In this paper we compare various anomaly detection techniques for cellular network data using a multivariate dataset from Milan and Trentino. We evaluate traditional statistical methods (Z-score, IQR), machine learning (One-Class SVM), and deep learning approaches (LSTM-Autoencoder, Transformer). Experimental results indicate that deep learning models significantly outperform traditional methods in terms of both accuracy and efficiency. The Transformer model achieved 96.5% accuracy in 23 epochs, compared to 93% by the LSTM-Autoencoder in 40 epochs and 86% by the One-Class SVM.

Nikol Gotseva; Atanas Vlahov; Roland Mfondoum; Antoni Ivanov; Vladimir Poulkov

2025 60th International Scientific Conference on Information, Communication and Energy Systems and Technologies (ICEST)

https://ieeexplore.ieee.org/abstract/document/11098230?casa_token=QTpN97pfdBcAAAAA:6-uXOZLfAocia49GqhmRSw_7niquGPRJmhZdb1kD_-iniF5Ii1qrN9eh4yorXd-qCK0xP53Kz3w

Cellular networks; Training; Deep learning; Solid modeling; Accuracy; Statistical analysis; Transformers; Solids; Data models; Anomaly detection