IMPLEMENTATION

An instance of the above architecture, called TClass, has been implemented. This implementation is designed in an object-oriented manner to allow different PEPs, global feature calculators, clustering algorithms and learners to be substituted.

Currently, the following global feature calculators are available: mean, median, mode, maximum value and minimum value for a channel. These can be applied to any channel.

Similarly, the PEPs implemented are:

• Increasing (an extended period with significant positive gradient) and Decreasing (an extended period with significant negative gradient). These both return four parameters: start time, duration, average value of the channel, and average gradient.
• Flat (no perceptible change in gradient). This returns three parameters: the start time, average value and duration.
• Localmax and Localmin. These both return two parameters: time of the maximum/minimum, and the value.

All of these have some simple heuristics for dealing with noise. For instance the local maximum applies a smoothing filter first to reduce noise before locating maxima.

Currently, TClass uses k-means clustering, with all values normalised by standard deviation. The confidence metric used for cluster membership is distance to the cluster's centroid. The learner can either be a naïve Bayes learner or C4.5 [Quinlan, 1993] with the default parameters.

 

 

Figure: Approach used in implementing to simplify clustering task. Classifier 1 through to Classifier N are voted to make a final classification.

Empirically, it was found that the k-means clustering algorithm did not perform well when all of the events from different classes were clustered simultaneously. Thus a slight modification was made to the architecture; the basic architecture was replicated from the event clustering stage onwards on a per-class basis, as shown in figure 7. Event clustering is now performed only on events coming from instances of the same class. These clusters are then used to extract the synthetic events as before. All training instances are then attributed with these synthetic events, creating a set of synthetic event attributes. These class-specific synthetic event attributes are combined with the class-independent global features, creating a set of features that are used by the learner to induce a classifier. Because one such classifier is constructed per class, if there are n classes, there are n classifiers induced.

To classify an unseen instance, event extraction and global feature calculation are applied. For each class, the instance is then attributed with that class' synthetic events, thus generating the synthetic event attributes required by the classifier associated with that class. The instance is then classified by each classifier. In this way, n classifications will be made. The most common classification is then taken as the final prediction.