avouros
commented
7 years ago introduced in version 3.0.0 (commit: https://github.com/RodentDataAnalytics/mwm-ml-gen/commit/1ebfe0c31c7524a032e3ae0da3a223774123904b).
Some research is still needed to achieve good tuning.
Up to version 2.0.0 For the classification of segments a classifier is created based on 1 optimal number of clusters. This classifier is selected because of the low classification error and the high trajectory coverage. The disadvantage of this method is that there may exist multiple 'good' classifiers but each one of them may lead to different results. Thus, which one is more appropriate?
After version 2.0.0 a) Majority rule: The software will generate a number of different classifiers (different number of default clusters) the perform on each segment a simple majority rule; "what strategy most of the classifiers say that this segment fall into? Then this segment will be fall into that strategy".
b) Different segmentations: It is important that we have at least two segmentations with different overlaps (usually 90% and 70%). The user will provide labels to both of them and afterwards 2 different sets of majority-rule-merged classifiers will be generated. If the results of these two groups of classifiers differ then statistics will be extracted from the common segments of the two segmentations (how each classifiers has label each of these segments?). It is very likely that the one set of classifiers has learnt some strategies worst than the other one thus the user will need to provide to this one some more labels for these strategies. Finally the the sets of classifiers are bound to agree on the results.
Note: Instead of adding more labels another way is to merge the two set of classifiers using majority rule again. In most cases the end results will be the correct one.