Research how we can make the automated categorization generic

[bartjkdp]

We would like to investigate how Den Bosch handles categories and how this relates to how Amsterdam does this.

• Get an overview of categories in Den Bosch
• Compare this with the categories in Amsterdam
• Get a dataset of signals from the past two years in Den Bosch
• Create an initial plan how we could create a general service

[jaccobrouwer]

https://www.mindsdb.com/ ... please check if this is something that could be beneficial(?)

[CBuiVNG]

Currently the ML endpoint only supports main and sub category. Is this sufficient for other municipalities? Or do we need to support an arbitrary tree of categories?

[CBuiVNG]

Initial findings:

• Duplicate text assigned to multiple categories. (sql join error?)

After removal of the duplicates:

• use stop words. Does increase prediction a bit
• use stemming. Not really noticeable
• the messages seems to be prefixed with "user_id:". This will cause issues when a single user enters multiple entries. Tested with a custom pre-processor that removes this user id. Improves prediction a bit
• some messages only have a number. There's probably a location stored somehwere (not exported). Makes it difficult to classify these.
• certain categories are overlapping ('illegale stort' and 'vervuiling' for example) which makes it hard to classify correctly
• Tried over/under sampling to avoid being biased towards the majority class (taking the data distribution into consideration). Results are worse with sampling.