Analytics maturity models typically contain four stages: descriptive or “what happened,” diagnostic or “why it happened,” predictive or “what will happen,” and prescriptive or “what should I do.”
Historically, these stages have required humans to define a specific problem, program specific algorithms, interpret the results, and perform an action. This has worked well for descriptive and diagnostic analytics where the problem is known, but has been less effective for predictive and prescriptive analytics where discovery is a goal.
With recent improvements to technologies, artificial intelligence (AI), machine learning (ML) and/or deep learning (DL) are now able to take analytics accuracy for predictive and prescriptive analytics to a new level by helping find new algorithm parameters and self-adjusting routines based on previous results.
AI, ML and DL: How do they relate to each other?
Artificial intelligence can describe a wide range of capabilities and encompasses machine learning and deep learning. Weak AI is the ability for self-learning with specific parameters and desired results. This level of AI can be found in robotics on an assembly line which works autonomously but within a specific set of parameters.
On the other end of the AI spectrum is Strong AI where robots learn and respond as humans do. This level of AI is still science fiction.
In order for artificial intelligence to work, there has to be the ability to learn.
This is where machine learning enters the picture. Machine learning at its core is algorithms that process data, create results, and use the results as feedback to adjust parameters to the algorithm in efforts to increase accuracy. The effectiveness of machine learning is related to the number of parameters programmed into the algorithm.
Deep learning takes machine learning to the next level by defining parameters based on large volumes of expected results.
In the use case of image identification, deep learning takes millions of images of a specific subject to define an expansive list of parameters that can identify that subject. This expansive parameter list is something humans could not feasibly create manually. The resulting algorithm using the expansive parameter list allows image recognition to be significantly more accurate than a manually defined algorithm.
ML and DL for Telecom
Customer retention is a good use case of machine learning in telecom. Historically, retention analysis has been tied to parameters like rate changes, service requests/truck rolls, and tenure. The weighting of each parameter was predefined and programmed into the algorithm.
With machine learning, companies can now add additional parameters to the predictive analysis while allowing the algorithms to learn and adjust weighting factors to increase accuracy. With an increased precision on churn prediction, companies can proactively provide precise targeted retention packages. This could be a change to the general retention packages that agents currently offer, where revenue is lost if the package is offered to customers that may not be at risk of leaving.
Troubleshooting network downtime is another use case for deep learning. With an exponential increase of network traffic from an increasing number of sources, companies could leverage deep learning to determine factors that lead to network downtime.
With a big data solution and the acquisition of network data, companies can use deep learning algorithms to identify parameters that resulted in network reliability issues in the past. Understanding these factors will allow companies can be more proactive in network maintenance resulting in less network downtime and increasing customer satisfaction.
How to move forward?
Machine learning and deep learning require both computational power and large datasets to process against. For companies to advance their analytics in the predictive and prescriptive stages using ML and DL, Big Data is a must.
A large data set does not provide results on its own. Data science, programming, and especially business/domain expertise are required to create effective predictive and prescriptive analytics.
Embracing new capabilities requires a new analytic operating model.
There isn’t a silver bullet for AI, ML, and DL, but ignoring their promise is no longer an option.