Feature Selection for Prediction of User-Perceived Streaming Media Quality

This paper considers the selection of features, measurements collected from an instrumented media player application, that most accurately predict the user-perceived quality of a media stream. The features are utilized by a nearest-neighbor stream quality prediction algorithm using a distance metric of dy- namic time warping. We explore three ways of selecting fea- tures from this data: manually, by observing how application- layer measurements change with changing network conges- tion conditions; correlation-based; and a mathematically- based technique using principal component analysis (PCA). We compare the prediction algorithm's accuracy obtained us- ing the features selected by each method, using a perfor- mance evaluation metric we term hit rate. Our results show that each method selects one feature set that, when used by our predictor, yields very high hit rates (typically 70-90%), and that each of these feature sets includes one particular fea- ture in common: retransmitted packets. We also show that the correlation-based and PCA-based methods of selecting fea- tures do not consistently select acceptable feature sets for our stream quality predictor, in terms of the hit rates generated by the predictor.