Exploiting Knowledge Across Distinct Domains: Learning Event Details from Network Logs

Modern wireless networks record large amount of network log data. Network administrators and central controllers utilize log data to monitor the health of networks and optimize system performance. Though log-structured data are a great source of data revealing recurrent patterns in networking, they have not been exploited for insights in a wider scope, compared to the recognition of time-series signal data and sequences in language and text. We design a data processing method to convert incomprehensible log data into intelligible vectors. We also propose to utilize recurrent models combined with an output layer of multi-label linear transformation to learn on-campus event details from exploitation on data in network domain. Our evaluation shows performance advantages (metrics of ROC, accuracy and F1 score)of recurrent models over SVM, by leveraging temporal nature in network logs.