I'm currently in the process of tagging a bunch of photos. I started off with some tags being binary variables (i.e. the tag in question was either present or absent) and some ordinal variables , however it seems switching as many as possible to ordinal variables (e.g. Very light, Light, Dark, Very dark) makes more sense as I get more info in the analysis later on. This switching off binary to ordinal is not possible for all tags. So I have a mixture of tags mostly ordinal (4pt scales) and some binary.
The initial tagging process may/will result in a lot of redundant tags, so I plan to do some dimension reduction on them (according to which tags drive particular metrics I'm concerned with and have the data for - these metrics are continuous). The trouble is I'm not sure which method is most appropriate to do this dimension reduction. From my understanding, PCA assumes the variables are all scale-level and also that variables are continuous. I came across 'Optimal Scaling' and CatPCA which sounds like it "converts" ordinal scales to continuous ones by imposing some sort of distance function on them. However I'm sure unsure as to if this method is appropriate for multi-scale data.
Does anybody have an idea as to which analyses is most appropriate here and which languages/library's/platforms support such analyses (if an appropriate way forward exists in Python I'll definitely go for that, however if not then hopefully R?).
In the end the goal is to use the reduced subset of score-driving tags as options for other people to use. Another goal is to use these reduced subset of tags along with the scores to train a Neural Net to predict the scores of a photo given a collection of tags.