A featured contribution from Leadership Perspectives: a curated forum reserved for leaders nominated by our subscribers and vetted by the Construction Tech Review Advisory Board.
Gjensidige
Lessons Learned on How to Succeed with Machine learning
In the last decade the demand for skilled data scientists have exploded. Businesses have realized that to stay competitive, they need to utilize their data with multivariate statistical methods – machine learning. The number of skilled data scientists has not been able to meet the business demand, thus being a major bottle neck in becoming more data driven. I predict this will change in the years to come, but many challenges still need focus.
The statistical processes data Scientists have been doing by building machine learning models are now quickly being automated with so called Auto-ML tools, as well as being supported by readymade generic models you could buy access to from cloud vendors (speech recognition, feature extraction from pictures and so on). The automation pro-cess is able to build faster and often more accurate models than the data scientist are able to do themselves.
In my many years working on machine learning within marketing automation, I have learned some lessons on how to succeed. Here are five key take always, and advices I can give on what I think is important to focus on in the journey to become a more data driven business.
1. Cross functional teams. The organizational part is actually of great importance for succeeding. If you are able to organize data scientists, data engineers and business oriented people working autonomous and agile in the same team, then you will be able to find better technical solutions, and easier being able to spot the most important/valuable tasks to automate. This will make you more productive with machine learning, as it is a never-ending process of test and learn. Each specialist is just a piece in the puzzle, and are dependent on the others professional competence to succeed. Cross functional teams, sharing the same goals, will over time give each team member enough understanding of the value of the other competencies. A traditional “waterfall” organization where each competence is organized in silos with their own goals and priorities not synchronized across, will complicate the process, and give less productivity.
2. Find suitable business processes to apply machine learning on. E.g. where the business-process could be enhanced by implementing automatic decisioning. This will change over time, based on how your business processes evolve, as well as how you are able to capture data for the machine learning. To enhance your business processes you will have to focus on applying more context specific models, so that the models are able to make precise decisions for specific business purposes. E.g. a model used for lead generation in a contact center does not have the same context as a personalized ad on a web page.
3. What data to gather for the machine learning process. Machine learning is fueled by data, and data is more important for the quality of the models than the other parts of the model building. Data capture, data cleansing and data governance is often regarded as the boring part of the process. But this part will be increasingly more important when the other parts of the model building are automated. Context data will be important to capture and use, when you are able to do real-time decisioning. The last minutes data will give you the best data for your decision models. Data Scientist needs to put emphasis on defining the right targets (dependent variable) for their models. The targets are defined by the precise purpose of the model. Many data scientists are fooled of a models statistical metrics (ROC, gain, lift, gini, F-score, confusion matrix, KS), because they often don’t have good metrics of how the model fits the purpose, and the business value it creates.
Machine learning is fueled by data, and data is more important for the quality of the models than the other parts of the model building
4. Set up automated data pipelines. The automation process needs a continuous data flow to be able to give business enhancement at scale. Many of today's IT Engineers do not understand what is needed for setting up pipelines for machine learning. Data Scientists should spend more time with the IT Engineers to teach them what a model building environment needs. The process of storing historical data for model building purposes, and the importance of real-time access of current data in the decisioning.
5. Model management and monitoring. Automated pipelines, gives you a scalable environment for setting up all sorts of models. The process of monitoring what is put into production is important. Data will change and so will the accuracy of your models. You will have to understand the decisions models make, and also understand their weaknesses and how to adjust them. It is not enough to focus on the standard statistical metrics used in model building. A way to handle temporary faults and loss in data sources must also be handled.
Machine learning is quickly becoming an integral part of everything in a business. It’s about intelligent automation of your business processes, and not just a mysterious thing data scientists do for themselves. Fast technical advancements makes it increasingly easy to assemble together readymade building blocks to achieve your goals by buying access to cloud services and integrating them in your own way to create business value. Good luck with your journey!
