Naive segmentation

One approach that has been previously explored and employed (e.g. [Kad95]) is naive segment-based feature extraction. This can be thought of as downsampling to a fixed size. Each channel is divided into $n$ segments temporally. For example if $n=5$ and the stream is 35 frames long, then the first 7 frames of each channel will be ``binned'' in first segment, the second 7 in the next segment and so on. The mean of each segment is calculated, and this becomes an attribute.

Hence, if there are $c$ channels, this will generate $nc$ features. These can be used as input to a conventional attribute-value learner. Four possible values of $n$: 3, 5, 10, 20 will be considered. In the following experiments, we also consider several different learners for the segmented attributes: J48, PART, IB1, AdaBoost with J48 and bagging with J48. Unless otherwise indicated the best result amongst the different learner-segment (there are 20 possible) combinations are presented. This does place a bias in the results towards the segmented learner, since it has more opportunities to randomly fit the test set. However, this was thought the fairest way to compare against TClass's performance, since an average over all learners and segment counts was not fair either.