HOW NOT TO OPERATIONALIZE A MACHINE LEARNING MODEL

Organizations today are generating and storing tremendous amounts of complex data. Data scientists are on the forefront of this landscape, responsible for improving business KPIs by operationalizing AI-based predictions and recommendations. However, few successfully deploy such models. And even fewer succeed in creating business impact.

In order to succeed you need to cross 3 chasms: the first is Business Relevance - avoiding model development before thinking the business value through. A clear vision of the desired business impact must shape the approach to data sourcing and model building.

The second is Operationalizing Models - migrating a predictive model from a research environment into production. This process can be a difficult because data scientists are typically not IT solution experts and vice versa.

The third, and most critical chasm is Translating predictions to business impact - where a data scientist ensures the decision makers understand the predictions and have enough wiggle room to take action and turn it into a competitive advantage. Management must possess the muscle to transform the organization so that the data and models actually yield better decisions. Additionally, model outputs need to be integrated into well-designed applications making them easy to consume.

In this talk, I will explain these three elements using a real-world use case. A European shipping company was looking to gain a competitive advantage by leveraging Machine Learning techniques. The aim was to create shipping-lane specific demand forecasting, and to implement it throughout its operations, in order to: save time and manual labor, adjust pricing and business agreements, and utilize smart resource allocation. Each percentage of environment is worth $1.5 million.

I will highlight common mistakes to avoid when operationalizing a Machine Learning model in an enterprise environment. Finally, I will demonstrate how we tackle such mistakes in this particular case.

This talk is ideal for Data Scientists, Product Managers, Development Managers and other business stakeholders that work with Data Scientists.