Novelty Prediction in Broadband Line Multi-variate Time Series Using a Deep Long Short-Term Memory Network

To ensure reliable broadband performance, particularly amongst customers working from home and reliant on home broadband performance, novelty prediction is an important part of the internet service providers (ISPs). It is ideal to identify the onset of poor broadband performance in real time before the customer experience, so that an ISP can take appropriate action. This paper proposes a Broadband (BB) line novelty prediction system. A deep bidirectional long short-term memory (BiLSTM) recurrent neural network is used for sequence-to-label classification using a telecoms dataset. This BiLSTM network, being able to understand historical information, is well-suited for sequence-to-label classification and the prediction of novelties before novelties occur. Time series observation points are retrieved from control hubs and faulty hubs over periods of five days, and a new feature is created to represent data incompleteness. The pre-processing phase produces descriptive features for each component to summarize the data characteristics. Both sequence-to-label decoding and our proposed sequence-to-label-rate decoding methods are then employed to make novelty predictions. A real telecoms dataset relating to customers is used to evaluate the performance of the system developed. The results are compared with traditional machine learning algorithms and show that the proposed system can predict novelties proactively, and with improved accuracy.