Preliminary Analysis Report of the Network Dataset

The measurements conducted by the probes, for example, the individual UDP-jitter tests, may be described in terms of the attributes of the probe (ranNode, apNode, hubType, headlineSpeed, location, and etc), the test (week day, hour, minute, target, and etc), and in fact anything else that we can monitor (the features). With this, if the distribution of a variable we are interesting in monitoring, say the “up-stream UDP jitter”, is dependent on a particular combination of feature values (or combinations of features), we can attempt to characterization this dependency, for example in terms of the correlation between a particular observation and some (set of) feature value(s). In this way, we can attempt to identify and explain the performance measurements collected, such as for the jitter, in terms of the features we observe. For example, we may identify that the up-stream UDP jitter measured by probes on DSL-lines in London to targets in New York is normally high/unstable in the evenings on weekdays, and should not be interpreted in the same way as say a measurements from the morning. Conversely then, anomalies here are observations that cannot be adequately explained by expected outcome of the feature combination that generates them. Our goal is therefore to build models that are able to explain measurements at the network level, in terms of the features we are able to monitor. something on flexibility, space, multi modal data ... To motivate the importance of considering the range in the values of features, and the combinations of features, we report below a small study that shows we are able to better characterize the distribution of the variable of interest (UDP jitter-up), if we explicitly account for the range of individual features and the impact of considering features jointly.