Venturing into Machine Learning – Start small and keep it real
The world of Advanced Analytics, Artificial Intelligence and Machine Learning is amazing, but also very intimidating. The big boys (Google, Facebook etc.) are doing some very clever, yet potentially unsettling things. However, I think bridging the gap between this hype and real-world applications in your business is hard!
Conversations with our clients are fascinating at the moment as they require a good understanding of both the business and the art of the possible. Typically, they are excited to get going, don’t really know where to start, and then end up boiling down into one of a few real world scenarios…:
- Using anomaly detection techniques to monitor thousands of operational data feeds, KPIs or machine level outputs in real time. Many systems will use crude threshold alarms, but it is possible to be much better with models that learn ‘normal’ and then look for anomalies in real time.
- Extending scenario one into process optimisation. What are the factors that drive the process outcome (customer satisfaction, sales or yield)? Can these be better managed to improve performance?
Reduce your dependency on manual checking by automating repetitive checks (e.g. invoice validation)
- Segment your customers based on their actual interactions with you. Use existing information to better target customers who are likely to respond to certain offers and progress to a sale. Stop wasting marketing budget on ‘no hopers’ and increase conversion rates.
- Create an evidence-based simulation of key aspects of your operation. Achieve unprecedented levels of accuracy by incorporating thousands of data points – creating a level of sophistication that was previously impossible. Use this model to rapidly test improvements to your business and improve your forecasting accuracy.
- Unlock the full potential of your social media channel. What are your customers (or employees) saying about you? Who or what is driving their behaviour? Get on the front foot by leading the conversation, not reacting to it.
- Implement low(er) frequency failure identification. Capture the data leading up to a failure (e.g. machine breaking down). Compare the current MI with this historical information and score the possibility of failure.
Aim for the stars but start with getting off the ground…
I’m hearing about so many data projects that promise the earth, inflate expectations (and costs) while delivering little.
We’re trying to focus on solving Data problems the way we do in the real world (and the way we’ve always done):
- Spend time defining the problem (and then think about it some more). What are the questions that you are looking to answer? Is this the real question or do you need to think further upstream? ‘Shift Left’ has been a mantra in process improvement for years – we should learn the lessons of the past.
- Understand what data do you need to solve the problem? It is fashionable to say there is no such thing as unstructured data – true, but it is an awful lot harder if your data is spread across numerous spreadsheets, paper records, or worse, in the heads of those around the business.
- Use the right tool for the job. If you are holding a hammer, it is tempting to think of everything as a nail. Machine learning techniques may well be the best option, but it is also possible some more traditional analysis may provide a more meaningful answer. Few of us are gifted to ask exactly the right question – work with the question owner to steer the analysis.
- Don’t even think of doing any of this unless you know how you are going to use the output – especially if you intend to integrate the outputs in the operational business. What is the real-world change required, what processes need to be modified, what behaviours need reinforcing?
It may be possible to crunch all the data in the world, but it is much harder to deliver on the last points if things are too complex.
Start small, build insight, make real change and add value – Repeat!
Posted in Latest News