Building a Case for Advanced Analytics in the Capacity Planning process

There has been significant data growth in the world of telecommunications during the modern era of mobile technologies, i.e., 4G and 5G. While data growth has often been marketed as a massive window of opportunity, in reality, the ever-expanding volumes of data are overwhelming for existing systems, thereby indirectly leading to issues concerning operational efficiency and network performance. To utilize the vast volume of structured and unstructured data, organizations must have the right innovation-driven solutions to maximize their revenue potential.

As a part of the network capacity management team, I would like to elaborate on the above from the perspective of capacity planning. In network capacity planning, it is essential to identify potential shortcomings or, in other words, factors that can impact network efficiency and performance. Some of these factors include the need for capacity expansion due to the unavailability of adequate network resources for handling the traffic demand, or an optimization requirement in the network investment plan for better utilization of existing resources. Traditionally, most of these decisions are made manually, which we all can agree, is not practical, efficient, or even error-proof. An AI-based capacity management system is the need of the hour to exploit the full potential of resources and network data.

How AI can enhance Capacity Management

AI and advanced analytics are terms that get used a lot, across industries, especially by marketing teams. However, beyond the ‘buzzword,’ these technologies have an essential role to play in enabling improvements from current, traditional processes. Advanced analytics-driven capacity management solutions can help users decrypt relevant information to derive useful insights by applying data analysis, predictive analytics, forecasting, and optimization models. These insights, backed by data, and fine-tuned through Machine Learning, can help organizations make the best possible decisions by anticipating and implementing necessary network changes ahead of the demand curve.

A significant component for providing reliable Network Investment Planning (NIP henceforth) recommendations towards Capacity Enhancement/Optimization is through the accurate forecasting of capacity KPIs. Forecasting is the foundation on which NIP recommendations are based.

A pool of forecasting models can be used to predict network performance. These models can be broadly classified as qualitative methods and quantitative methods. Qualitative methods are used when historical data is not available or when data is noisy. Quantitative methods, on the other hand, make forecasts based on mathematical models rather than subjective judgement. Choosing the right forecasting model depends on many factors like the context of forecast, relevance, availability of historical data and the desired level of accuracy.

Which is the right forecasting model to adopt for your business?

Different forecasting methods have distinct advantages and disadvantages. Therefore, selecting the right forecasting method is of critical importance to all decision-makers.

My data science team at Subex and I have built a proprietary machine-learning model for forecasting called CapMan4C. During this experience, we also collaborated with a team of domain experts to have clarity on the problem which businesses are trying to resolve and to understand data specific to the network domain.

I bring up this experience to highlight that the thoughts involved behind building a proprietary framework for forecasting are based on two criteria:

• What does our expertise in the telecom domain tell us?
• What do our customers say?

Based on these insights, it is evident that traditional forecasting time series models are observed to not perform well on network data. Using my prior data science experience in the healthcare industry, I applied different machine learning models like decision tree, random forest, etc., for forecasting on telecom network data. I observed that plugging in these existing models do not provide the desired level of accuracy. The approach of leveraging a domain-driven model eliminates the problems mentioned above. Its framework is designed so that the selection of a base model for forecasting does not solely depend on traditional time series models.

To further enhance the accuracy of the model, this kind of framework also needs to include three key components:

1. Recency factor – The general assumption while working with time series data is that all the historical data points have equal weightage. However, when the event ‘recency’ is accounted for, this notion does not always hold. In theory, an event that has taken place more recently is more indicative of the current situation than a distant past.
2. Feature engineering –Feature engineering is a creative process in which we experiment and extract features that can be useful for enhancing the model. Several features, if extracted at a granular level from the data (including data from external sources), can help identify hidden trends and seasonality in data and make more insightful forecasts from the machine learning model.
3. Auto-hyperparameter tuning– Hyperparameters are parameters of ML-models specified by “hand” based on expertise and domain experience. These parameters can be used to control the learning process, which can significantly impact the performance of the trained model. The more flexible and powerful an algorithm is, the more design decisions and adjustable hyper-parameters it will have. Traditionally, Grid Search or Random search is used to perform hyperparameter optimization. Grid Search works by searching exhaustively through a specified subset of hyperparameters. This can be very time consuming and computationally expensive. Random search is slightly better but is similar to Grid Search because it pays no attention to past evaluation. As a result, network teams often spend a significant amount of time evaluating “bad” hyperparameters.

In summary, operating a network is a complicated endeavor, and regardless of how much planning takes place, problems do occur. Identifying the right kind of AI-driven approach toward capacity management provides network planners the ability to analyze “what-if scenarios” based on forecasts from ML models. It helps them take calculated decisions on Network Investment Planning based on AI-based recommendations. This will help them to mitigate risks proactively and maximize their revenue potential.

To understand these key pillars, and unlock the full potential of your network

Read This Whitepaper

Sourav Modi

Sourav has 9+ years of experience in the IT industry. He joined Subex on 15 April,2019 and has been involved in bringing AI/ML capabilities in the network Analytics Domain for telecommunications industry.

Before joining Subex, Sourav had been involved in bringing AI/ML capabilities in US healthcare industry. In his previous role, he has built ML models like classifying cancer as benign or malignant, predicting for chronic kidney disease given various blood features as indicators, etc. He has also worked with Deloitte, IBM in the past.