PaperlayerView exact revision on arXiv

Source-linked AI summary

Veridical Data Science

Bin Yu, Karl Kumbier

arXiv:1901.08152v5stat.MLcs.LG

TL;DR

The paper addresses limited reliability and transparency in data science when human judgment calls shape analyses across the data science life cycle. It proposes PCS, a workflow and documentation framework that combines predictability, computability, and stability with perturbation-based inference. PCS is illustrated in scientific applications and performs favorably against other methods across simulated settings, while physical experiments remain necessary for causal validation.

  • Problem

    Limited transparency about human judgment calls throughout the data science life cycle can blur evidence and contribute to false discoveries.

  • Method

    PCS combines a life-cycle workflow and reproducible documentation with predictability checks, computability considerations, and stability analyses based on data and model perturbations.

  • Results

    PCS compares favorably to other methods across simulated settings, and 80% of second-order interactions identified by iRF had been verified in the literature through physical experiments.

  • Takeaways & Limitations

    PCS supports reproducible, transparent assessment of how human decisions affect data results and can recommend hypotheses for downstream scientific investigation.

  • Takeaways & Limitations

    Stability assumptions require alternative decisions to be appropriate, meaning well-justified by domain knowledge and understanding of the data-generating process.

Abstract

from arXiv · show

Building and expanding on principles of statistics, machine learning, and scientific inquiry, we propose the predictability, computability, and stability (PCS) framework for veridical data science. Our framework, comprised of both a workflow and documentation, aims to provide responsible, reliable, reproducible, and transparent results across the entire data science life cycle. The PCS workflow uses predictability as a reality check and considers the importance of computation in data collection/storage and algorithm design. It augments predictability and computability with an overarching stability principle for the data science life cycle. Stability expands on statistical uncertainty considerations to assess how human judgment calls impact data results through data and model/algorithm perturbations. Moreover, we develop inference procedures that build on PCS, namely PCS perturbation intervals and PCS hypothesis testing, to investigate the stability of data results relative to problem formulation, data cleaning, modeling decisions, and interpretations. We illustrate PCS inference through neuroscience and genomics projects of our own and others and compare it to existing methods in high dimensional, sparse linear model simulations. Over a wide range of misspecified simulation models, PCS inference demonstrates favorable performance in terms of ROC curves. Finally, we propose PCS documentation based on R Markdown or Jupyter Notebook, with publicly available, reproducible codes and narratives to back up human choices made throughout an analysis. The PCS workflow and documentation are demonstrated in a genomics case study available on Zenodo.

1. Introduction

The paper introduces veridical data science through the PCS framework, which integrates predictability, computability, and stability across the data science life cycle. It addresses limited transparency around human judgment calls by combining workflow guidance with reproducible documentation.

  • The data science life cycle moves from a domain question through data collection, management, cleaning, exploration, modeling, and interpretation.
  • Limited transparency about human judgment calls can blur evidential support and contribute to false discoveries.
  • PCS uses predictability as a reality check, computability to assess tractability, and stability to evaluate reproducibility relative to judgment calls.
  • The framework combines a PCS workflow with documentation in R Markdown or Jupyter Notebook containing narratives, code, and visualizations.

2. PCS principles in the DSLC

PCS applies predictability, computability, and stability throughout the data science life cycle, beginning with justified problem and data decisions and extending through modeling and evaluation. The framework emphasizes checking whether results remain consistent under appropriate perturbations and documenting the judgments involved.

  • The PCS workflow applies its principles at every life-cycle step, while PCS documentation records workflow decisions in a six-step format.
  • Stability asks whether alternative appropriate decisions would produce similar data conclusions and treats appropriate perturbations as domain-justified changes to data or models.
  • Problem formulation and data collection require translating domain questions into measurable analyses and considering whether collected data represent future settings.
  • B. Predictability as reality check: Predictability provides empirical validation by evaluating how well a model represents relationships using a prediction function, testing data, and an evaluation metric.
  • B. Predictability as reality check: The prediction loss function measures similarity between predictions and targets, with larger values defined as worse predictive accuracy.

B.2. Cross validation.

The framework treats computability as a condition for feasible data science and stability as a broader assessment of how results change under data and model perturbations. It also identifies weakly justified model-distribution assumptions as a major threat to reliable knowledge extraction.

  • C. Computability: Computability is the gatekeeper of data science because data must be generated, stored, managed, and analyzed efficiently and scalably.
  • C. Computability: Computing resources determine which analyses can be performed and are necessary for PCS predictability and stability analyses.
  • D. Stability at the modeling stage: Stability measures changes in data results under data or model perturbations, extending statistical sampling variability beyond a single assumed distribution.
  • D. Stability at the modeling stage: Statistical uncertainty relies on a model distribution that must be supported by domain knowledge, an understanding of data generation, and downstream utility.
  • D. Stability at the modeling stage: Evaluating stability across problem formulation, preprocessing, exploration, randomized algorithms, and model choices supports reliable and reproducible conclusions.

D.1. Formulating stability at the modeling stage.

PCS formulates modeling-stage stability by tracking a target across defined data and model perturbations and summarizing the resulting target distribution. The framework highlights that perturbation choices are subjective and that unstable targets can motivate alternative analyses or new data collection.

  • A stability target is a data result or estimand determined by input data and a specific model or algorithm.
  • PCS evaluates stability by measuring how target T changes across collections of data perturbations D and model or algorithm perturbations Λ.
  • Appropriate perturbations can include bootstrap sampling, block bootstrap, domain-supported generative models, and equally appropriate prediction functions.
  • The trustworthiness of a data result depends on how perturbations are justified because their selection involves subjective human judgment calls.
  • Stability metrics summarize target distributions using quantities such as feature-selection proportions or saliency-map ranges.
  • An unstable target may suggest an alternative analysis or target, while held-out data or new data can help address multiplicity, overfitting, reproducibility, and external validity.

D.2. Data perturbation.

Data perturbations can reduce variability in estimated targets, while model perturbations assess how alternative analyses of the same data affect target estimates.

  • Unjustified probabilistic models make sampling interpretations inapplicable.
  • Data perturbations can reduce variability in estimated targets.Random forests use subsampling, while GANs use synthetic adversarial examples to improve prediction robustness.
  • Model perturbations compare how alternative analyses of the same data affect target estimates.Examples include robust statistics and Bayesian sensitivity analysis.

D.3. Algorithm or model perturbation.

Algorithm and model perturbations use randomness and alternative models to assess or reduce instability in target estimates. Generative models additionally summarize prior data and knowledge while producing synthetic data perturbations.

  • Random components and multiple appropriate algorithms provide model perturbations for assessing target stability.Examples include random projections, randomized initialization, and competing models or algorithms.
  • Model perturbations can reduce target variability or instability.Lasso coefficients can be selected for stability across regularization parameters, while dropout improves neural-network generalizability.
  • Generative models both summarize past data and prior knowledge and generate synthetic observations for data perturbation.
  • Synthetic data can augment observations as domain-inspired regularization.The amount combined with observed data reflects the degree of belief in the generative models.
  • Computability issues can limit which predictive models and algorithms are tractable for large, high-dimensional datasets.Scalable optimization methods such as GD and SGD are used to address these constraints.

3. PCS inference through perturbation analysis

PCS inference extends uncertainty assessment beyond sampling variability by examining how data, problem translations, preprocessing, and modeling choices perturb target estimates. It uses prediction screening, perturbation distributions, and stability summaries to identify reliable conclusions.

  • Perturbation intervals quantify target-estimate variability across data cleaning, problem translations, preprocessing, and modeling decisions.They differ from confidence intervals by being explicitly connected to documented and reviewed perturbations.
  • PCS perturbation intervals: PCS inference first defines a target, perturbations, prediction functions, data split, evaluation metric, stability metric, and stability target.The choices should be documented and justified in the domain context.
  • PCS perturbation intervals: Prediction screening removes models that fail to fit the data according to a threshold τ.Thresholds may use accepted baselines, the top k models, or accuracy close to the best model.
  • PCS perturbation intervals: The surviving models are evaluated under data perturbations to form a joint target distribution.Requiring stability across all perturbations is more conservative for type I error than requiring it for any single perturbation.
  • PCS perturbation intervals: PCS reports perturbation distributions using stability metrics such as the 10th and 90th percentiles or low-dimensional projections.
  • PCS inference identifies the stable portion of accurate models and can simplify unstable predictive results.This stable portion is intended to make conclusions more reliable and easier to interpret.
  • PCS hypothesis testing: PCS hypothesis testing defines constrained perturbations representing plausible data-generating processes under a null hypothesis.These may include well-founded probabilistic models, PDE-based simulations, or selected controls, with choices documented and debated.

B.1. Formalizing PCS hypothesis testing.

PCS hypothesis testing compares observed data with data generated under constrained perturbations representing a null hypothesis. Its formalization is illustrated through neuroscience and genomics examples and sparse linear-model simulations.

  • A constrained perturbation is documented and justified using domain knowledge before comparing perturbation intervals for D0 and D.The comparison evaluates whether observed data are consistent with the hypothesis embedded in D0.
  • Neuroscience PCS inference can generate data sharing primary features with observations while remaining otherwise random.A maximum entropy model uses the number of moments as its constraint for comparing population-level findings.
  • Sparse linear-model simulations considered six data-inspired generative models and compared PCS inference with selective inference and asymptotic normality.

4. PCS documentation

PCS documentation integrates narratives, code, and analyses to record the assumptions and human judgment calls underlying each stage of the data science life cycle. This transparency helps readers assess the evidence and reliability of resulting conclusions.

  • PCS documentation combines narratives with reproducible code and analyses in an R Markdown or Jupyter Notebook.The narratives describe the domain problem and support choices about computation, preprocessing, visualization, modeling, evaluation, and stability.
  • PCS documentation connects mathematical models and algorithms with the real-world phenomena they approximate by explaining the assumptions made throughout the life cycle.This connection is illustrated in Fig. 3 as a bridge between reality and models or algorithms.
  • Narratives record assumptions and judgment calls across the data science life cycle, including data collection, cleaning, modeling, and interpretation.They should be grounded in prior knowledge and an understanding of how the data were collected.
  • The documentation gives readers information needed to make informed judgments about the evidence and process behind a data conclusion.The genomics case study is documented on Zenodo.

5. PCS recommendation system for scientific hypothesis generation

PCS inference uses predictability and stability to rank target estimates for further study and to provide recommendations for downstream human decisions. These recommendations can suggest relationships or interactions for follow-up experiments without replacing physical tests of causality.

  • PCS inference ranks target estimates using stability and predictability to guide follow-up studies and experiments.The framework is used as a recommendation system for downstream human decisions.
  • PCS inference identified candidate relationships between visual-cortex neurons and stimuli, along with higher-order biomolecular interactions potentially regulating gene expression.These examples come from applications involving iterative random forests and signed iterative random forests.
  • 80% of second-order interactions identified by iRF had been verified in the literature through physical experiments.The authors present this verification as support for using highly predictable and stable analyses to suggest higher-yield hypotheses or intervention experiments.
  • Predictability and stability do not replace physical experiments for proving or disproving causality.PCS recommendations are intended to suggest hypotheses and intervention experiments, not establish causal effects computationally.

6. Conclusion

The conclusion presents PCS as a framework combining predictability, computability, stability, workflow, and documentation across the data science life cycle. It extends stability and PCS inference while emphasizing transparent, reproducible decisions and the need for further empirical and theoretical work.

  • PCS unifies predictability, computability, and stability into a workflow and documentation framework for reliable, reproducible, and transparent data science.Prediction checks how models capture natural phenomena, computability addresses tractability, and stability examines reproducibility under perturbations.
  • PCS extends stability across the entire data science life cycle, including problem formulation, data collection, data cleaning, and exploratory data analysis.The framework also develops perturbation intervals and hypothesis testing to evaluate variability under a broad range of perturbations.
  • PCS documentation integrates narratives justifying judgment calls with reproducible code and visualizations so users can assess whether data-driven decisions are reliable.The documentation is intended to make decisions as transparent as possible.
  • The framework still requires broader practical vetting, additional case studies, and further theoretical analysis of PCS procedures and the connections among its principles.The authors identify these as open problems beyond the paper’s case studies and previous work.

8. Supporting Information: Simulation studies of PCS inference in the linear model setting

The simulation studies evaluate PCS perturbation intervals for sparse linear-model feature selection across data-inspired generative settings. PCS compares favorably with classical inference, especially under high-dimensional, heteroskedastic, and misspecified models.

  • Simulation design: PCS perturbation intervals were evaluated for sparse linear-model feature selection across data-inspired simulation settings.The simulations were designed to assess PCS inference using ROC analysis and to illustrate its generalizability to settings where data or model perturbations can be defined.
  • Simulation design: The genomic feature matrix contained 630 standardized features, with 25 randomly selected active features used to generate responses.The feature set combined 35 genomic assay measurements with second-order pairwise interactions, excluding quadratic terms.
  • Simulation design: The study considered six generative settings spanning four noise distributions and two misspecified models.The misspecified models included removing 12 active features before fitting and generating responses through active-feature interactions.
  • PCS procedure: PCS screened models by held-out prediction accuracy, retained the 10 most accurate penalty settings or used a pre-specified prediction threshold, and repeated analyses across data halves.Each retained setting was evaluated over 100 bootstrap samples, with selected features summarized by stability intervals and one stability score per feature.
  • Results: Across all settings, PCS compared favorably with asymptotic normality and selective inference in ROC analyses averaged over 100 replicates.The ROC curves assess true- and false-positive rates for selecting active features while limiting false discoveries.
  • Results: PCS was particularly advantageous when n < p, errors were heteroskedastic, or the model was misspecified, recovering more active features while distinguishing them from inactive features.The authors identify generalizability and the ability to define data or model/algorithm perturbations as principal advantages over procedures that do not typically handle multiple models easily.
Loading 1901.08152v5…