Every data-driven company has machine learning use cases. Don’t miss out on finding the best in your company – with our machine learning use case checklist.
TLDR: Check out the checklist here: www.datarevenue.com/en-resources/ml-usecase-checklist
1. Where to Find Ideas
- What are your team's 3 most important strategic goals?
Clarify what counts at this moment. This will focus your ideas and increase the likelihood that the solution you build is valuable.
Idea source A: Manual data analysis
- List all the processes that require your team to spend a lot of time on data analysis.
Machine learning is just a pattern-finding machine. It finds patterns in a dataset and then uses that knowledge to make predictions.
That's exactly what we humans usually do when we analyse data: We try to understand – to identify a pattern.
So if someone from your team analyses a particular type of data over and over again, then machine learning could save you time and energy.
- Is this data analysis highly repetitive?
Machine learning models are limited – they can only learn to do one specific thing. That means you can only automate repetitive tasks.
You will also need a lot of training examples for the algorithm – at least 200. So make sure you have enough data available.
- Do you need to scale this process to achieve one of your strategic goals?
Machine learning makes predictions cheaply and quickly. If it can’t automate your predictions completely, it will give your team considerable support and make them more effective at a particular task.
This is particularly valuable if you want to process a much higher volume of predictions in the future.
- Have all the important facts been captured in datasets?
Machines can only read digital data. If the task involves information that is not available in a dataset somewhere (things like conversations, manual inspections, etc.), then the machine will be at a disadvantage, and it will probably not be able to complete the task to your satisfaction.
>> Important: You don't need a perfect database where everything is nicely structured in one place. It's not very hard for a data scientist to combine datasets. What's important is that you already have the data somewhere.
Idea source B: Software decision-making
Contrary to public perception, machine learning is usually best used to improve existing software – not to replace a human employee.
- List all the software you use that repeatedly makes a decision or a prediction (even Excel sheets) based on a set of rules.
Just like with manual processes, this needs to be a high-volume process so you can provide the algorithm with enough examples for training.
In particular, identify places where the software is implementing certain rules – and where you suspect that these rules don't fully capture the real, underlying complexity of the connections. Machine learning can capture that complexity for you by finding thousands of rules.
Can you can answer “yes” to these 2 questions?
- Could a human expert make better decisions?
If a human being could make a better decision, then that's a good sign that the current rules don't capture the true complexity of the decision.
If that's the case, then you can interview this expert and figure out what data you need to feed the algorithm so it can try to learn to do the same thing.
- Have you saved > 500 examples of the “correct” predictions for a given situation?
The algorithm requires a lot of examples to learn these patterns. You need examples of the input factors, and especially of the correct output – the values you expect the algorithm to predict in a given situation.
If you don't have enough correct predictions, then you could ask a human expert to annotate the data with the expected output.
2. How to Prioritize Ideas
Now that you've narrowed down your list of ideas, you can compare them according to the following factors:
- Value – In a best-case scenario, how valuable would the solution be?
How much time and resources do you save with automated decision support? What else will be possible once you have automated this task and/or drastically improved its accuracy?
- Data availability – Do you already have all the data you need to feed an algorithm – right now?
This is not the data you might have in the future, but the data history you have access to now. How much is there? How relevant is it?
- Reference availability – Have other teams already built similar systems?
Is this a research project with few comparable solutions? Or have other teams – maybe in other industries – already solved a similar problem?
A good reference can reduce the risk of the project considerably.
- Ease of validation – You should be able to build a proof of concept in 6 weeks or less.
If you don’t have a clear idea of how to test your idea within 6 weeks, then be careful. Successful projects are usually simple and easy to validate.
Download the checklist
Sign up for the “AI for Managers” mailing list, and download the checklist here:
The checklist also includes a Google Docs template — so you can start filling it out right away.