InteractEva: A Simulation-Based Evaluation Framework for Interactive AI Systems

Yannis Katsis, Maeda F. Hanafi, Martín Santillán Cooper, Yunyao LiAAAI 2022 · DOI

Introduction

Evaluating Interactive AI (IAI) systems presents a unique challenge compared to traditional machine learning models. In standard AI, performance is measured against static datasets. However, IAI systems evolve through real-time feedback from Subject Matter Experts (SMEs), creating a "moving target" problem. The system's effectiveness is intrinsically tied to the specific sequence and nature of user interactions. Traditionally, developers have relied on manual user testing, which is resource-intensive, difficult to replicate, and often yields only qualitative insights.

The InteractEva Framework

Presented at AAAI-22, InteractEva is a simulation-based framework designed to automate the evaluation of IAI systems. By mimicking diverse user behaviors at scale, it enables systematic, quantitative debugging and optimization of human-in-the-loop systems.

The architecture of the InteractEva framework, showing the interaction between the simulation backend, the evaluated IAI system, and the analysis frontend.

The framework is divided into two primary components: a simulation backend and an interactive frontend.

1. Simulation Backend

The core of the framework is the User Simulation Engine, which replaces manual human interaction with automated agents.

  • User Models: These interchangeable modules simulate various user personas, ranging from highly cooperative users to those providing minimal feedback. This allows developers to test system robustness across a spectrum of human behaviors.
  • Use Cases: A library containing datasets and ground truth information representing specific real-world tasks.
  • Simulation Loop: The engine iteratively queries a user model, submits the resulting action to the IAI system, and retrieves the updated model's predictions. This process is repeated $k$ times to generate a statistically significant set of "runs."

2. Interactive Frontend

Simulation results are exported as JSON reports and visualized in a dedicated UI designed for three key analytical tasks:

  • Performance Analysis: The UI displays the min-max range of performance metrics (e.g., Precision, Recall, F1), helping developers identify "long-tail" edge cases where the system might fail under specific conditions.
  • Error Analysis: Developers can "drill down" into individual simulation runs to inspect the model's internal state and logic at each step of the interaction.
  • What-if Analysis: The framework facilitates side-by-side comparisons of different AI backend versions or user models, serving as a powerful tool for regression testing.

Case Study: Pattern Induction

The framework was validated using Pattern Induction, an industrial IAI system within the IBM Watson Discovery suite used for Natural Language Processing (NLP) extraction. By integrating InteractEva into the development cycle, the team was able to identify backend issues that were previously difficult to detect through manual testing. It provided a systematic way to track performance improvements and ensured that new features did not introduce regressions in existing interaction paths.

Insights and Future Directions

A key insight from the development of InteractEva is the importance of the "plurality of users." Simulating an "average" user is insufficient; robustness requires testing against a variety of interaction styles. While the framework currently focuses on algorithmic robustness rather than user interface (UI) intuitiveness, it successfully transforms "vibe-based" feedback into hard, actionable metrics.

Future work involves expanding the library of user models to increase simulation fidelity and applying the framework to other domains, such as interactive image labeling and recommendation systems.

Summary updated

BibTeX
@inproceedings{katsis2022interacteva,
  author = {Yannis Katsis and Maeda F. Hanafi and Martín Santillán Cooper and Yunyao Li},
  title = {{InteractEva: A Simulation-Based Evaluation Framework for Interactive AI Systems}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21721},
}

Loading related content...

Figure