Transition from clustering to classification?

Question

To date, I have done several ad-hoc text clustering projects which use combinations of topic modeling, k-means, and other algorithms. Basically, the point of these projects was to produce themes for different events based on associated text. The themes were named manually after the appropriate levels of clustering were determined, and are now stored in a csv in the following format:

event_id    majortheme    minortheme         majortheme_id    minortheme_id
12          Job Failure   TWS Issue          1                major1minor1
14          Job Failure   TWS Issue          1                major1minor1
15          Job Failure   Job Abend          1                major1minor2
16          Access Issue  Unable to Login    2                major2minor1
17          Access Issue  Unable to Connect  2                major2minor2

I want to transition from clustering to classification (from descriptive to prescriptive analytics), i.e. being able to take new events (with new event_ids) and classify them based on previous clusterings. This would be sort of an iterative training dataset, as new classifications would be added to previous clusterings, after verifying that the model hasn't gone completely awry. Using Python, what is the best approach to implement this sort of classification pipeline? Is it as simple as saving my initial clustering results and then just using that data as a training set going forward? Then saving the test predictions to the original training dataset, and so on?

Answer 1

In order to build a model to make predictions you need a labeled training set, that is, a training set in which each training example is assigned a class label. Training sets are usually labeled by human experts that use their domain knowledge to manually classify the examples in the training set. You have already done that, as described in your first paragraph.

However, sometimes this process is expensive. In order to decrease cost sometimes semi-supervised learning is applied. In semi-supervised learning, and under some assumptions, a small amount of labeled, together with a large amount of unlabeled data, are used to build a predictive model. The unlabeled data are assigned labels during the training process.

This seems to match your idea in your last paragraph, and it is already implemented in Python. I think that if you found clear clusters in your data it may work very well. Worth a try at least.