π» Using genAI to generate labels for clusters for use as magnets
If we are going to use some set of labels as magnets, we face a tension:
- on the one hand want them to express the generality we intend: the label should express the fact that this is a group, like 'health behaviours'; we are expressing the fact that we do NOT expect the raw labels to express this generality but to express specific examples.
- but this will make them perform worse as actual magnets because the best magnets should remain in single-case formulations and not try to generalise.
So if we have many labels like "school creativity project in North district implemented", then it is better that the cluster label is also of that form, and not for example "creativity projects implemented in multiple schools"
So if the raw labels often express ideas like "girls responded to the training" and "boys responded to the training", if we provide a pair of labels like this as magnets, they should perform well, whereas "children responded to the training" will not perform so well.