Rules/Methods to use for determining user groups with no review data

I’m working on an analytical problem where I have customers who deal with end consumers. End consumers can leave customers reviews based on their experience (Star ratings, text etc). I’m trying to segregate new/old customers into different quality categories: good, average, bad, for an online advertising platform. I’m looking for insights on rules/methods on how to do this

  1. How to set a quantifiable threshold for calling a customer good, average, bad for business? There is consumer review data available (Imagine a table with customer, and the star rating they received, and/or textual reviews from consumer)
  2. How to classify a new customer on the platform with no or minimal review data available about them particularly, while there’s also some customers who have many reviews available?

Looks like you are trying to triage customers into 3 groups, given 2 pieces of information: star rating, text review.

To deal with the quantifiable threshold issue, the simplest metric to mind is just use the rating and ignore the text review. (like how Net Promoter Score just looks at a 1-10 scale). Star rating should be accurate unless someone misclicks, then you can use text analysis to check if someone wrote a good review and clicked 1 when 10 should mean good. (UI problem). Text review can also help expand the rating scale. If the scale is 1-5, people who want to give -5 or 10 can only express it through text and not the rating. But if you only want to delineate good vs average and average vs bad, star rating seems most convenient.

To deal with the customer with varying number of consumers issue, you can average the scores from consumers for each customer if you only want an overall view and don’t mind losing information in such aggregations. For customers with no reviews, if you are only modelling based on review data, then I see no way to make any judgement, if you had additionally customer features (such as which end consumers they serve), you may infer such missing reviews from other customers. For short reviews, you may look for common keywords or phrases and try to find the sentiment from customer-consumer pairs with longer reviews and who also have those keywords. Such words could be verbs, expression of emotion, or product features. eg. If someone mentions “Butterfly keyboard” with “Mac” you can be quite certain they are unhappy, since a commonly hated feature has come to their attention, or caused enough frustration during use to warrant a review on it to express how bad it is.