Topology reconstruction in telecommunication networks: Embedding operations research within deep learning

Abstract

We consider the task of reconstructing the cabling arrangements of last-mile telecommunication networks using customer modem data. In such networks, downstream data traverses from a source node down through the branches of the tree network to a set of customer leaf nodes. Each modem monitors the quality of received data using a series of continuous data metrics. The state of the data, when it reaches a modem, is contingent upon the path it traverses through the network and can be affected by, e.g., corroded cable connectors.

We train an encoder to identify irregular inherited events in modem quality data, such as network faults, and encode them as discrete data sequences for each modem. Specifically, the encoding scheme is obtained by using unsupervised contrastive learning, where a Siamese neural network is trained on a positive (true) topology, its modem data, and a set of negative (false) topologies. The weights of the Siamese network are continuously updated based on a new modified version of the Maximum Parsimony optimality criterion. This approach essentially integrates an optimization problem directly into a deep learning loss function.

We evaluate the encoder’s performance on simulated data instances with randomly added events. The performance of the encoder is tested both on its ability to extract and encode events as well as whether the encoded data sequences lead to accurate topology reconstructions under the modified version of the Maximum Parsimony optimality criterion.

Promising computational results are reported for trees with a varying number of internal nodes, up to a maximum of 20. The encoder identifies a high percentage of simulated events, leading to nearly perfect topology reconstruction. Overall, these results affirm the potential of embedding an optimization problem into a deep learning loss function, unveiling many interesting topics for further research.