Combining Clustering and AI for Congestion-Free Mobile Networks

• AI
• Machine learning
• Networks

Share on Twitter : Combining Clustering and AI for Congestion-Free Mobile Networks (New window) Share on Linkedin : Combining Clustering and AI for Congestion-Free Mobile Networks (New window)

I. Mobile Network Service Quality: Improved Connectivity Thanks to AI

Mobile networks are becoming increasingly complex as traffic demand grows and new services and technologies emerge. Carriers must overcome considerable challenges when managing these networks to ensure a high QoS (Quality of Service) while also reducing operational costs. Mobile network management automation is implemented through SON (Self-Organizing Network) functions, optimizing network performance and simplifying associated management tasks.

In recent years, the power of artificial intelligence has been harnessed to automate mobile network management, especially in the context of predictive maintenance.

Anomaly prediction makes it possible for mobile applications or mobile network carriers to be proactive in preventing anomalies before they appear, and to set up their services in a more efficient way.

A learning model based on labeled data is put in place to predict anomalies—especially congestion —in mobile networks. This congestion prediction model improves QoS, but some challenges remain:

• If learning is carried out cell by cell, the number of models to be trained is equal to the number of cells in the network, resulting in substantial implementation complexity.
• If learning is carried out on all cells, the resulting model is inaccurate because the cells do not all behave in the same way, especially between rural and urban areas.

To resolve this dilemma, Orange suggests carrying out a clustering stage before creating the learning model, where cells with similar behavior will be grouped into homogeneous groups or clusters.

Orange has adopted the term “hybrid model” to describe this combination of cell clustering with a supervised AI model. With the hybrid model, a specific learning model can be created for each group or cluster of cells.

It’s an approach that has several advantages:

• It reduces complexity by creating a reduced number of learning models, which is more manageable than having one model per cell, and so addresses our first issue.
• Each model uses a rich dataset due to the fairly large number of cells per cluster.
• The generated models are accurate as they are trained based on a set of homogeneous cells, tackling the second issue.

Experiments based on a real dataset have lent credence to this approach, demonstrating improved prediction of future network congestion. Adding clustering improves model precision and contributes to better congestion prediction.

Another advantage of this approach is that the models are portable from one city to another.  Clusters are also stable over time. This translates into ease of use, while also saving time and resources by avoiding regular updates to clusters.

II. AI and Congestion Prediction in Mobile Cells

1. Principle

An AI-based approach combining two techniques is used to group homogeneous cells into clusters. First, the LBM (Latent Block Model) [1] method is applied to a dataset containing several cell performance indicators (KPIs). LBM is a clustering technique that is able to group both rows and columns in a dataset. KPIs are related to traffic volume, number of active users, latency and cell load, and are selected to detect and analyze congestion in a 4G network. If a cell is congested, the KPIs will exceed certain thresholds. The optimal number of clusters is then obtained by calculating the likelihood between clusters using the ICL (Integrated Completed Likelihood) criterion [2].

2. Improved Performance

An initial experiment was conducted to predict congestion based on data from cells in Paris Nord-Est, France. This clustering model identified a dozen groups, divided into three categories: high, medium and low congestion rates. Next, a supervised learning model based on logistic regression was created for each group.

For the highly congested category, which accounts for 70% of the congestion in the initial dataset, the prediction model’s performance was improved compared to the performance of the global model created with all cells. Over a prediction horizon of one hour, the model is able to predict 90.4% (recall) of congestion with a precision of 90.5%.

For moderately congested clusters, the prediction model maintains the same performance. However, the precision of the models was increased by providing adaptive configuration of the logistic regression.

Cells in the low-congestion category, which accounts for nearly 20% of congestion anomalies, were eliminated from congestion prediction. This is because it is impossible to predict congestion for this type of cell, as they experience so few instances of congestion. It would be difficult to distinguish these cells without using the clustering technique as they generate a large number of false congestion predictions. A global model without the clustering technique would generate more false alarms than good predictions, which would be problematic since the operational team or SON would have to correct non-existent congestion in a cell.

The precision of mobile network congestion prediction is greater than 90% thanks to AI.

III. Applying the Hybrid Model

Portable learning models have several advantages. For one, existing models can be reused, saving time and resources while improving performance. This section demonstrates that the hybrid model is portable from one city to another, meaning it can be used in different geographic contexts. For another, this model is able to accurately predict future congestion, allowing carriers to proactively monitor and manage anomalies before they appear.

The model has been applied for the following French cities: Paris Nord-Est, Strasbourg, Belfort and Lille. In Figure 1, the different clusters are represented by individual colors. It shows that distant cells located in different cities behave in a similar way toward the anomaly. At the same time, geo-clustering techniques that group together geographically close cells in the same cluster are not always accurate. The example in Figure 2 illustrates two adjacent cells, which are only 0.01 km apart, that do not belong to the same group. This result is linked to the different orientation of the cells: the blue cell faces the forest, while the red cell faces toward the buildings. As a result, the blue cell belongs to a group of low-congestion cells, while the red cell is part of a cluster of highly congested cells. These observations call into question the effectiveness of approaches that cluster cells based on geographical proximity alone, and underline the importance of taking other factors—such as cell behavior, as reflected in KPI measurements—into account.

Figure 1: Clusters obtained for the cities of Paris Nord-Est, Lille, Strasbourg and Belfort.

Figure 2: An example of two neighboring cells that do not belong to the same group.

To illustrate the portability of the hybrid model, let us take a cluster called A—which includes 83 cells in Lille, a handful in Belfort, 36 in Paris Nord-Est and 65 in Strasbourg—as our example. After examining cluster A cells located in Belfort, it is difficult to create a precise learning model as only a small number of cells are available. This underscores the importance of having a portable model; it means the information and knowledge that has already been acquired in other cities can be leveraged to compensate for the lack of Belfort-specific data.

A learning model is created based on data associated with the cluster A cells in Lille. Next, this model is used to predict congestion within cluster A cells located in Strasbourg. The results are remarkable: The learning model transferred from Lille is able to accurately predict future congestion in Strasbourg. Over a prediction horizon of one hour, the model is able to predict 87% of future congestion with a precision of 90%. This proves how effective the hybrid model’s portability is and how the model is able to generalize knowledge from one city to another.

What’s more, cluster creation is stable over time, indicating that the hybrid model is not complex. The clustering approach is able to form near-identical groups of cells over two consecutive years, with only slight differences. These minor differences are mainly due to errors during data collection. This stability serves to consolidate the reliability of our approach, and affirms the model’s ability to form coherent clusters.

IV. Conclusion

An AI-based hybrid model improves congestion prediction in cellular networks. This model combines a clustering technique with a supervised learning algorithm, allowing it to predict anomalies more accurately to guarantee high QoS. The hybrid model is capable of distinguishing cells that react similarly to network congestion. By adapting this model, prediction performance is improved while the number of learning models needed is reduced.

The approach outlined is geographically portable, meaning that learning models can be transferred from one city to another. This has proved particularly useful in cases where limited data is available for certain cities, serving to compensate for the lack of specific data. Furthermore, this approach opens up a path to accurate prediction of future congestion in cells located in different cities.

In terms of possible next steps, Orange suggests experimenting with the hybrid model on a live Orange network. This expansion would allow the model to be tested in a variety of contexts and could collect data from different cities.