RATE-Analytics: Next Generation Predictive Analytics for Data-Driven Banking and Insurance

Dennis Collaris, Mykola Pechenizkiy, Jarke J. van WijkCommit2Data 2024 · OpenReview

TL;DR

This paper introduces the RATE-Analytics project, a collaboration focused on building trustworthy and explainable AI for banking and insurance. They tackle the critical issue of low adoption of ML models due to a lack of trust by developing interactive visualization tools that bring human experts into the loop for model diagnostics, refinement, and justification.

What Problem Does This Paper Address?

In the banking and insurance sectors, vast amounts of customer data are available, offering immense potential for developing new products and improving services. However, the adoption of advanced machine learning (ML) models is hampered by a fundamental problem: a lack of reliability and trustworthiness. When domain experts don't trust a model or its predictions, they are unlikely to use it, regardless of its technical performance. This paper addresses this gap by focusing on making ML models more transparent, explainable, and ultimately, trustworthy, enabling better human-AI collaboration.

Methods

The core of the RATE-Analytics project's technical contribution lies in developing interactive visualization systems designed to facilitate human-in-the-loop ML. They propose three main visual analytics approaches:

  • EXPLAINEXPLORE: This system provides local explanations for individual predictions. It offers context by showing similar predictions and the impact of small input perturbations. This is particularly useful for decision support and justification.
  • Contribution-Value Plots (CV Plots): These plots offer an intermediate perspective, showing how feature importance changes across different feature values. This helps diagnose unexpected model behavior by visualizing model behavior on a per-feature basis.
  • STRATEGYATLAS: This approach enables a global understanding of complex ML models by identifying and interpreting "model strategies." These strategies are visualized in a "StrategyMap," allowing data scientists to analyze feature values and contributions to understand different model behaviors for subgroups.

Complementary to these visualization tools, they also developed a software library for interoperability and an algorithmic contribution (LEMON) to improve the faithfulness and speed of explanation computations.

Datasets

The paper mentions that real-world data from the project partners, Achmea and Rabobank, were used. Specific datasets are not detailed, but they are described as containing a wealth of information about customer behavior and events relevant to banking and insurance use cases like fraud detection, loan default prediction, and insurance claim prediction.

Key Insights and Ideas

  • Trust is paramount: The paper strongly emphasizes that technical accuracy alone is insufficient. If domain experts don't trust a model, its practical utility plummets.
  • Human-in-the-loop is key: Integrating human experts into the ML lifecycle is crucial for building trust, enabling model refinement, and supporting decision-making.
  • No silver bullet: The project learned that different problems and perspectives require different explanation approaches, ranging from local to global.
  • Visualizations bridge the gap: Interactive visual analytics tools can effectively translate complex ML model behavior into understandable insights for domain experts and data scientists.
  • Beyond prediction scores: For high-stakes decisions, simply relying on a prediction score is inadequate. Explanations are needed to verify fairness, privacy, and trust.

Key Results

The project developed and evaluated interactive visualization systems (EXPLAINEXPLORE, CV Plots, STRATEGYATLAS) that demonstrated practical utility.

  • EXPLAINEXPLORE was used by data scientists in Achmea's debtor management department to diagnose models, identify areas for improvement, and support decision-making.
  • STRATEGYATLAS was applied to an automatic insurance acceptance model at Achmea, helping data scientists understand and refine the model by revealing unexpected model behaviors.
  • A quantitative survey showed that CV plots increase understanding of model behavior.
  • The project also explored techniques for handling non-IID data and anomaly detection in financial networks, demonstrating effectiveness on both synthetic and real-world datasets.

Limitations and Future Work

The authors acknowledge that there is no single solution for all interpretability problems. They also note that the research contributions were evaluated in practice, achieving a Technology Readiness Level (TRL) of 5.

Future work includes:

  • Exploring the application of their generic technology in other domains like medical diagnostics and predictive maintenance.
  • Investigating the feasibility of a startup company to help businesses understand and explain complex ML models, focusing on customer interviews, business model validation, and identifying initial customers.
  • Continuing to develop solutions that reduce the risk of unethical decisions and ensure compliance with regulations like GDPR and the AI Act.
  • Tailoring prototypes for different customer needs and preparing a convincing demonstration for the future startup.

Summary updated

Abstract

We conducted the RATE-Analytics project: a unique collaboration between Rabobank, Achmea, Tilburg and Eindhoven University. We aimed to develop foundations and techniques for next generation big data analytics. The main challenge of existing approaches is the lack of reliability and trustworthiness: if experts do not trust a model or its predictions they are much less likely to use and rely on that model. Hence, we focused on solutions to bring the human-in-the-loop, enabling the diagnostics and refinement of models, and support in decision making and justification. This chapter zooms in on the part of the project focused on developing explainable and trustworthy machine learning techniques.
BibTeX
@inproceedings{collaris2024rate,
  author = {Dennis Collaris and Mykola Pechenizkiy and Jarke J. van Wijk},
  title = {{RATE-Analytics: Next Generation Predictive Analytics for Data-Driven Banking and Insurance}},
  booktitle = {Commit2Data},
  year = {2024},
  url = {https://openreview.net/forum?id=kWhtvNxF1c},
}
Figure