This blog is written in collaboration with Chirag Patel, guest author.
From the beginning of this decade, we have seen the unprecedented growth of systems using machine learning to analyze sheer volumes of data to predict market behaviors at scale. We have seen the rise of using Internet of Things (IoT)-based sensors for acquiring and streaming data from every corner of our life. This sets the stage for the upcoming years, where new experiences would be shaped by a burgeoning IoT, Advanced Machine Learning-based AI systems, and the 5G industry, all of which promise to accelerate the "experience economy" by bringing people, profiles, and preferences together at scale.
A recent report revealing how T-Mobile surged ahead of Verizon in the Wireless Purchase Experience category serves as a testimonial for laying the early groundwork in 5G experience mapping. With the onset of COVID-19, telecom companies are further challenged to meet the surge and sudden shift in user experience, from online content collaboration to an incremental volume of streaming profiles over fixed wireless access (FWA), which is an important first use case to emerge from the rollout of 5G. This observation further bolstered the business argument for validating the industrial perception of baselining 5G experience off of user profiles exhibiting high utilization of 4G.
A small team of machine learning engineers and cloud architects came together to help a tier 1 service provider in envisioning, designing, and implementing a 5G experience determination project. They aspired to build a model that could predict 5G consumer profiles in the presence of fuzzy, uncertain, non-consumable, and missing information in a limited time period. Moreover, this task had a conspicuous lack of 5G operational data, which further challenged the notion of building an intelligence rendering solution on a cloud-native infrastructure that rendered the necessary KPIs for accurate 5G positioning, with room for future adjustments through rapid prototyping.
The objective of achieving 5G experience consistency needed to have two key components. It should provide a way to preserve one or more experience-based calculations derived from two families of data - Consumer Determinants and Operational Determinants - to help answer the following question: What type of users and businesses are likely to adopt 5G on FWA first, and what is the expected cost of laying the infrastructure to support the consistency in 5G service rendered to those users?
While the first family could be sourced from external providers, it was the second one, in the face of missing 5G operational data, that was hard to build. The data science team adopted a creative approach to value adjacency by deploying APIs to curate proxy determinants that were key to the success of existing high-speed fiber connections. What emerged out of this exercise was a fascinating group of next-generation ground truth mined and Deep Learning-scored KPIs for an accurate 5G experience mapping.
The team curated proxy variables such as Distance to Nearest Fiber Line scores and Fiber-connectedness, along with 30 odd labeled attributes from fist family, and passed them through a Neural Network (CNN) written to optimize the ‘Fiber-lit’ attribute, i.e. locations and profiles associated with fiber connection. After a few passes through a 3-layered CNN and subsequent cross-validation with a Support Vector Machine, the team discovered approximately 6% of net-new user profiles, which would have been overlooked based on 4G baselining, exhibited high affinity for 5G FWA within Travis County, Texas.
In terms of business value, this approach rendered room for the consumption and labeling of a new generation of datasets expected to emerge from the actual rollout of 5G. This ensures that early adopters of this methodology will sustain an early mover advantage over competition.
Oracle Cloud Data Science Services were the backbone of the above mentioned project for creating and validating high-quality models faster. We also used Oracle Cloud NVIDIA GPU shapes for massive distributed training. Subsequently, we deployed for inference using Oracle Functions and API Gateway. For data acquisition, we used Oracle Cloud Streaming, Oracle IOT Services and Oracle 5G solutions.
Want to start your own data science project? Check out the Oracle Data Science Quickstart Automation Template.