In this article Oliver Tearle, Head of Research, The ai Corporation explains why efficient management processes and process automation are the key to managing today’s machine learning models. Enabling organisations to innovate and drive increased efficiencies in today’s digital world.
The use of machine learning (ML) and artificial intelligence (A.I.) has earned its place in many core business functions. With many businesses utilizing the technologies as their base value generation tools. Those businesses have seen how effective ML and A.I. can be and are rapidly expanding its use, further developing business value. However, while this strategy is very effective for the first few model implementations, the more machine learning projects that are undertaken, the more manual effort is required to maintain the models, to ensure they are optimized.
These manual processes may be manageable at first, but as the number of models and channels increase, it can rapidly become untenable to add more and more staff to manage and monitor all the processes. Managing ML and A.I. can become very expensive, which is why efficient management processes are so important. So, what is the solution? AutopilotML is an emerging technology, which resolves this problem by automatically maintaining and optimising the implemented ML models. Using ML to manage the models removes the manual effort involved, enabling the data scientists and subject matter experts (SMEs) to focus on more valuable work.
Learn More: Creating Business Value with Intelligent Automation
Here are 5 distinct components how AutopilotML automates the process of managing machine learning models:
Many data science problems start with significant manual effort selecting the data to present to an algorithm. Throwing all the data you have at a system will usually result in a sub-par model. As there can be many different, often conflicting, signals in the data which need to be targeted and modelled individually. This is especially true with fraud, where different geographical regions, payment channels etc. have vastly different problems and patterns.
Understanding which data to train a model on is a major step in the machine learning journey. As many practical problems involve data, which changes over time – finance, security, weather etc. Therefore, it is important that predictive models are trained on the most relevant, up to date data possible. Some problems may be complex enough that many different models are ‘warranted’ to work together to cover different parts of the data – based on regional or geographical trends, for example.
This becomes further complicated when the data is subject to rapid changes. Such as with fraud, where fraudsters find a unique method and go on a spending spree, before they are stopped. It is essential that this behaviour is detected and stopped as quickly as possible – something which is very difficult with a manually managed solution.
The model creation step comprises of two sub-steps: Information ‘mining’ and model building.
The information mining step is performed to assist with increasing performance of the generated models, usually by creating more information for the model to learn from. This is very time consuming to do manually, as the data scientist needs to uncover relationships between the data elements and devise ways of exploiting the insight as additional data fields for the machine to pick up on, during training. This is important as this additional data very often means the difference between an unsuitable or an excellent model.
The model building stage takes the enhanced data – the base data, plus the additional mined data, and many different models are built on it. A data scientist may try many different machine learning techniques and configurations for this step, which is often where most of the time for the entire process is spent, until a satisfactory model is created.
Learn More: Robotic Process Automation: What is it good for?
The final phase is bulk performance testing and selection of the best performer(s). It is at this stage where some manual input is required, not least, because it is imperative that the user selects the right model for the task. It’s no use having a fraud risk model, which detects 100% of a fraud problem, but challenges every authorisation.
AutopilotML will suggest changes to the overall strategy to maintain or increase performance. This could include a long list of changes, including additions and removal of models, as well as threshold changes due to rapid trend changes in the data; or the list may be empty, as the current strategy remains optimised for the most recent data.
An essential part of the AutopilotML process is deploying and monitoring the performance of individual models, as well as the strategy comprised of several models working together on a common problem.
Once a strategy has been developed and is in production (hopefully providing lots of value!) it is incredibly important this strategy remains relevant and effective on the live data.
Models which are no longer effective (as dictated by some SME pre-set metrics, such as false positive rates) are retired or go back into the process of modification and evaluation to determine if anything can be done to improve performance, such as modifying rule parameters or model thresholds.
This stage is essential for providing information on the strategy performance, such that SMEs and business leaders can quickly review how effective their machine learning strategy is and can understand the trends in the data better. Enabling a more agile response for both positive and negative cases.
This is where nearly all an SMEs time is spent. Does changing a model’s threshold yield better results? What if a new model is created under a new set of constraints? How will all these changes affect the overall performance of the system?
This is where an AutopilotML driven approach comes into its own, as these questions are continually asked and tested, without any involvement from the data scientists or SMEs. The result is a simple list of recommended changes, along with the effects of making those changes. An Autopilot tool continues developing the strategy using the technology available without any human input.
Using best of breed machine learning technology and AutopilotML significantly reduces the amount of time it takes to analyse data models, increasing accuracy whilst reducing decision lag time, leading to a more agile approach to machine learning implementation and increased value from data, faster. When it comes to fraud, this means reduced false positive rates, less declines and more transactions.
Process automation is continuing to innovate and provide increased efficiency and profit gains in the places it’s implemented. The automation revolution isn’t coming, it’s here, so prepare your business for streamlining, more effective, engaged staff and increased profit.
Let us know if you liked this article on LinkedIn, Twitter, or Facebook. We would love to hear from you!
Credit: Google News