Building a Decisioning Engine for Data Scientists: A Practical Guide

Track: Data Science

Type: Talk

Room: Lefthand Room (Seminar Room 3)

Time: Oct 03 (Thu): 14:45

Duration: 0:45

Machine learning offers a modern replacement for human-curated predictive decisioning. However, in practice, predictive rules manually created from expert domain knowledge remain extremely relevant in many industries, most notably finance. These rules are straightforward, highly explainable, and enable experts to incorporate valuable domain knowledge. However, the open-source ecosystem until recently, lacked comprehensive tooling for creating decision flows that can be operationalized following modern-day machine learning best practices.

In this talk, we will introduce a new open-source decisioning framework built on Hamilton, a versatile framework designed to streamline the creation and management of dataflows using standard Python functions. Hamilton simplifies the development process by converting functions into nodes in a Directed Acyclic Graph (DAG), allowing for efficient execution and visualization of decision flows. Our new open-sourced framework enables scientists to create rules around Statistical, Machine Learning, and Large-language models. We will showcase the framework's ability to address fraud detection, credit granting, and integration with LLMs. This session is tailored for decision creators (e.g., data scientists, decision scientists, and data analysts), and operations engineers.

The talk will delve into the key components of our Python-based decisioning solution applicable to finance, including decision trees, credit risk scorecards, and decision tables. We will emphasize using Directed Acyclic Graphs (DAGs) to achieve interoperability, traceability, and visual insights into decision-making processes. We will then show how these decisioning processes can be productionised for real-time and batch use cases. The talk will provide listeners with practical insights for developing decisioning flows using best practices.