Predicting Session-Level Video Streaming Connection Failures by Natural Language Understanding of Historical Tabular Data

Video streaming has gained widespread acceptance, but connection failure is still an critical quality of service problem and hurts user QoE. Accurate and timely prediction of sessions' connection failure enable video service providers to take actions to solve problems. This paper first measures a real-world large-scale online video streaming system's session connection failure problems, and then proposes an accurate session-level connection failure prediction model for this system. The model's novelties are three folds: first, to handle the sparse data distribution of categorical data, such as region, ISP, and device, it builds their deep representations to solve the curse of dimensionality problem. Second, to obtain timely prediction, it proposes a sliding window based temporal average feature. Third, to capture the latent relationships between miscellaneous feature values in the massive tabular data, it draws inspiration from deep natural language processing (NLP) model and proposes a novel feature representation learning method. It first transforms the historical connection tabular data into natural language text and then conducts NLP pre-training and finetuning tasks. Experimental results demonstrate that the developed methods are efficient. With the introduction of above three methods, the prediction performance (AUC) is improved by 2%, 2%, 3%, respectively. Our method is a general method to build prediction model based on tabular data, and can be used to solve similar problems in other areas.