90% of corporate strategies explicitly mention Data and AI algorithms as differentiators to extend the company’s value proposition in the market.

Why do only 20% of the analytical insights have a direct impact on the business?

Why do only 13% of Data Science projects reach production?

Many reasons contribute to the failure of Machine learning projects, from initial data issues to inadequate monitoring.

Other reasons we have found for failure include :

1. The translation problem.  Business objectives and user needs are often mistranslated into data science metrics. 

A clear difference exists between the language of business and the language of data science.

2. Is AI really needed to solve the business problem? Which user problems is AI uniquely positioned to solve? 

Artificial intelligence isn’t always the answer. Business problems can be solved without AI. 

3. Solving the business problem requires data availability.

Lack of variety and not having the right kind of data can affect the performance of the Machine Learning model.

4. Siloed data teams.

5. In some business cases, ML production is not mandatory, but it should be.

Often, the goal of analytics projects is to build just models rather than deliver value.

6. Thinking too much about the ML model rather than the business value it provides

 

To reduce Machine Learning failures, here’s how we approach it, the 10 steps below :

 

1. Data audit. Which are the main schemas, features and labels that are defining the main data sources who model the business.

2. Understanding business problems and customer needs.

What’s the main business problem and user need we are addressing?

3. Identifying potential use cases to approach the business problem by analyzing what ML algorithm might help solve the problem.

4. Match customers needs with data needs. Understanding how data available might solve the business problem.

5. Prioritization. Which use case should we implement first? 

Evaluate algorithm complexity vs customer benefits.

6. Prototyping AI model with ML no-code or exploratory data analysis tools before implementation.

This step increases the data interpretability of the process and reduces the translation problem mentioned previously.

If we are able to model, prototype, the solution with the current data ecosystem and data available then we can start the implementation phase. If not, we need to define another business case.

7. First, Build a simple model and iterate. Machine Learning success is an iterative process.

8. Product definition based on the model’s output.

What is the ML’s output? How´s the output (UI/UX) presented?

9.Success criteria. 

How we qualify success in an offline / online environment. 

10. Data actionability. What actionable next steps can we generate based on model´s output.

Avatar
Posted by:Fran Castillo

@francastillo