Create mapping between training data and labels

I was just wondering, whether such a mapping is the right way to do our task as we have to map the business to one of the following properties:

0: good_for_lunch 1: good_for_dinner 2: takes_reservations 3: outdoor_seating 4: restaurant_is_expensive 5: has_alcohol 6: has_table_service 7: ambience_is_classy 8: good_for_kids

In the mapping I uploaded last week, each photo of the same business has the same labels. E.g. a picture of a cockail might not only have label 5 but also others that are not directly associated with the picture (e.g. 3). So I think that we would not teach the machine optimally if we do the mapping this way. Shouldn't it rather be like the following: We have several photos of a business and each photo 'contributes' one (or several) label(s) to the business? (This information is not contained in the data though....)

I would have two ideas in case the photo-id mapping doesn't work as desired:

1.)

For each label type check the businesses that are classified only at this label (and maybe other labels at the same time?)
let the machine find similiarities in their pictures and classify this 'picture-type' as the corresponding label

2.)

gather photos for each business and its labels
teach the machine to associate specific sets of photos with a specific label combination

Sorry for deluging you with so much stuff... I just had some ideas coming to my mind, and wanted to fix them and ask for your opinions.

sacdallago / dataminer

Create mapping between training data and labels #4