|
Abstract:
|
The ability to collect (near) real-time metric data from set-top boxes (STBs) creates opportunities for early problem detection and proactive intervention using statistical models; however, the volume and complicated nature of the data create several challenges for the data analyst. This paper studies data collected by a large U.S. TV service provider compiled over a two month period. Our objective is to build a model to identify patterns related to hard to solve video problems that are likely to lead to multiple repair attempts. Once known, these patterns can be shared with technical experts who can help identify root cause and take action. It is expected this will help shorten the time to repair, reduce costs associated with multiple repair attempts, and ultimately improve the overall customer experience. In this talk we focus on the difficulties that arise in developing an effective model. We propose several strategies related to training sample construction, dimension reduction, and transformations to the raw data that help improve performance.
|
