complement while balancing effect size, size, and interpretability. Slice Finder uses statistical testing and ranking criteria to select top problematic slices, with extensions for model comparison and interactive search.
- Slice definition: A slice is a conjunction of feature-value predicates, with fewer literals making it more interpretable; numeric features can be discretized into ranges.Slices may use operators such as equality, inequality, and numeric comparisons.
- Problem formulation: Problematic slices have higher loss than their complements and must balance statistical significance, effect magnitude, sufficient size, and a concise feature-based description.The framework avoids both overly broad slices with weak differences and overly specific slices with little impact.
- Model comparison: The same framework can compare two models by defining slice loss as the second model’s loss minus the first model’s loss.The paper focuses its remaining development on validating a single model.
- Statistical testing: Slice Finder treats each candidate slice as a hypothesis, using Welch’s t-test for unequal variances or sample sizes and effect size to measure practical magnitude.Statistical significance establishes that a difference exists, while effect size measures how large it is.
- Slice selection: The system ranks top-k slices by increasing literal count, decreasing slice size, and decreasing effect size, subject to effect-size and significance thresholds and minimality constraints.Returned slices may overlap, and users can vary the effect-size threshold to explore different degrees of problematicness.
3 SYSTEM ARCHITECTURE
Slice Finder combines automated slice discovery with interactive visualization, offering clustering, decision-tree, and lattice-search approaches for finding interpretable problematic slices. The lattice method searches overlapping feature combinations using breadth-first expansion, statistical filtering, and false-discovery control, while scalability remains a concern for high-dimensional data.
- System architecture: Slice Finder integrates automated data slicing with interactive visualization and supports clustering, decision-tree, and lattice-search approaches.The system processes validation data, identifies problematic slices, and lets users explore them interactively.
- Clustering: Clustering provides an intuitive baseline but produces high-dimensional, high-variance groups that are difficult to interpret and requires choosing the number of clusters.The cluster-count choice also affects slice metrics and sizes, making the parameter difficult to tune for large problematic slices.
- Decision-tree searching: The decision-tree approach partitions examples into interpretable leaves and searches breadth-first for significant, large-effect slices, but may miss problematic slices outside its learned partition.Its efficiency improves when only a few tree levels are needed, while optimization for classification can prevent exhaustive coverage.
- Lattice searching: Lattice searching discretizes numeric features, limits high-cardinality categorical values, and traverses overlapping feature combinations breadth-first to identify top-k slices.Candidate slices are prioritized by literal count, slice size, and effect size, then filtered using effect-size thresholds and significance testing.
- Lattice searching: The lattice algorithm’s returned slices satisfy the paper’s definition of problematic slices through effect-size filtering and statistical significance testing.The algorithm accepts a lattice root, slice limit, effect-size threshold, and significance level, and returns problematic slices.
- Scalability and statistical control: Slice Finder parallelizes effect-size evaluation across literal levels, but searching can still require an exponential number of slices in high-dimensional datasets.The paper identifies effect-size computation as the performance bottleneck and uses asynchronous workers to populate the hypothesis-testing queue.
4 USING Slice Finder FOR MODEL FAIRNESS
Slice Finder frames fairness analysis as identifying interpretable slices where model performance is relatively worse, without requiring sensitive features to be specified in advance.
- Sensitive applications require checking whether models perform equally well across groups, but performance can vary because of data, coverage, or training limitations.
- Fairness is difficult to formalize because definitions vary by application, no standard is widely accepted, and some definitions conflict.
- Slice Finder can identify interpretable slices with fairness issues without requiring sensitive features to be specified in advance.
- The framework can support equalized-odds analysis and any fairness metric expressible as a scoring function, with further fairness analysis performed afterward.
- Under this fairness definition, problematic slices have higher loss, corresponding to lower accuracy for log loss, than their counterparts.
5 EXPERIMENTS
The experiments compare lattice search, decision trees, and clustering across Census Income, Credit Card Fraud, and synthetic data to assess slice quality, interpretability, efficiency, and false-discovery control.
- The experiments examine accuracy, efficiency, technique trade-offs, effect-size thresholds, interpretability, and false-discovery control.
- The evaluation compares lattice search, decision trees, and k-means clustering on recommended slice quality and interpretability.
- Census Income and Credit Card Fraud use random forest classifiers, while the Census dataset has 15 features and 30K examples and the balanced fraud dataset has 984 transactions.
- Synthetic data provides additional insight into Slice Finder’s operations, while real datasets compare behavior under different data characteristics.
5.2 Problematic Slice Identification
Slice Finder is evaluated on synthetic and real data for recovering artificially introduced problematic slices, with lattice search outperforming decision trees and clustering in synthetic accuracy comparisons.
- The evaluation adds problematic slices by randomly flipping labels with 50% probability and measures recovery of those newly introduced slices.
- The experiments compare lattice search, decision trees, and clustering across synthetic and real datasets, beginning with synthetic data and then examining real-data behavior.
- Lattice search consistently achieves higher synthetic-data accuracy than decision trees as recommendations increase, while clustering performs much worse.Lattice search can identify overlapping problematic slices, whereas decision trees search only non-overlapping slices.
- On Census Income, lattice search and decision trees show similar comparison results, but both have lower accuracy than on synthetic data because existing problematic slices are unknown.
5.3 Large Problematic Slices
On real datasets, lattice search and decision trees find smaller slices with effect sizes above T = 0.4, while clustering produces large clusters with near-zero or negative average effect sizes.
- Lattice search and decision trees outperform clustering on average slice size and average effect size in the real-dataset comparisons.
- Clustering starts with large clusters whose average effect size is around 0.0 and sometimes negative, whereas lattice search and decision trees find smaller slices above T = 0.4.
- On Census Income, lattice search and decision trees find up to k = 10 problematic slices with similar effect sizes, while lattice search finds larger slices because it considers overlaps.
- On Credit Card Fraud, decision trees find no more than 7 problematic slices because deeper leaves become too small to split, requiring more tree levels for additional recommendations.Decision trees may search one level faster than lattice search but may need more levels to produce the same number of recommendations.
5.4 Adjusting Effect Size Threshold T
Adjusting the effect-size threshold changes the size and strength of recommended slices, while sampling and parallelization provide scalability with high relative accuracy. Runtime comparisons between LS and DT depend on the number of recommendations requested.
- Threshold effects: As T increases, LS searches smaller, higher-effect-size slices; on Census Income, it surpasses DT in effect size once T reaches at least 0.4.
- Threshold effects: For Credit Card Fraud, increasing T causes DT’s average slice size to drop abruptly while average effect size rises as deeper tree levels yield smaller problematic slices.
- Scalability: At sample fraction 1/128, LS and DT retain relative accuracy of 0.88 while runtime increases almost linearly with sample size, enabling about two orders of magnitude faster discovery.
- Scalability: Adding parallel workers improves LS runtime on the full Census Income data, but marginal gains decrease as more workers are added; DT lacks parallel-training support.
- Scalability: DT is faster for fewer than 5 recommendations, LS is relatively faster at intermediate k, and DT becomes faster again when k reaches at least 70.
5.6 Interpretability
Slice Finder produces human-interpretable problematic slices using a few common features, supporting direct diagnosis of model behavior across datasets and methods.
- Interpretability: Without Slice Finder, users may need to inspect misclassified examples or clusters manually to determine whether the model is biased or failing.
- Interpretability: LS identifies Census Income slices with a few common features, including married spouses and high capital gains, that reveal where the model can be improved.
- Interpretability: DT’s Census Income slices remain straightforward to interpret, although their greater number of literals makes interpretation more tedious.
- Interpretability: Credit Card Fraud slices are harder but still reasonably interpretable because anonymized feature names such as V14 obscure their meaning.
5.7 False Discovery Control
False-discovery control is needed because small samples can make non-problematic slices appear problematic. The evaluation compares Bonferroni, Benjamini–Hochberg, and α-investing in terms of false discovery rate, power, and interactive use.
- False Discovery Control: Small samples can make many slices appear problematic by chance, motivating false-discovery control when selecting large problematic slices.
- False Discovery Control: As α or AI wealth increases up to 0.01, AI and BH achieve higher false discovery rate alongside higher power than the conservative alternatives.
- False Discovery Control: α-investing slightly outperforms BH and Bonferroni in slice accuracy by allocating α through the Best-foot-forward policy.
- False Discovery Control: α-investing is the only evaluated technique described as supporting an interactive setting, whereas Bonferroni is conservative and BH ignores slice-order information.
6 RELATED WORK
Related tools support manual slicing, bias discovery, model analysis, or broader data exploration, whereas Slice Finder automatically searches for interpretable model-underperformance slices.
- Model validation tools: Existing validation tools such as Facets, TFMA, and MLCube support bias discovery or granular analysis but rely on manual slice exploration or predefined feature dimensions.
- Model validation tools: Slice Finder complements these tools by automatically finding slices useful for model validation.
- Data exploration: OLAP methods optimize interesting or broad data rules, whereas Slice Finder targets slices where the model underperforms without evaluating every possible slice.
- Model understanding: PALM identifies influential training examples using an interpretable meta-model, while Slice Finder addresses problematic validation-data slices.
7 CONCLUSION
Slice Finder is presented as an efficient tool for finding large, significant, interpretable slices, with complementary search approaches and an interactive visualization front-end. Future work targets improved numeric-feature discretization, slice merging and summarization, production deployment, and user studies.
- Slice Finder efficiently and accurately finds large, significant, interpretable slices using decision trees for non-overlapping slices and lattice search for possibly overlapping slices.
- An interactive visualization front-end helps users quickly browse through a handful of problematic slices.
- Future work includes better numeric-feature discretization, slice merging and summarization, production deployment, and a user study of explanatory and debugging value.