AI Sonalake

Predicting the future values of the RSL can assist the network operators to maintain optimal network performance and prevent Qos violations before they happen. We worked closely with one of the leading mobile operators and the domain experts to build a solution that predicts the QoS violations based on RSL drop up to seven days prior.

Outline of Innovative Business Solution
Dataset Overview:
To do this we used three months’ worth of real world dataset from a mobile operator in Ireland. The dataset consisted of more than 15 million records in the form of the daily snapshots of the maximum and minimum RSL for over 16,000 transmission links spread all across the island. The dataset also contained descriptive network data such as the types of transmitting antenna, height of antenna, the precise geolocation of the transmitting and receiving antenna etc.

In addition, we also used the weather data for the location of each of the transmitting and receiving antenna using the coordinates from an online weather service using an API that reports live and forecasted weather data for any geo-coordinate.

Model:
To forecast QoS violations for a transmission link, we converted the regression problem into a classification problem. We assigned class labels as ‘Critical’ when the RSL value was below -70 dB and ‘Normal’ otherwise. Following which, the entire time series was broken down into sequences of seven days and we calculated the following statistic on those features for seven days.

Some of the features we extracted have been listed below:
– Mean
– Standard deviation
– Average absolute deviation
– Minimum value
– Maximum value
– Difference of maximum and minimum values
– Median
– Median absolute deviation
– Interquartile range
– Number of values above mean
– Number of peaks
– Skewness
– Kurtosis
– Energy

After the class labels were assigned and the features were extracted, we tried various class balancing techniques as the dataset was skewed in favor of normal in the ratio of 39:1 (i.e. for each critical we had 39 normal classes in the dataset). To solve the issue of imbalanced class distribution we tried Synthetic
Minority Oversampling Technique (SMOTE) and undersampling and found that controlled undersampling of the majority class was the most efficient solution in terms of performance and scalability.

After Data set preparation, for the modeling phase we used the XGBoost model given its scalable nature and its ability to handle imbalanced data.

Detail on Analytics-Based Approach
We needed to forecast the RSL values seven days in advance so that the Network operators have enough time to take corrective measures when the model predicts QoS violations on a specific transmission link. In order to achieve this, we used the concept of direct multi-step forecasting.
Direct multi-step forecasting consists of training a different model for each step of the forecast horizon. For example, to predict the next three values of a time series, three different models are trained, one for each step. As a result, the predictions are independent of each other.

As illustrated in the diagram we have one model for each time step ahead. Model 1 makes the predictions for time t+1 by taking the rolling features from the previous time three time steps and combining them with the forecasted weather features for the closest geo location to the receiver of the transmission link. We observed that the RSL values are drastically affected by the prevailing weather conditions such as temperature, wind, wind direction, precipitation etc., and a lot of the variation in the RSL values can be explained using the weather features.

(IMAGE: 2 – Sonalake – Predicting the impact of Weather conditions on QoS in
Mobile Networks.png)