From the course: Machine Learning in Telecommunication: From Basics to Real-World Cases

Supervised learning: Learning from labeled data

From the course: Machine Learning in Telecommunication: From Basics to Real-World Cases

Start my 1-month free trial Buy for my team

Supervised learning: Learning from labeled data

“

(bright music) - [Instructor] Ever wish you could predict network performance or customer behavior before it happens? With supervised learning, we can actually do that. Let's take one example of 4G network, where we have data showing how signal quality or also termed as signal to interference and noise ratio, SINR in technical terms, can affect the throughput or data speed. We have collected this data over months from different sites, and the machine analyze the data and discovered the correlation between SINR and the throughput, like how higher SINR typically leads to higher throughput or faster speeds, and it builds a mathematical equation based on this correlation, and this becomes the model. So the next time we enter a new SINR value, the model uses this equation to predict the new throughput. So x in this case is the input, which is signal to interference and noise ratio, and y is the output for any new given x or any new given input. This is called as regression problem in supervised learning, where the output, like throughput or speed, changes continuously with the input, which is SINR. And while sometimes this relationship is linear here, supervised models can also learn more complex or nonlinear patterns. So if we have more than one inputs, if it is more than SINR which is contributing or impacting the throughput, could be a signal strength, could be a modulation, in that case, we have many different input factors to be included in this model. In that case, the patterns, which is currently looking like a linear pattern, could be a nonlinear. So that is also being supported. But supervised learning isn't just about the numbers, it's also about the decisions. Let's check another example of customer churn. We want to predict whether a user is likely to stay or leave the network. And this is where the classification problems comes into the picture, where output is a category. Like yes or no, it's a binary. It's 0 or 1. We feed the model data, like how often they experienced dropped calls, how is their average data speed, or how high their monthly bill is. So if there is a deviation from the average between these different factors, the machine learns to classify that each customer would leave or not, and it'll classify them into two groups, those likely to stay and those likely to leave the network. And this lets telecom companies take action before it's too late, like offering better plans to high-risk users. Where else can supervised learning be used in telecom? It can be used in network performance prediction. You can use past SINR and throughput data to predict user speeds in a new area or new network deployment. And this helps telecom operators offer realistic services to the consumers, and then they can make some marketing plan, providing realistic data plans based on what users can actually expect. Another example is to detect certain issues in the network. So supervised models can also be trained to classify problem areas, like sites with poor signal quality. If many signals show low SINR and throughput in one location, the model can flag the site for investigation and that help engineers target the improvement area faster. Another use case could be traffic forecasting. Here where it is getting a bit more advanced, we can also predict future traffic patterns using supervised learning, especially when we treat the task as a time series prediction. So by feeding in the past traffic data, we got the different sites and their traffic for maybe six months, the model learns the trend and can predict things for upcoming weeks or months, that how the peak load time at specific sites would look like. And then on the basis of that, engineers can make certain plans to add capacity or to mitigate this problem of high traffic or high utilization in some sites. So even though this touches the world of forecasting model, it still falls under supervised learning if the past data includes labeled timestamps and usage levels. So whether it is predicting speed, spotting churn risks, or forecasting traffic, it is one of the most powerful tools in the AI toolbox, especially in telecom. So the next time you look at network data, remember if you know the output you want to predict, supervised learning is the way to go.

Contents