The case for machine learning transforming the telecoms industry

The telecommunications sector is riding high on the waves of the tech revolution and digital transformation.
24 June 2021

Ways machine learning is transforming the telecommunications industry. AFP PHOTO / JONATHAN NACKSTRAND (Photo by JONATHAN NACKSTRAND / AFP)

  • The telecoms industry needs machine learning to be able to process and regain control over what’s done with available data.
  • The technology’s use cases in telecoms have shown great potential in assisting with anomaly detection, root cause analysis, managed services, and network optimization.

As technologies like artificial intelligence (AI) and machine learning (ML) become ubiquitous, it will be almost impossible to come across any industry not capitalizing on the benefits they can provide. The telecoms industry has traditionally navigated quite well through tech change.  Globally, they managed to transform from landline to mobile carriers, then move from voice calls to messaging and data-centric networks. In most of the developed markets, telecoms are creating ecosystems for the data-driven economy.

The reality on ground is that the telecoms industry is one of the fastest-growing industries as well as one that uses AI and ML in many aspects of their business from enhancing the customer experience to predictive maintenance to improving network reliability. ML in particular has numerous potential use cases since telecommunications companies deal with vast amounts of data and need to drive conclusions from it – an overwhelmingly difficult proposition to do manually. 

According to Ericsson in a blog post, “In the area of system monitoring, anomaly detection systems are crucial for identifying performance issues and problematic network behavior. Proactively predicting the degradation of key performance indicators, and identifying the likely root cause, can help reduce and prevent outages.”

As for the area of managed services, Ericsson said ML models can improve trouble ticket management by effectively classifying, prioritizing, and escalating incidents. Capacity planning and customer retention can be improved through explainable churn prediction.  “Furthermore, in the area of intelligent networks, the incorporation of ML tools can enable self-healing radio networks, which automatically detect issues and take corrective actions,” the report said, adding that new technologies such as deep learning and reinforcement learning can be used to automate the network design process and optimize network performance in real time.  

Common ML system components & use cases

Data is the lifeline of any ML system and telecoms data is complex, multimodal, and plentiful. It comprises numerical metrics and text-based logs collected from many thousands of devices. The computational and communication costs of processing the data, as well as the latency and performance requirements, determines how the data components should be designed and implemented. 

Another use case would be the offline and online predictions. ML predictions can be made in either periodically scheduled batches (offline), or in a dynamic streaming manner in real time (online) and batch prediction may be suitable when some delay is acceptable. In batch prediction, model prediction requests are accumulated over time, and the model responds to each batch of requests at an appropriate, predetermined time. 

The report stated that for mission-critical tasks such as predicting service outages, however, real-time predictions may be required. In this mode of operation, the ML model service immediately returns a prediction output upon receiving input data. This execution mode can have challenging requirements from an operational standpoint because real-time prediction may need to support a large and unpredictable number of requests, the model service may need to scale dynamically and provision more resources at peak request times. 

Then there’s the workflow management use case as well whereby to orchestrate the entire end-to-end ML pipeline, workflow management tools can help immensely. “An ML pipeline consists of a number of inter-dependent tasks including data collection, transformation, validation, training, and serving. Workflow management tools can help effectively chain these tasks together, such that unexpected delays or issues in one step do not break subsequent steps,” Ericsson said. 

While ML solutions are complex systems composed of several components that may differ from the existing infrastructure organizations have in place, the report said, depending on the particular use case, each of these sub-components may be implemented in a different manner.