TLDR: The paper introduces an updated Predictability-Computability-Stability (PCS) framework for “veridical” (truthful) data science, addressing the challenge of irreproducibility in AI-driven findings. It provides a streamlined workflow for practitioners, integrating the use of generative AI while emphasizing reality checks, computational efficiency, and stability against judgment calls throughout the data science life cycle. A case study highlights how data cleaning choices significantly impact prediction uncertainty, underscoring the framework’s importance for trustworthy data science.
In today’s world, data science and artificial intelligence (AI) are transforming nearly every aspect of our lives. From scientific discoveries to medical advancements, data-driven insights are incredibly powerful. However, a significant challenge remains: many findings are difficult or even impossible to reproduce. This issue often stems from the many choices, or ‘judgment calls,’ made throughout the data science process, which introduce uncertainty that traditional methods often overlook.
The Challenge of Trustworthy Data Science
Data analysis, while ancient in concept, has evolved dramatically with technology. Today, software automates complex calculations, and with the rise of generative AI (GenAI) systems like ChatGPT and Gemini, human-machine collaboration in data science is entering a new era. Yet, this evolution also brings new complexities. The data science life cycle (DSLC), a structured process for investigation, involves numerous human judgment calls. These decisions, often subjective, can lead to different analytical paths and outcomes, making results hard to replicate. This lack of reproducibility is a major concern, highlighted by studies showing that a large percentage of scientific discoveries cannot be replicated.
Introducing the PCS Framework
To tackle this challenge, the Predictability-Computability-Stability (PCS) framework for veridical (truthful) data science has been developed. PCS offers a principled way to address uncertainty throughout the DSLC, aiming to make data-driven conclusions more trustworthy and reproducible. This framework integrates best practices from both machine learning and statistics, fostering a unified approach to ensure the safety and reliability of data science and AI.
PCS Principles: Predictability, Computability, and Stability
The PCS framework is built on three core principles:
- Predictability (P): This acts as a ‘reality check,’ ensuring that the data science process accurately captures reality and aligns with the problem’s goals. In supervised learning, this means checking predictions against unseen data. For other tasks, it involves using domain knowledge or cross-validation to verify discovered patterns.
- Computability (C): Beyond just speed and memory, computability in PCS includes using data-inspired simulations to explore different scenarios and understand the computational aspects of the analysis.
- Stability (S): This crucial principle ensures that data-driven results remain consistent even when reasonable changes or ‘perturbations’ are made to the data science process. These perturbations could involve using different datasets, cleaning procedures, or modeling algorithms. If results are unstable, the process needs to be improved. Stability checks are documented rigorously, providing a deeper understanding of uncertainty than traditional statistical methods.
Navigating the Data Science Life Cycle with PCS
The DSLC consists of six iterative stages, and PCS principles are applied throughout each one:
Stage 1: Problem Formulation and Data Collection
This initial stage involves working with experts to define a clear problem and gather sufficient, well-documented data. PCS emphasizes ensuring the data is suitable for the task and that its collection process is understood to avoid downstream issues. Generative AI can assist here by quickly providing high-level understanding of unfamiliar subjects, but its output must always be cross-referenced and verified with trusted external sources.
Stage 2: Data Cleaning and Exploratory Data Analysis
Often underestimated, data cleaning is critical. PCS advocates for an iterative approach, where insights from exploratory data analysis (EDA) guide cleaning decisions. It’s recommended to create multiple versions of cleaned data by varying judgment calls (e.g., how missing values are handled). A case study in the paper demonstrates how different data cleaning choices made by students led to significant variations in predictive model performance, highlighting the substantial uncertainty introduced at this stage. GenAI can help generate starter code for cleaning and documentation, but human verification is essential.
Stage 3: Uncovering Data Structure
This optional stage uses unsupervised learning techniques like dimensionality reduction or clustering to find hidden patterns. PCS ensures the validity of these findings through stability and predictability checks. For instance, clustering results should be consistent across different algorithms or data samples. While GenAI can assist in generating code for these complex techniques, understanding their mechanics is crucial for verifying the generated code.
Stage 4: Predictive Modeling
PCS focuses on ensuring models are aligned with reality and stable during development. This involves pre-selecting stability evaluation metrics and using data perturbations to test model sensitivity. Instead of relying on a single ‘optimal’ model, PCS encourages using multiple models (different algorithms, hyperparameters, or cleaned datasets) to better understand prediction uncertainty. These models can then be combined into a ‘PCS ensemble’ to create more stable predictions and provide prediction perturbation intervals (PPIs) for uncertainty quantification. GenAI can help with parameter tuning and model selection, but the entire modeling pipeline should not be delegated to it, and random seeds should be used for reproducibility.
Stage 5: Evaluation of Results
Scrutinizing findings is an ongoing process. PCS suggests measures to avoid confirmation bias, such as comparing results to ‘dummy’ labels or showing ‘fake’ results alongside real ones to domain experts. It also emphasizes checking result consistency across different visualizations and metrics. GenAI can assist in creating comparison plots, but analysts and domain experts must draw the conclusions.
Stage 6: Communication of Results
The final step is sharing findings effectively with the target audience, aiming for real-world impact. This means tailoring presentations, creating user-friendly software, or publishing accessible code. The paper highlights the importance of making code publicly available and easy to use, noting that the PECARN team, while publishing their findings and data, did not release code for data cleaning or modeling. GenAI can draft documentation for code and data repositories and help create simple web applications for demonstrations, but its output must always be verified.
Generative AI and PCS: A Powerful Partnership
The paper provides the first formal guidance on incorporating GenAI into the DSLC. While GenAI offers powerful assistance, its tendency to ‘hallucinate’ (present plausible but false information) necessitates adherence to PCS principles. This includes cross-referencing multiple GenAI systems, perturbing prompts to check consistency, and always verifying GenAI output with trusted external sources. The general spirit of these recommendations, focused on reality and stability checks, will remain relevant even as GenAI capabilities rapidly advance.
Also Read:
- Language Models Reshape Tabular Data Preparation Workflows
- A New Definition for Interpretable AI: Bridging the Gap Between Models and Human Understanding
Why PCS Matters
The PCS workflow provides a practical framework for creating reproducible and trustworthy results by systematically addressing the uncertainty introduced by judgment calls. By integrating reality checks, computational efficiency, and stability quantification, PCS enables more reliable data science across various domains, from social sciences to biomedical research. For those interested in learning more, the full research paper can be found at this link, and additional resources are available through the Yu Group website and GitHub.
